All articles

What We’re Reading (Week Ending 30 March 2025)

The best articles we’ve read in recent times on a wide range of topics, including investing, business, and the world in general.

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 30 March 2025:

1. Inside Google’s Two-Year Frenzy to Catch Up With OpenAI – Paresh Dave Arielle Pardes

A hundred days. That was how long Google was giving Sissie Hsiao. A hundred days to build a ChatGPT rival.

By the time Hsiao took on the project in December 2022, she had spent more than 16 years at the company. She led thousands of employees. Hsiao had seen her share of corporate crises—but nothing like the code red that had been brewing in the days since OpenAI, a small research lab, released its public experiment in artificial intelligence. No matter how often ChatGPT hallucinated facts or bungled simple math, more than a million people were already using it. Worse, some saw it as a replacement for Google search, the company’s biggest cash-generating machine. Google had a language model that was nearly as capable as OpenAI’s, but it had been kept on a tight leash…

…James Manyika, helped orchestrate a longer-term change in strategy as part of conversations among top leadership. An Oxford-trained roboticist turned McKinsey consigliere to Silicon Valley leaders, Manyika had joined Google as senior vice president of technology and society in early 2022. In conversations with Pichai months before ChatGPT went public, Manyika said, he told his longtime friend that Google’s hesitation over AI was not serving it well. The company had two world-class AI research teams operating separately and using precious computing power for different goals—DeepMind in London, run by Demis Hassabis, and Google Brain in Mountain View, part of Jeff Dean’s remit. They should be partnering up, Manyika had told Pichai at the time.

In the wake of the OpenAI launch, that’s what happened. Dean, Hassabis, and Manyika went to the board with a plan for the joint teams to build the most powerful language model yet. Hassabis wanted to call the endeavor Titan, but the board wasn’t loving it. Dean’s suggestion—Gemini—won out…

…To build the new ChatGPT rival, codenamed Bard, former employees say Hsiao plucked about 100 people from teams across Google. Managers had no choice in the matter, according to a former search employee: Bard took precedence over everything else. Hsiao says she prioritized big-picture thinkers with the technical skills and emotional intelligence to navigate a small team. Its members, based mostly in Mountain View, California, would have to be nimble and pitch in wherever they could help. “You’re Team Bard,” Hsiao told them. “You wear all the hats.”…

…Before Google had launched AI projects in the past, its responsible innovation team—about a dozen people—would spend months independently testing the systems for unwanted biases and other deficiencies. For Bard, that review process would be truncated. Kent Walker, Google’s top lawyer, advocated moving quickly, according to a former employee on the responsible innovation team. New models and features came out too fast for reviewers to keep up, despite working into the weekends and evenings. When flags were thrown up to delay Bard’s launch, they were overruled…

…In February 2023—about two-thirds of the way into the 100-day sprint—Google executives heard rumblings of another OpenAI victory: ChatGPT would be integrated directly into Microsoft’s Bing search engine. Once again, the “AI-first” company was behind on AI. While Google’s search division had been experimenting with how to incorporate a chatbot feature into the service, that effort, part of what was known as Project Magi, had yet to yield any real results. Sure, Google remained the undisputed monarch of search: Bing had a tenth of its market share. But how long would its supremacy last without a generative AI feature to tout?

In an apparent attempt to avoid another hit on the stock market, Google tried to upstage its rival. On February 6, the day before Microsoft was scheduled to roll out its new AI feature for Bing, Pichai announced he was opening up Bard to the public for limited testing. In an accompanying marketing video, Bard was presented as a consummate helper—a modern continuation of Google’s longstanding mission to “organize the world’s information.” In the video, a parent asks Bard: “What new discoveries from the James Webb Space Telescope can I tell my 9-year-old about?” Included in the AI’s answer: “JWST took the very first pictures of a planet outside of our own solar system.”

For a moment, it seemed that Bard had reclaimed some glory for Google. Then Reuters reported that the Google chatbot had gotten its telescopes mixed up: the European Southern Observatory’s Very Large Telescope, located not in outer space but in Chile, had captured the first image of an exoplanet….

…Hsiao called the moment “an innocent mistake.” Bard was trained to corroborate its answers based on Google Search results and had most likely misconstrued a NASA blog that announced the “first time” astronomers used the James Webb telescope to photograph an exoplanet. One former staffer remembers leadership reassuring the team that no one would lose their head from the incident, but that they had to learn from it, and fast. “We’re Google, we’re not a startup,” Hsiao says. “We can’t as easily say, ‘Oh, it’s just the flaw of the technology.’ We get called out, and we have to respond the way Google needs to respond.”

Googlers outside the Bard team weren’t reassured. “Dear Sundar, the Bard launch and the layoffs were rushed, botched, and myopic,” read one post on Memegen, the company’s internal messaging board, according to CNBC. “Please return to taking a long-term outlook.” Another featured an image of the Google logo inside of a dumpster fire. But in the weeks after the telescope mixup, Google doubled down on Bard. The company added hundreds more staff to the project. In the team’s Google Docs, Pichai’s headshot icon began popping up daily, far more than with past products…

…Meanwhile, GDM’s responsibility team was racing to review the product. For all its added power, Gemini still said some strange things. Ahead of launch, the team found “medical advice and harassment as policy areas with particular room for improvement,” according to a public report the company issued. Gemini also would “make ungrounded inferences” about people in images when prompted with questions like, “What level of education does this person have?” Nothing was “a showstopper,” said Dawn Bloxwich, GDM’s director of responsible development and innovation. But her team also had limited time to anticipate how the public might use the model—and what crazy raps they might try to generate.

If Google wanted to blink and pause, this was the moment…

…But despite the growing talk of p(doom) numbers, Hassabis also wanted his virtual assistant, and his cure for cancer. The company plowed ahead.

When Google unveiled Gemini in December 2023, shares lifted. The model outperformed ChatGPT in 30 of 32 standard tests. It could analyze research papers and YouTube clips, answer questions about math and law. This felt like the start of a comeback, current and former employees told WIRED. Hassabis held a small party in the London office. “I’m pretty bad at celebrations,” he recalls. “I’m always on to thinking about the next thing.”…

…One year on from the code-red moment, Google’s prospects were looking up…

…But just when Google employees might have started getting comfortable again, Pichai ordered new cutbacks. Advertising sales were accelerating but not at the pace Wall Street wanted. Among those pushed out: the privacy and compliance chiefs who oversaw some user safeguards. Their exits cemented a culture in which concerns were welcome but impeding progress was not, according to some colleagues who remained at the company.

For some employees helping Hsiao’s team on the new image generator, the changes felt overwhelming. The tool itself was easy enough to build, but stress-testing it was a game of brute-force trial and error: review as many outputs as possible, and write commands to block the worst of them. Only a small subset of employees had access to the unrestrained model for reviewing, so much of the burden of testing it fell on them…

…The image generator went live in February 2024 as part of the Gemini app. Ironically, it didn’t produce many of the obviously racist or sexist images that reviewers had feared. Instead, it had the opposite problem. When a user prompted Gemini to create “a picture of a US senator from the 1800s,” it returned images of Black women, Asian men, or a Native American woman in a feather headdress—but not a single white man. There were more disturbing images too, like Gemini’s portrayal of groups of Nazi-era German soldiers as people of color…

…The Project Magi team had designed a feature called AI Overviews, which could synthesize search results and display a summary in a box at the top of the page. Early on, responsible innovation staffers had warned of bias and accuracy issues and the ethical implications for websites that might lose search traffic. They wanted some oversight as the project progressed, but the team had been restructured and divided up.

As AI Overviews rolled out, people received some weird results. Searching “how many rocks should I eat” brought up the answer “According to UC Berkeley geologists, eating at least one small rock per day is recommended.” In another viral query, a user searched “cheese not sticking to pizza” and got this helpful tip: “add about 1/8 cup of non-toxic glue to the sauce to give it more tackiness.” The gaffes had simple explanations. Pizza glue, for example, originated from a facetious Reddit post. But AI Overviews presented the information as fact. Google temporarily cut back on showing Overviews to recalibrate them.

That not every issue was caught before launch was unfortunate but no shock, according to Pandu Nayak, Google’s chief scientist in charge of search and a 20-year company veteran. Mostly, AI Overviews worked great. Users just didn’t tend to dwell on success. “All they do is complain,” Nayak said. “The thing that we are committed to is constant improvement, because guaranteeing that you won’t have problems is just not a possibility.”…

…This past December, two years into the backlash and breakthroughs brought on by ChatGPT, Jeff Dean met us at Gradient Canopy. He was in a good mood. Just a few weeks earlier, the Gemini models had reached the top spot on a public leaderboard. (One executive told WIRED she had switched from calling her sister during her commutes to gabbing out loud with Gemini Live.) Nvidia CEO Jensen Huang had recently praised NotebookLM’s Audio Overviews on an earnings call, saying he “used the living daylights out of it.” And several prominent scientists who fled the caution-ridden Google of yesteryear had boomeranged back—including Noam Shazeer, one of the original eight transformers inventors, who had left less than three years before, in part because the company wouldn’t unleash LaMDA to the public.

As Dean sank into a couch, he acknowledged that Google had miscalculated back then. He was glad that the company had overcome its aversion to risks such as hallucinations—but new challenges awaited. Of the seven Google services with more than 2 billion monthly users, including Chrome, Gmail, and YouTube, all had begun offering features based on Gemini. Dean said that he, another colleague, and Shazeer, who all lead the model’s development together, have to juggle priorities as teams across the company demand pet capabilities…

…Google faces one challenge that its competitors don’t: In the coming years, up to a quarter of its search ad revenue could be lost to antitrust judgments, according to JP Morgan analyst Doug Anmuth. The imperative to backfill the coffers isn’t lost on anyone at the company. Some of Hsiao’s Gemini staff have worked through the winter holidays for three consecutive years to keep pace. Google cofounder Brin last month reportedly told some employees 60 hours a week of work was the “sweet spot” for productivity to win an intensifying AI race. The fear of more layoffs, more burnout, and more legal troubles runs deep among current and former employees who spoke to WIRED.

2. 10 Biggest Ideas in “How NOT to Invest” – Barry Ritholtz

1. Poor Advice: Why is there so much bad advice? The short answer is that we give too much credit to gurus who self-confidently predict the future despite overwhelming evidence that they can’t. We believe successful people in one sphere can easily transfer their skills to another – most of the time, they can’t. This is as true for professionals as it is for amateurs; it’s also true in music, film, sports, television, and economic and market forecasting…

…3. Sophistry: The Study of Bad Ideas: Investing is really the study of human decision-making. It is about the art of using imperfect information to make probabilistic assessments about an inherently unknowable future. This practice requires humility and the admission of how little we know about today and essentially nothing about tomorrow. Investing is simple but hard, and therein lies our challenge…

…7. Avoidable Mistakes: Everyone makes investing mistakes, and the wealthy and ultra-wealthy make even bigger ones. We don’t understand the relationship between risk and reward; we fail to see the benefits of diversification. Our unforced errors haunt our returns.

8. Emotional Decision-Making: We make spontaneous decisions for reasons unrelated to our portfolios. We mix politics with investing. We behave emotionally. We focus on outliers while ignoring the mundane. We exist in a happy little bubble of self-delusion, which is only popped in times of panic.

9. Cognitive Deficits: You’re human – unfortunately, that hurts your portfolio. Our brains evolved to keep us alive on the savannah, not to make risk/reward decisions in the capital markets. We are not particularly good at metacognition—the self-evaluation of our own skills. We can be misled by individuals whose skills in one area do not transfer to another. We prefer narratives over data. When facts contradict our beliefs, we tend to ignore those facts and reinforce our ideology. Our brains simply weren’t designed for this.

3. AI Boom Reshapes Power Landscape as Data Centers Drive Historic Demand Growth – Aaron Larson

Enverus, an energy-dedicated software-as-a-service (SaaS) company that leverages generative AI across its solutions, released its 2025 Global Energy Outlook in late January. Like many industry observers, Enverus predicts power demand growth fueled by the AI race will dominate the energy narrative.

“The energy narrative in 2024 shifted from focusing on the urgency of the energy transition to the urgency of energy security,” the report says. “What stands out in this evolving narrative is the role of demand, led by data center hyperscalers who appear almost agnostic to price. For this group, the energy trilemma prioritizes reliability as No. 1, environmental concerns as No. 2, cost as No. 3. This has placed the quest for 24/7 reliable baseload power at the forefront, with natural gas-fired capacity competing with nuclear and geothermal to meet the challenge.”

Enverus forecasts U.S. load to increase 1.2% in 2025 compared to 2024, and 38% by 2050…

…When Deloitte’s team publishes its annual Power and Utilities Industry Outlook around the beginning of the year, it typically tries to identify five key trends…

…“To meet the rising demand from data centers, utilities will likely continue enhancing grid efficiency, enlisting reliable and clean power sources, and implementing equitable tariffs and cost allocation through collaborative partnerships,” the Deloitte report says. Supporting that, the report says utilities are likely to continue embracing nuclear power (Figure 1); integrating distributed energy resources; adapting workforce strategies to address skills gaps; and exploring first-of-a-kind projects in carbon capture and storage, offsets, and removal strategies…

…“I’ve been in this industry a long time, and I joke that for the first 34 years of my career, every utility was basically satisfied with 2% growth, and cutting operations and maintenance costs, which combined to make the economics work,” Keefe said. “Now, some utilities are talking 100% growth in the next five years. I mean, it’s just mind-boggling that it’s changed so fast, and it seemed like it’s overnight.”…

…For its report, Enverus Intelligence Research (EIR), a subsidiary of Enverus, analyzed breakeven economics across nine technologies to assess the risk of Inflation Reduction Act (IRA) credit elimination, comparing them with and without IRA incentives against industry incumbents…

…“Across the Lower 48 [the continental U.S.], a staggering 76% and 37% of queued solar and wind capacity, respectively, are dependent on tax incentives to be economically viable,” Corianna Mah, an analyst at EIR, said.

Without subsidies, onshore wind, EOR, solar, and blue hydrogen technologies cost from 29% to 63% more than incumbents, but with incentives, costs range from a 13% premium to a 35% discount.”…

…In contrast, the PTC for green hydrogen and ITC for geothermal face higher risks for tax credit elimination, with unsubsidized breakeven premium ranges of 205% to 310%, dropping to 103% to 135% when subsidized, highlighting their limited competitiveness…

…Enverus expects markets with high battery energy storage system (BESS) adoption to see a significant transformation in battery operations. Its analysts suggested ancillary market adjustments may be needed, which could reshape revenue streams and grid dynamics. The Electric Reliability Council of Texas’ (ERCOT’s) market provides a glimpse of this evolution, with battery capacity surging 237% since early 2023…

…The report notes that ERCOT currently has 8,374 MW of operating storage capacity, with 5,201 MW under construction and 8,244 MW with signed interconnection agreements set to come online by 2025—a 160% increase over today’s already saturated levels. By 2025, EIR expects this additional capacity will heavily influence energy markets, pushing prices lower.

4. Historical analogies for large language models – Dynomight

How will large language models (LLMs) change the world?

No one knows. With such uncertainty, a good exercise is to look for historical analogies—to think about other technologies and ask what would happen if LLMs played out the same way.

I like to keep things concrete, so I’ll discuss the impact of LLMs on writing. But most of this would also apply to the impact of LLMs on other fields, as well as other AI technologies like AI art/music/video/code.

1. The ice trade and freezers

We used to harvest huge amounts of natural ice and ship them long distances. The first machines to make ice were dangerous and expensive and made lousy ice. Then the machines became good and nobody harvests natural ice anymore.

In this analogy, LLMs are bad at first and don’t have much impact. Then they improve to match and then exceed human performance and human writing mostly disappears…

…4. Horses and railroads

At first, trains increased demand for horses, because vastly more stuff was moving around over land, and horses were still needed to get stuff to and from train stations.

In this analogy, giving human writers LLMs makes them more efficient, but it doesn’t put anyone out of work. Instead, this new writing is so great that people want more of it—and more tailored to their interests. Instead of 8 million people paying $20 per month for 5000 people to create Generic Journalism Product, groups of 100 people pay $200 per month for one person to create content that’s ultra-targeted to them, and they’re thrilled to pay 10× more because it makes their lives so much better. Lots of new writers enter the market and the overall number of writers increases. Then LLMs get even better and everyone is fired…

…8. Site-built homes and pre-manufactured homes

We can build homes in factories, with all benefits of mass production. But this is only used for the lowest end of the market. Only 6% of Americans live in pre-manufactured homes and this shows no sign of changing.

In this analogy, LLMs make text cheaper. But for some reason (social? technical? regulatory?) AI writing is seen as vastly inferior and doesn’t capture a significant part of the market…

…13. Human calculators and electronic calculators

Originally a “computer” was a human who did calculations.

In this analogy, LLMs are an obvious win and everyone uses them. It’s still understood that you need to know how to write—because otherwise how could you understand what an LLM is doing? But writing manually is seen as anachronistic and ceases to exist as a profession. Still, only a tiny fraction of writing is done by “writers”, so everyone else adopts LLMs as another productivity tool, and soon we’ve forgotten that we ever needed humans to do these things…

…To predict the impact of LLMs we also need to understand:

  • Will LLMs act more as competitors or complements to human writing?
  • How will people react to LLMs? Maybe LLMs will write amazing novels and people will love them. Or, maybe, people just can’t stand the idea of reading something written by an AI.
  • If people decide they don’t like LLMs, to what degree are countermeasures possible? Can we build machine learning models to detect LLM-generated text? Will we force LLM providers to embed some analogy to yellow dots in the text? Can we create a certification process to prove that text was created by a human? (You could record a video of yourself writing the entire book, but how do you certify the video?)

Beyond all that, I wonder to what degree these analogies are useful. One big difference between writing to these other domains is that once writing is created, it can be copied at near-zero cost. The closest historical analogy for this seems to be the printing press disrupting hand copying of books, or maybe computers disrupting paper books. But it’s also possible that this shift is something fundamentally new and won’t play out like any of these analogies suggest.

5. Jim Millstein on the Massive Risks of Any ‘Mar-a-Lago Accord’ (Transcript here) – Joe Weisenthal, Tracy Alloway, and Joe Millstein

Tracy: Shall I just jump into it and ask the obvious question or one of the obvious questions. Where is this suggestion coming from a debt restructuring as part of a potential Mar-a-Lago Accord? What is the problem we’re trying to solve?

Jim: I don’t want to engage in sanewashing. There’s clearly an impetus by the President to impose tariffs. He’s tariff man, and around him, through Bessant and Miran, there is some intellectual architecture that suggests that’s just a tactic towards an end, and the end is to bring manufacturing back to the United States.

Obviously during this period of globalization, we’ve been running massive trade deficits, particularly in manufacturers, where we’re importing a number of critical systems to both our defense industry and to our manufacturing industry. We once dominated the semiconductor trade – we actually created that industry in the 1960s, through a series of government policies, research and development grants to IBM and AT&T that created the semiconductor technology. Then a series of procurement policies at NASA and the Defense Department to commercialize that industry, and eventually we created the calculator industry, and the computer industry, and the TV industry, and all of that. That was all a byproduct of a coordinated set of federal policies. Fast forward 40 years, 50 years later, and semiconductor manufacturing is mostly being done, particularly at the high end, in a strategically vulnerable country across the straits of China in Taiwan. That has created a sense, now going back 10 years, in the defense establishment that we have a problem, and not just in semiconductors but in a number of advanced industries where we’re really reliant as a country on the importation of critical technologies and critical intermediate inputs.

If you piece together some of the things that Bessant has said and some of the things that Miran has said, the goal of the tariff play, which is really just a tactic, is to bring manufacturing back to the United States to hollow-in, or build out the communities that were hollowed out by the wave of globalization that occurred after China’s admission to the WTO in the early 2000s.

One of the critical elements, or transmission mechanisms that they’re trying to affect, is the exchange rate of the dollar. A high dollar means that our exports are more expensive and our imports are less expensive. We have been the beneficiary with a strong dollar of very cheap imports, moderating the inflation that might otherwise occur from domestic manufacturing. But that said, we’ve lost manufacturing. 40 years ago, we represented 25% of the manufacturing industry. Now we’re a mere 15% of global manufacturing. China was nowhere to be seen, now they’re 35% of global manufacturing. The goal of this Mar-a-Lago accord is to really weaken the dollar without upsetting the financial flows that finance our debt…

…Joe: You use the word sanewashing, which is a good word because there’s this intellectual architecture around Trump. It’s not clear that Trump himself sees it this way, that this works, that you can re-accelerate US manufacturing simply via some weakening of the dollar in a coordinated way, or tariffs. What is the gap between what you see is actually going on and the white papers that people put out on this?

Jim: This is all coming out of Miran’s paper as Tracy indicated at the beginning. He’s put together the most comprehensive strategy, and he acknowledges there’s a very narrow corridor within which this might work. In some sense, the president has already gotten out ahead with his tariff tactics and also his threatening to withdraw the security umbrella from NATO. Those are the two critical sticks that Miran advocated we use to induce foreign central banks and foreign investors to continue to buy treasuries at favorable rates so as to continue to finance what is really a growing and potential – as Dalio said in your podcast – debt crisis.

Maybe to frame that problem, today federal debt to GDP is one-to-one. Federal debt is equal to GDP. We’re running deficits at 7% of GDP, and the economy is kind of growing at 1%, 2%, little north of 2%. So the debt is growing faster as a result of the imbalance in the federal budget where deficits are growing at the rate of 7% of GDP, which means the debt’s growing at the rate of 7% of GDP, where our debt is growing now faster than GDP and is becoming an increasing overhang to the extent that when you look at the federal budget, interest expense has become the second largest category of federal spending.

Tracy: issuing bonds to pay off bonds.

Jim: That’s right. We’re now issuing bonds to pay the interest on our bonds. This is a classic recipe for disaster. We’re not even treading water. We’re now slowly sinking under a huge pile of debt. So, we have to get that fiscal imbalance corrected. And as you were saying at the beginning of the podcast, Joe, some very tough allocation decisions need to be made with regard to federal spending because someone joked that when you look at the federal government, it’s really a retirement program attached to an army.

Joe: I’ve heard it called an insurance program, but it’s the same thing.

Jim: Yeah, exactly. You have income security in the form of social security for retirement, and you have medical security in the form of Medicare for retirement. When you add it all up, the parts of the budget that Elon Musk and his merry band of pranksters are off trying to slash, is a relatively small part of federal spending. But it is the stuff that actually supports education, transportation, housing, infrastructure. The sort of stuff that is building human capital, building physical, public capital, building housing structure. That part of the budget is a mere $700 billion out of a total spending of $6.75 trillion. The rest of it is interest on the debt, retirement security, defense, and healthcare support. So we’re really in a pickle.

We’re going to see in the Fall, or maybe sooner, when the reconciliation bills finally make their way to a vote on the floor of the House and the Senate, we’re going to see whether or not this Congress really has the courage to deal with the allocation issues that you mentioned. Because in the framework for the House Reconciliation Bill, they call for $880 billion – that’s over 10 years – so, it’s really not a lot. It’s about $100 billion of spending cuts annually in Medicaid, transportation, housing, and education. Out of that, Medicaid is about $600 billion a year and the housing, transportation, education, that part of the budget is about $700 billion. So they’re calling for a reduction of $100 billion a year against that $1.3 trillion of Medicaid and the other social spending. So it’s not a big ticket and it’s not going to make a massive change in the deficit, particularly if they add incremental tax cuts on tips, on overtime, on social security as they’ve talked about. They’re not really attacking the deficit. So we’re going to continue to need to sell a lot of debt.

Tracy: So you’ve laid out the pickle problem very well. The idea embedded in the Mar-a-Lago Accord is that the US could bring down its debt costs by getting foreign investors to swap some of their current treasuries into century bonds that would be less expensive for the US to actually pay back.

Jim: That’s right. How do we induce them to engage in that exchange? The way you do exchange offers in the private markets that I traffic in is with carrots and sticks. You offer a sweetener and you threaten doom and gloom. The two primary tactics here that you foreign country are going to face, on the one hand a high tariff wall unless you play ball, and on the other hand, the withdrawal of our security umbrella. So if you want the protection of the largest and most powerful military in the world to protect your borders against a Russian invasion, you’re going to have to swap your debt that you currently hold, which is generally short-term bills, into what they’re calling century bonds, a 100-year bond at a low interest rate, which takes the refinancing risk of indebted country away from it, because we don’t have to touch that debt for 100 years.

Tracy: Terming out duration.

Jim: Terming out duration… on the one hand, and reducing the interest burden of servicing that debt over time. There are a couple of problems with this. One problem is that when you look at who holds US government debt, not more than 15% of it today is held offshore.

Tracy: It’s come down a lot.

Jim: Yea it’s come down a lot. Much of that 15% is not in the hands of government instrumentalities, but rather in foreign private investors. So inducing that crowd to come into this exchange offer, even if you could succeed, you’re touching a very small part of the debt. So where’s the rest of it? Where’s the other 85% of our $36 trillion of outstanding debt? It’s basically owned by us. Some of it’s owned in government accounts and the Social Security and Medicare trust funds, but some of it is owned by banks and insurance companies. Some of it’s owned by endowments and wealthy individuals. Some of it’s in the bond, in the mutual fund market, underwriting our money market funds. The reality is, to get this done, we’re really doing it with ourselves.

What we really need to do is term out our debt, and the problem we’re facing right now is that the interest cost of our debt is relatively high. You know the 10-year is at 4.3%, the 30-year – put aside as to what you’d pay for a century bond – 30-year is even higher. The current average interest rate on our outstanding $36 trillion of debt is 3.3%. To term it out in this market would take that $1.1 trillion of annual interest expense up – if we had to term it out at 4.3% or 4.6%, we’d be talking about increasing the interest expense we’re facing. This intellectual architecture around the so-called Mar-a-Lago Accord has many flaws, not least among which is, in targeting foreign holders of our debt, we’re targeting a relatively small part of it.

If the game plan here of that Mara Lago accord is to weaken the dollar so as to improve the competitiveness of US domestic manufacturing, there is another approach. That you’ve also heard rumor of from the Trump Administration, and that is the creation of a sovereign wealth fund, to take assets that the US government currently owns, dump them in a central fund managed by the Treasury Department and allowing the Treasury Department then to intervene directly into the foreign exchange markets to try and push the dollar down…

… Joe: But when you get into existential questions about the safety and risk-freeness of US debt, what are we talking about here?

Jim: Once we went off the gold standard, once our currency and our debt was not convertible into gold, into a hard commodity, the reliability of the US government debt is really a bet on the US government, that the economy is going to be so strong and generate the capacity to pay taxes to support the repayment of the debt. So these two things now, it’s a confidence game and they’re intricately linked. The dynamism of the US economy is ultimately what supports the creditworthiness of the debt. But as your debt – and this is what Dalio was talking about – as your debt levels increase to the point where your ability to service the debt is called into question, or your ability to service the debt is squeezing out the role that the government plays in buttressing, undergirling the dynamism of the economy, you get to a point where investors start to worry about the durability of the debt, the ability of the government to pay the debt. So the debt overhang itself becomes a retardant to economic growth and if the dynamism of the economy is what undergirds people’s confidence in our ability to repay our debts when due, we’re in a world of hurt…

…Tracy: We’ve talked a lot about creative ways for the US government to raise money and pay off its debt. There’s one we haven’t talked about, which is one of my favorite financial topics of all time, and that is the bonds owned by the US issued by other countries, really old ones, like Chinese imperial debt. Did you know the UK owes the US a lot of money from World War II loans?

Jim: Oh, still, I didn’t know that.

Joe: I didn’t know that.

Tracy: As an intellectual curiosity, I find it really interesting to think about the question of what would happen if Trump decided to go after those as a way of raising money. This actually came up in the first Trump Administration. The Treasury was looking at ways to get a payout on the Chinese bonds. And funnily enough, it was doing that at the same time that the SEC was prosecuting someone for selling those bonds to investors and promising a payout. That’s fun. That could be fun.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have a vested interest in Alphabet (parent of Google) and Microsoft. Holdings are subject to change at any time.

More Of The Latest Thoughts From American Technology Companies On AI (2024 Q4)

A collection of quotes on artificial intelligence, or AI, from the management teams of US-listed technology companies in the 2024 Q4 earnings season.

Earlier this month, I published the two-part article, The Latest Thoughts From American Technology Companies On AI (2024 Q4) (see here and here). In them, I shared commentary in earnings conference calls for the fourth quarter of 2024, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. 

A few more technology companies I’m watching hosted earnings conference calls for 2024’s fourth quarter after I prepared the article. The leaders of these companies also had insights on AI that I think would be useful to share. This is an ongoing series. For the older commentary:

Here they are, in no particular order:

Adobe (NASDAQ: ADBE)

Adobe’s management will be offering new Firefly web app subscriptions that will support both Adobe’s Firefly AI models and 3rd-party models; management envisions the Firefly app as the umbrella destination for ideation; management recently introduced Adobe’s new Firefly video model into the Firefly app offering; management will be introducing Creative Cloud offerings with Firefly tiering; the Firefly video model has been very well-received by brands and creative professionals; users of the Firefly video model can generate video clips from a text prompt or image; the Firefly web app allows users to generate videos from key frames, use 3D designs to precisely direct generations, and translate audio and video into multiple languages; the Firefly web app subscription plans include Firefly Standard, Firefly Pro, and Firefly Premium; more than 90% of paid users of the Firefly web app have been generating videos; Firefly has powered 20 billion generations (16 billion in 2024 Q3) since its launch in March 2023, and is now doing more than 1 billion generations a month; management thinks the commercially-safe aspect of Firefly models is very important to users; management thinks the high-level of creative control users get with Firefly models is very important to them; the adoption rates of the Firefly paid plan signals to management that Firefly is adding value to creative professionals

In addition to Creative Cloud, we will offer new Firefly web app subscriptions that integrate and are an on-ramp for our web and mobile products. While Adobe’s commercially safe Firefly models will be integral to this offering, we will support additional third-party models to be part of this creative process. The Firefly app will be the umbrella destination for new creative categories like ideation. We recently introduced and incorporated our new Firefly video model into this offering, adding to the already supported image, vector and design models. In addition to monetizing stand-alone subscriptions for Firefly, we will introduce multiple Creative Cloud offerings that include Firefly tiering…

…The release of the Adobe Firefly Video model in February, a commercially-safe generative AI video model, has been very positively received by brands and creative professionals who have already started using it to create production-ready content. Users can generate video clips from a text prompt or image, use camera angles to control shots, create distinct scenes with 3D sketches, craft atmospheric elements and develop custom motion design elements. We’re thrilled to see creative professionals and enterprises and agencies, including Dentsu, PepsiCo and Stagwell finding success with the video model….

…In addition to generating images, videos and designs from text, the app lets you generate videos from key frames, use 3D designs to precisely direct generations, and translate audio and video into multiple languages. We also launched 2 new plans as part of this release, Firefly Standard and Firefly Pro and began the rollout of our third plan, Firefly Premium, yesterday. User engagement has been strong with over 90% of paid users generating videos…

…Users have generated over 20 billion assets with Firefly…

…We’re doing more than 1 billion generations now a month and 90% of people using Firefly the app also saw — are generating video as well as part of that…

…For Firefly, we have imaging, vector, design, video, voice, video and voice coming out just a couple of weeks ago, off to a good start. I know there have been some questions about how important is commercially safety of the models. They’re very important. A lot of enterprises are turning to them for the quality, the breadth but also the commercial safety, the creative control that we give them around being able to really match structure, style, set key frames for precise video generation, 3D to image, image to video…

…If we look at the early adoption rates of the Firefly paid plan, it really tells us both of these stories. We have a high degree of conviction that it’s adding value and being used by Creative Professionals,

Adobe’s management thinks that marketing professionals will need to create and deliver an unprecedented volume of personalised content and that marketing professionals will need custom, commercially safe AI models and AI agents to achieve this, and this is where Adobe GenStudio and Firefly Services can play important roles; management is seeing customers turn to Firefly Services and Custom Models for scaling on-brand marketing content production; there are over 1,400 custom models created since launch of Firefly Services and Custom Models; Adobe GenStudio for Performance Marketing has won leading brands recently as customers; Adobe GenStudio for Performance Marketing has partnerships with leading digital advertising companies

Marketing professionals need to create an unprecedented volume of compelling content and optimize it to deliver personalized digital experiences across channels, including mobile applications, e-mail, social media and advertising platforms. They’re looking for agility and self-service as well as integrated workflows with their creative teams and agencies. To achieve this, enterprises require custom, commercially safe models and agents tailored to address the inefficiencies of the content supply chain. With Adobe GenStudio and Firefly Services, Adobe is transforming how brands and their agency partners collaborate on marketing campaigns, unlocking new levels of creativity, personalization and efficiency. The combination of the Adobe Experience Platform and apps and Adobe GenStudio is the most comprehensive marketing platform to deliver on this vision…

…We had another great quarter in the enterprise with more customers turning to Firefly Services and Custom Models to scale on-brand content production for marketing use cases, including leading brands such as Deloitte Digital, IBM, IPG Health, Mattel and Tapestry. Tapestry, for example, has implemented a new and highly productive digital twin workflow using Custom Models and Firefly…

…Strong demand for Firefly Services and Custom Models as part of the GenStudio solution with over 1,400 custom models since launch.

GenStudio for Performance Marketing wins at leading brands including AT&T, Lennar, Lenovo, Lumen, Nebraska Furniture Mart, Red Hat, Thai Airways, and University of Phoenix.

Strong partnership momentum with GenStudio for Performance Marketing supporting ad creation and activation for Google, Meta, Microsoft Ads, Snap, and TikTok and several partners including Accenture, EY, IPG, Merkle and PWC offering vertical extension apps.

Adobe’s generative AI solutions are infused across the company’s products and management sees the generative AI solutions as a factor driving billions in annualised recurring revenue (ARR) for the company from customer acquisition to customer retention and upselling; Adobe has AI-first stand-alone and add-on products such as Acrobat AI Assistant, the Firefly App and Services, and GenStudio for Performance Marketing; the AI-first stand-alone and add-on products already accounted for $125 million in book of business for Adobe in 2024 Q4 (FY2025 Q1), and management expects this book of business to double by the end of FY2025; management thinks that the monetisation of Adobe’s AI services goes beyond the $125 million in book of business and also incorporates customers who subscribe to Adobe’s services and use the AI features

Our generative AI innovation is infused across the breadth of our products, and its impact is influencing billions of ARR across acquisition, retention and value expansion as customers benefit from these new capabilities. This strength is also reflected in our AI-first stand-alone and add-on products such as Acrobat AI Assistant, Firefly App and Services and GenStudio for Performance Marketing, which have already contributed greater than $125 million book of business exiting Q1 fiscal ’25. And we expect this AI book of business to double by the end of fiscal ’25…

…A significant amount of the AI monetization is also happening in terms of attracting people to our subscription, making sure they are retained and having them drive higher-value price SKUs. So when somebody buys Creative Cloud or when somebody buys Document Cloud, in effect, they are actually monetizing AI. But in addition to that, Brent, what we wanted to do was give you a flavor for the new stand-alone products that we have when we’ve talked about introducing Acrobat AI Assistant and rolling that out in different languages, Firefly, and making sure that we have a new subscription model associated with that on the web, Firefly Services for the enterprise and GenStudio. So the $125 million book of business that we talked about exiting Q1 only relates to that new book of business.

Adobe’s management is seeing every CMO (Chief Marketing Officer) being very interested in using generative AI in their content supply chain

Every CMO that we talk to, every agency that we work with, they’re all very interested in how generative AI can be used to transform how the content supply chain works.

Adobe’s management sees AI as bringing an even larger opportunity for Adobe

I am more excited about the larger opportunity without a doubt as a result of AI. And we’ve talked about this, Kash. If you don’t take advantage of AI, it’s a disruption. In our particular case, the intent is clearly to show how it’s a tailwind.

Adobe’s management is happy to support 3rd-party models within the Firefly web app or within other Adobe products so long as the models deliver value to users

We’ll support all of the creative third-party models that people want to support, whether it’s a custom model we create for them or whether it’s any other third-party model within Firefly as an app and within Photoshop, you’re going to see support for that as well. And so think of it as we are the way in which those models actually deliver value to a user. And so it’s actually just like we did with Photoshop plug-ins in the past, you’re going to see those models supported within our flagship applications.

Adobe’s management is seeing very strong attach rate and adoption of generative AI features in Adobe’s products with creative professionals

This cohort of Creative Professionals, we see very strong attach and adoption of the generative AI features we put in the product partially because they’re well integrated and very discoverable and because they just work and people get a lot of value out of that. So what you will see is you’ll start to see us integrating these new capabilities, these premium capabilities that are in the Firefly Standard, Pro and Premium plans more deeply into the creative workflow so more people have the opportunity to discover them.

Meituan (OTC: MPNGY)

Meituan’s autonomous vehicles and drones have accumulated 4.9 million and 1.45 million in orders-fulfilled by end-2024; Meituan’s drones started operating in Dubai recently

By year end of 2024, the accumulated number of commercial orders fulfilled by our autonomous vehicles and drones have reached 4.9 million and 1.45 million, respectively. Our drone business also started commercial operation in Dubai recently.

Meituan’s management wants to expand Meituan’s investments in AI, and is fully committed to integrating AI into Meituan’s platform; management’s AI strategy for Meituan has 3 layers, which are (1) integrating AI into employees’ work, (2) infusing AI into Meituan’s products, and (3) building Meituan’s own large language model

We will actively embrace and expand investment in cutting-edge technologies, such as AI or unmanned aerial delivery or autonomous delivery service vehicles, and accelerate the application of these technologies. And we are committed to fully integrating AI into consumers’ daily lives and help people eat better, live better…

…Our AI strategy builds upon 3 layers. The first one is AI at work. We are integrating AI in our employees’ day-to-day work and our daily business operations and to significantly enhance the productivity and work experience for our over 400,000 employees. And then second layer is AI in products. So we will use AI to upgrade our existing products and services, both 2B and 2C. And we will also launch brand-new AI-native products to better serve our consumers, merchants, couriers and business partners…

…The third layer is building our own in-house large language model, and we plan to continue to invest and enhance our in-house large language model with increased CapEx.

Meituan’s management has developed Meituan’s in-house large language model named Longcat; management has rolled out Longcat alongside 3rd-party models to improve employees’ productivity; Longcat has been useful for AI coding, conducting smart meetings, short-form video generation, for AI sales assistance, and more; Longcat has been used to develop an in-house AI customer service agent, which has driven a 20% improvement in efficiency and a 7.5 percentage points improvement in customer satisfaction; the AI sales assistant reduced the workload of Meituan’s business development (BD) team by 44% during the Spring Festival holidays; 27% of new code in Meituan is currently generated by its AI coding tools

On the first layer, AI at work, on the employee productivity front, we have our — we have developed our in-house large language model. It’s called longcat. By putting longcat side by side with external models, we have rolled out our very highly efficient tools for our employees, including AI coding, smart meeting and document assistant, and also, it’s quite useful in graphic design and short-form video generation and also AI sales assistance. These tools have substantially boost employee productivity and working experience…

…We have developed an intelligent AI customer service agent using our in-house large language model. So after the pilot operation, the results show more than 20% enhanced efficiency. And moreover, the customer satisfaction rate has improved over 7.5 percentage points…

…During this year’s Spring Festival holidays, we gathered an updated business information of our 1.2 million merchants on our platform with AI sales assistant. So it very effectively reduced the workload of our BD team, yes, by 44% and further enhanced the accuracy of the listed merchant information on our platform…

…Right now, in our company, about 27% of new code is generated by AI coding tools.

Meituan’s management is using AI to help merchants with online store design, information enhancement, and display and operation management; management is testing an AI assistant to improve the consumer experience in their search and transactions; management will launch a brand-new advanced AI assistant later this year that will give everyone a free personal assistant; the upcoming advanced AI assistant will be able to satisfy a lot of consumer-needs in the physical world because in order to bring AI to the physical world, physical infrastructure is needed and Meituan has that

We use AI across multiple categories by providing various tools such as smart online store design and smart merchant information enhancement and display and operation management…

…On the consumer side, we have already started testing AI assistant in some categories to enhance customer — consumer experience for their search and transaction on our platform. And for example, we have rolled out a restaurant assistant and travel assistant — reservation assistant. They can chat with the users, either by text or voice, making things more convenient and easier to use for users. And right now, we are already working on a brand-new AI native product. We expect to launch this more advanced AI assistant later this year and to cover all Meituan services so that everyone can have a free personal assistant. So based on our rich off-line service offerings and efficient on-demand delivery network, I think we will be able to handle many personalized needs in local services. And whether it’s ordering food delivery or making a restaurant reservation or purchasing group deals or ordering groceries or planning trips or booking hotels, I think we have got it covered with a one-stop, and we are going to deliver it to you on time…

…Our AI assistant will not only offer consumer services in the digital world, not just a chatbot, but it’s going to be able to satisfy a lot of their needs in the physical world because in order to bring AI to the physical world, you need more than just very smart algorithms or models. You need infrastructure in the physical world, and that’s our advantage…

…We have built a big infrastructure in the physical world with digital connections. We believe that, that kind of infrastructure is going to be very valuable when we are moving to the era of physical AI.

Meituan’s management expects to incur a lot of capex to improve Meituan’s in-house large language model, Longcat; to develop Longcat, management made the procurement of GPUs in 2024 a top priority, and expects to further scale GPU-related capital expenditure in 2025; Longcat has quite good evaluation results in China; Longcat’s API core volume has increased from 10% at the beginning of 2024 to 68% currently

On the algorithm model and compute side, it’s going to need a lot of CapEx and a very good foundation model. So in the past year, to ensure adequate supply of GPU resources has been a top priority for us. And even as we allocate meaningful resources in shareholder return and new initiatives, we keep investing billions in GPU resources. So our capital — CapEx this year has been substantial. And this year, we plan to further scale our investment in this very critical area. And thanks to our infrastructure and large language model team, we have made significant optimization, both in efficiency and effectiveness. And as a result, our in-house large language model, longcat, has achieved quite good evaluation results comparable to the top-tier models in China…

…The API core volume for Longcat has increased from 10% at the beginning of last year to 68% currently, so — which further validates the effectiveness of our in-house foundation model.

Meituan’s management believes that AI is going to give a massive push to the robotics industry; Meituan has been researching autonomous vehicles since 2016 and drones since 2017; management has made several investments in leading robotics and autonomous driving start-ups; management expects Meituan’s efforts in robotics and AI to be even more tightly integrated in the future

I think AI is going to give a massive push to the development of robotics. So we have been a very early mover when it comes to autonomous delivery vehicles and drones. So actually, we started our R&D in autonomous vehicles in late ’26 (sic) [ late ’16 ]. And we started our R&D in drones in 2017. So we have been working on this for many years, and we are making very good progress. So right now, we are looking to ways to apply AI in the on-demand delivery field. So apart from our in-house research — in-house R&D, we have also made quite several investments in leading start-ups in the robotics and autonomous driving sector to support their growth…

…In future, our robotics and AI will be even more tightly integrated, and we will keep improving in the areas such as autonomous delivery and logistics and automations because right now, apart — besides the last-mile delivery of on-demand delivery, we also operate a lot of rather big warehouses, and that will be very good use cases for automation technologies.

MongoDB (NASDAQ: MDB)

MongoDB’s management expects customers to start building AI prototypes and AI apps in production in 2025 (FY2026), but management expects the progress to be gradual, and so MongoDB’s business will only benefit modestly from AI in 2025 (FY2026); there are high-profile AI companies building on top of MongoDB Atlas, but in general, customers’ journeys with building AI applications will be gradual; management thinks that customers are slow in building AI applications because they lack AI skills and because there are still questions on the trustworthiness of AI applications; management sees the AI applications of today as being fairly simplistic, but thinks that AI applications will become more sophisticated as people become more comfortable with the technology

In fiscal ’26, we expect our customers will continue on their AI journey from experimenting with new technology stacks to building prototypes to join apps and production. We expect the progress to remain gradual as most enterprise customers are still developing in-house skills to leverage AI effectively. Consequently, we expect the benefits of AI to be only modestly incremental to revenue growth in fiscal ’26…

…We have some high-profile AI companies who are building on top of Atlas. I’m not at liberty to name who they are, but in general, I would say that the journey for customers is going to be gradual. I would say one is a lack of AI skills in their organizations. They really don’t have a lot of experience and it’s compounded by the rapid evolution of AI technology that they feel like it’s very hard for them to kind of think about like what’s stack to use and so on and so forth. The second, as I mentioned earlier, on the Voyage question, there’s also a real worry about the trustworthiness of a lot of these applications. So I would say the use cases you’re seeing are fairly simplistic — customer chat bots, maybe document summarization, maybe some very simple [indiscernible] workflows. But I do think that, that is we are in the early innings, and I expect a sophistication to increase as people get more and more comfortable,

In 2024 (FY2025), MongoDB started demonstrating that the modernisation of the technology-stack for applications (i.e. MongoDB’s Relational Migrator service) can be reduced with the help of AI tools; management will expand customer engagement for the modernisation so that it can contribute meaningfully to MongoDB’s business in 2026 (FY2027) and beyond; management will start with Java apps that run on Oracle; management sees a significant revenue opportunity in the modernisation of apps; MongoDB has successfully modernised financial applications for one of Europe’s largest ISVs (independent software vendors); management is even more confident of Relational Migrator now than in the past; Relational Migrator is tackling a very tough problem because it involves massive legacy code, and the use of AI in deciphering the code is very helpful; management is seeing a lot of interest from customers for Relational Migrator because the customers are in pain from their technical debt, and their legacy technology stack cannot handle AI applications

In fiscal ’25, our pilots demonstrated that AI tooling combined with services can reduce the cycle time of modernization. This year, we’ll expand our customer engagements so that app monetization can meaningfully contribute to our new business growth in fiscal ’27 and beyond. To start with, and based on customer demand, we are specifically targeting Java apps running on Oracle, which often have thousands of complex store procedures that need to be understood, converted and tested to successfully monetize the application. We addressed this through a combination of AI tools and agents along with inspection verification by delivery teams. Though the complexity of this work is high, the revenue opportunity for modernizing those applications is significant. For example, we successfully modernize our financial application for one of the largest ISVs in Europe, and we’re now in talks to modernize the majority of the legacy estate…

…[Question] What sort of momentum have you seen with relational migrator. And maybe how should we be thinking about that as a growth driver going forward?

[Answer] Our confidence and bullish on the space is even higher today than it was before…

…When you’re looking at a legacy app that’s got hundreds — tens of thousands, if not thousands, not tens of thousands of store procedures being able to reason about that code, being able to decipher that code and then ultimately to convert that code takes — is a lot of effort. And — but the good news is that we are seeing a lot of progress in that area. We see a lot of interest from our customers in this area because they are in so much pain with all the technical debt they’ve assumed. Second is that when they think about the future and how they enable AI in these applications, there’s no way they can do this on their legacy platforms. And so they’re motivated to try and modernize as quickly as possible.

MongoDB’s management sees AI transforming software from a static tool into a decision-making partner, but the rate of change is governed by the quality of the software’s data infrastructure; legacy databases cannot keep up with the requirements of AI and this is where MongoDB’s document-model database is advantageous; MongoDB’s database simplifies AI development by providing an all-in-one solution incorporating all the necessary pieces, including an operational data store, a vector database, and embedding and reranking models; MongoDB’s database provides developers with a structured approach when they are building AI applications; management sees AI applications being much better than traditional software in scenarios that require nuanced understanding, sophisticated reasoning and interaction and natural language

AI is transforming software from a static tool into a dynamic decision-making partner. No longer limited to predefined tasks, AI-powered applications will continuously learn from real-time data, but this software can only adapt as fast as the data infrastructure is built on and legacy systems simply cannot keep up. Legacy technology stacks were not designed for continuous adaptation. Complex architectures, batch processing and rigid data models create friction at every step, slowing development, limiting organization’s ability to act quickly and making even small updates time consuming and risky. AI will only magnify these challenges. MongoDB was built for change. MongoDB was designed from the outset to remove the constraints of legacy databases, enabling businesses to scale, adapt and innovate at AI speed. Our flexible document model handles all types of data while seamless scalability ensures high performance for unpredictable workloads…

…We also simplify AI development by natively including vector and tech search directly in the database providing a seamless developer experience that reduces cognitive load, system complexity, risk and operational overhead, all with the transactional, operational and security benefits intrinsic to MongoDB. But technology alone isn’t enough. MongoDB provides a structured solution-oriented approach that addresses the challenges customers have with the rapid evolution of AI technology, high complexity and a lack of in-house skills. We are focused on helping customers move from AI experimentation to production faster with best practices that reduce risk and maximize impact…

…AI-powered applications excel where traditional software often falls short, particularly in scenarios that require nuanced understanding, sophisticated reasoning and interaction and natural language…

…MongoDB demarcatizes the process of building trustworthy AI applications right out of the box. Instead of cobbling together all the necessary piece parts and operational data store, a vector database and embedding and reranking models, MongoDB delivers all of it with a compelling developer experience…

…We think architecturally, we have a huge advantage of the competition. One, the document model really supports different types of data structured, semi-structured and unstructured. We embed a search and Vector Search onto a platform. No one else does that. Then we’ve now with the Voyage AI, we have the most accurate embedding and reranking models to really address the quality and trust issue. And all this is going to be put together in a very elegant developer experience that reduces friction and enables them to move fast.

MongoDB acquired Voyage AI for $220 million, $200 million of which was paid in MongoDB shares; Voyage AI helps MongoDB’s database solve the hallucination issue – a big problem with AI applications – and make AI applications more trustworthy; management thinks the best way to ensure accurate results with AI applications is through high-quality data retrieval, and high-quality data retrieval is enabled by vector embedding and reranking models; Voyage AI’s vector embedding and reranking models have excellent ratings in the Hugging Face community and are used by important AI companies; Voyage AI has an excellent AI team; through Voyage AI, MongoDB can offer best-in-class embedding and reranking models; ISVs (independent software vendors) have gotten better performance when they switched from other embedding models to Voyage AI’s models; Voyage AI’s models increase the trustworthiness of the most demanding and mission-critical AI applications; Voyage AI’s models will only be available on Atlas

With the Voyage AI acquisition, MongoDB makes AI applications more trustworthy by pairing real-time data and sophisticated embedding and retreatment models that ensure accurate and relevant results…

…Our decision to acquire Voyage AI addresses one of the biggest problems customers have when building and deploying AI applications, the risk of hallucinations…

…The best way to ensure accurate results is through high-quality data retrieval, which shows that not only the most relevant information is extracted from an organization’s data with precision, high-quality retrieval is enabled by vector embedding and reranking models. Voyage AI has embedding and reranking models and are among the highest rated in the Hugging Face community for retrieval, classification, clustering and reranking and are used by AI leaders like Anthropic, LangChain, Harvey and Replit. Voyage AI led by Stanford professor, Tang Yuma, who has assembled a world-class AI research team from AI Labs at Stanford, MIT, Berkeley and Princeton. With this acquisition, MongoDB will offer best-in-class embedding and reranking models to power native AI retrievable…

…Let me address how the acquisition of Voyage AI will impact our financials. We disclosed last week that the total consideration was $220 million. Most Voyage shareholders received a consideration in MongoDB stock with only $20 million being paid out in cash…

…We know a lot of ISVs have already reached out to us since the acquisition saying they switched to Voyage from other model providers and they got far better performance. So the value of Voyage is being able to increase the quality and hence the trustworthiness of these AI applications that people are building in order to serve the most demanding and mission-critical use cases…

…Some of these new capabilities like Voyage now that will be available only on Atlas.

Swisscom was able to deploy a generative AI application in just 12 weeks using MongoDB Atlas

Swisscom, Switzerland’s leading provider of mobile, Internet and TV services deployed in new GenAI app in just 12 weeks using Atlas. Swisscom implemented Atlas to power a RAG application for the East Foresight library transforming unstructured data such as reports, recordings and graphics into vector bettings that large language models can interpret. This enables Vector Search to find any relevant contact resulting in more accurate and tailored responses for users.

If an LLM (large language model) is a brain, a database is memory, and embedding models are a way to find the right information for the right question; embedding models provide significant performance gains when used with LLMs

So think about the LLM as the brain. Think about the database is about your memory and the state of where how things are. And so — and then think about embedding as an ability to find the right information for the right question. So imagine you have a very smart person, say, like Albert Einstein on your staff and you’re asking him, in this case, the LLM, a particular question. While Einstein still needs to go do some homework based on what the question is about finding some information before he can formulate an answer. Rather than reading every book in a library, what the embedding models do is essentially act like a library and pointing Einstein to the right section, the right aisle, the right shelf, the right book and the right chapter on the right page, to get the exact information to formulate an accurate and high-quality response. So the performance gains you get a leveraging embedding models is significant.

Okta (NASDAQ: OKTA)

The emergence of AI agents has contributed to the growing importance to secure identity; management will provide access to Auth For GenAI on the Auth0 platform in March 2025; 200-plus startups and large enterprises are on the waitlist for Auth For GenAI; Auth For GenAI allows AI agents to securely call APIs; management is seeing that companies are trying to build agentic systems, only to run into problems with giving these agents access to systems securely; within AI, management sees agentic AI as the most applicable for Okta’s business in the medium term

With the steady rise of cloud adoption, machine identities and now AI agents, there has never been a more critical time to secure identity…

…On the Auth0 platform, we announced Auth For GenAI. We’ll begin early access this month. We already have a wait list of eager customers ranging from early startups to Fortune 100 organizations. Auth for GenAI is developed to help customers securely build and scale their Gen AI applications. This suite of features allows AI agents to securely call APIs on behalf of users while enforcing the right level of access to sensitive information…

…People are trying to stitch together agentic platforms and write their own agentic systems and what they run smack into is, wait a minute. How am I going to get these agents access all these systems if I don’t even know what’s in these systems and I don’t even know the access permissions that are there and how to securely authenticate them, so that’s driving the business…

…I’ll focus on the agentic part of AI. That’s probably the most, in the medium term, that’s probably the most applicable to our business…

…On the agent side, the equivalent of a lot of these deployments have like passwords hardcoded in the agent. So if that agent gets compromised, it’s the equivalent of your monitor having a bunch of sticky notes on it with your passwords before single sign-on. So Auth for GenAI gives you a protocol in a way to do that securely. So you can store these tokens and have these tokens that are secured. And then if that agent needs to pop out and get some approval from the user, Auth for GenAI supports that. So you can get a step-up biometric authentication from the user and say, “Hey, I want to check Jonathan’s fingerprint to make sure before I book this trip or I spend this money, it’s really Jonathan.” So those 3 parts are what Auth for GenAI is, and we’re super, super excited about it. We have a waitlist. Over 200-plus Fortune 100s and startups are on that thing.

Okta’s management thinks agentic AI is a real phenomenon and will turbocharge machine identity for Okta by 2 orders of magnitude higher; already today, a good part of Okta’s business is providing machine identity; management is the most excited about the customer identity part of Okta’s business when it comes to agentic AI because companies will start having agentic AIs as customers too; management thinks Okta will be monetise agentic AI from both people building agents, and people using agents

The agenetic revolution is real, and the power of AI and the power of these language models, the interaction modalities that you can have with these systems these machines doing things on your path and what they can do and how they can infer next actions, et cetera, et cetera. You all know it’s really real. But the way to think about it from an Okta perspective, it is like machine identity on steroids, turbocharged to like 2 orders of magnitude higher. So that’s like really exciting for us because what do we do. A good part of our business is actually logging in machines right now. Auth0 has the machine-to-machine tokens where people, if they build some kind of web app that services other machines, they can use Auth0 for the login for that. Okta has similar capabilities. And now you have not only that basic authentication challenge but you have the — all of these applications as you get 2 orders of magnitude, more things logging in, you have to really worry about the fine grain authorization into your services…

…[Question] Which side of the business are you more excited about from an agentic AI perspective?

[Answer] I think the customer identity side is more exciting. I think it’s a little bit of a — my answer is a little bit of a — I’m kind of like having both ways because a lot of the — when you talk about developers building agentic AI, they’re doing it inside of enterprises. So like the pattern I was talking about earlier, there’s these teams and these companies that have been tasked with we hear about this [ agent ] and make it work. And the first thing they have to do is I’ve had many conversations with customers where they’ve been in these discussions and we want — we did a POC and now we’re worried about doing it broadly, but the task was basically hook everything up to our existing — hook these agents up to all of our existing systems. And before we could do that inside of enterprise, we had to get a good identity foundation in front of all these things. And so it’s kind of like similar to your building something and you’re a developer, you’re exposing APIs, you’re doing fine grain authorization. You’re taking another — you’re using another platform or you’re building your own agentic AI platform, and you’re having to talk to those systems and those APIs to do things on user’s behalf, so you’re a developer, but it’s kind of like a workforce use case, but I think people building these systems and getting the benefit from that is really exciting…

…We can monetize it on “both side”, meaning people building the agents and people using the agents. The agents have to log in and they have to log into something. So I think it’s potential to monetize it on both sides.

Okta’s management thinks the software industry does not yet know how to account for AI agents in software deals; management thinks that companies will eventually be buying software licenses for both people and AI agents

One of the things that we don’t have today is the industry doesn’t have a way to like identify an agent. I don’t mean in the sense of like authenticating or validated agent. I mean to actually a universal vernacular for how to record an agent, how to track it and how to account for it. And so I think that’s something you’ll see coming. You’ll see there will be actually a type of account, an Okta that’s an agent account. You’ll see companies starting to — when they buy software, they say, hey, I buy these many people and these many agentic licenses. And that’s not quite there yet. Of course, platforms that are coming out with agent versions have this to some degree, but there isn’t a common cross-company, cross enterprise definition of an agent, which is an interesting opportunity for us actually.

Sea Ltd (NYSE: SE)

Sea’s management is using AI in Shopee to understand shoppers’ queries better and to help sellers enhance product listings, and these AI initiatives have improved purchase conversion rates and sellers’ willingness to spend on advertising; management has upgraded Shopee’s chatbots with AI and this led to meaningful improvement in customer service satisfaction score and customer service cost-per-contact; management is using AI to improve the shopper return-refund process and has seen a 40% year-on-year decrease in resolution times in Asia markets; management thinks Shopee is still early in the AI adoption curve

We continue to adopt AI to improve service quality in a practical and effective manner. By using large language models to understand queries, we have made search and discovery more accurate, helping users find relevant products faster. We provide our sellers with AI tools to enhance product listings by improving descriptions, images, and videos. These initiatives have improved purchase conversion rates while also making sellers more willing to spend on ads, boosting our ad revenue…

… After upgrading our chatbots with AI, we saw a meaningful increase in our customer service satisfaction score over the past year, and a reduction in our customer service cost-per-contact by nearly 30% year-on-year. We also used large language model capabilities to enhance our buyer return-refund process, addressing a key e-commerce pain-point. In the fourth quarter, we improved resolution times in our Asia markets by more than 40% year-on-year, with nearly six in ten cases resolved within one day. We believe we are still early in the AI adoption curve and remain committed to exploring AI-driven innovations to improve efficiency and deliver better experiences for our users.

Sea’s management thinks the use of AI is helping Sea both monetise its services better, and save costs

[Question] I just wanted to get some color with regard to the benefit from AI. Are we actually seeing cost efficiency, i.e., the use of AI actually save a lot of the manual labor cost? So that helps to achieve a lot of cost savings? Or are we actually seeing the monetization is getting better coming from AI?

[Answer] We are seeing both, in fact, for example, in our search and recommendations, we actually use the large language model to better understand user queries, making certain discovery a lot more accurate and helping users find relevant faster… We are also using the AI to understand the product a lot better like historically, it was a fintech matching, but now we can use existing pictures and the descriptions and the reviews to generate a lot more richer understanding of the product. And all those help us essentially matching, our product users’ intention a lot better. 

We are also having a lot of AIGC, AI-generated content in our platform. We provide that as a tool to our sellers to be able to produce image, a description of the product or the videos, especially a lot better compared to what they had before.

And both of this increased our conversions meaningfully in our platform.

On the other side, on the cost savings side, I think in Forrest’s opening, we talked about the chatbot, the — if you look at our queries, about 80% of the queries are answered by the chatbot already, which is a meaningful cost savings for the — for our operations. I think that’s also why you can see that our cost management for e-commerce is doing quite well. Even for the 20% answered by the agent, we have an AI tool for the agent to be able to understand the context a lot better, so can help them to respond a lot faster to the customers,

Tencent (OTC: TCEHY)

Tencent’s AI initiatives can be traced to 2016; management has been investing in Tencent’s proprietary foundation model, HunYuan, since early 2023; management sees HunYuan as the foundation for Tencent’s consumer and enterprise businesses

Our AI initiatives really trace back to 2016 when we first established our AI lab. Since 2023, early part of that, we have been investing heavily in our proprietary HunYuan foundation model, which forms an important technology foundation for our consumer and enterprise-facing businesses and will serve as a growth driver for us in the long run. Our investments in HunYuan enable us to develop end-to-end foundation model capabilities in terms of infrastructure, algorithm, training, alignment and data management and also to tailor solutions for the different needs of internal and external use cases.

Tencent’s management has released multimodal HunYuan foundation models across image, video, and 3D generation; the multimodal HunYuan foundation models have received excellent scores in AI benchmarking

In addition to LLMs, we have released multimodal HunYuan foundation models with capabilities that span across image, video and 3D generation. HunYuan’s image generation models achieved the highest score from FlagEval in December of last year. In video generation, our model excels in video output quality and ranked first on Hugging Face in December of last year. 

Tencent’s management has been actively releasing Tencent’s AI models to the open source community

Our 3D generation model was the industry’s first open source model supporting text and image to 3D generation. In addition to that, we also contribute to the open source community actively and have open sourced a series of advanced models in the HunYuan family for 3D generation, video generation, large language and image generation. Several of these models have gained great popularity among developers worldwide.

For Tencent’s consumer-facing AI products, management has been utilising different AI models because they believe that a combination of models can handle complex tasks better than a single model; Tencent’s native AI application, Yuanbao, provides access to multiple models; Yuanbao’s DAU (daily active users) increased 20-fold from February 2025 to March 2025; management has been testing AI features in Weixin to improve the user experience and will be adding more AI features over time; management will be introducing a lot more consumer-facing AI applications in the future; management thinks consumer AI is in a very early stage, but they can see Yuanbao becoming a strong AI native assistant helping with deep research, and the Ema Copilot being a personal and collaborative library; management is looking to infuse AI into each of Tencent’s existing consumer products

Going to our consumer-facing AI products. We adopt a multimodal strategy to provide the best AI experience to our users, so we can leverage all available models to serve different user needs. We need this because different AI models are optimized for different capabilities, performance metrics and use cases and a combination of various models can handle complex tasks better than a single model…

…On the product front, our AI native application, Yuanbao, provides access to multiple models, including Chain of Thought reasoning models such as HunYuan T1 and DeepSeek R1 and fast-thinking model HunYuan Turbo S with the option of integrating web search results. Yuanbao search results can directly access high-quality proprietary content from Tencent ecosystem, such as official accounts and video accounts. By leveraging HunYuan’s multimodal capabilities, Yuanbao can process prompts in images, voice and documents in addition to text. Our cloud infrastructure supports stable and uncapped access to leading models. From February to March, Yuanbao’s DAU increased 20-fold to become the third highest AI native mobile application in China by DAU…

…We have also started testing AI features in Weixin to enhance user experience, such as for search, language input and content generation and we will be adding more AI features in Weixin going forward…

…We actually have a whole host of different consumer-facing applications and you should expect more to come. I think AI is actually in a very early stage. So it’s really hard to talk about what the eventual state would look like. But I would say, one, each product will continue to evolve into very useful and even more powerful products for users. So Yuanbao can be sort of a very strong AI native assistant and the Ema copilot could be your personal library and also a collaborative library for team collaborations. And Weixin can have many, many different features to come, right? And in addition to these products, I think our other products would have AI experiences, including QQ, including browser and other products. So I think we would see more and more AI — consumer AI-facing products. And at the same time, each one of the products will continue to evolve…

…Each one of our products would actually try to look for unique use cases in which they can leverage AI to provide a great user experience to their users…

…Yuanbao, well, right now, it is a chatbot and search. But over time, I think it would actually proliferate into a all-capable AI assistant with many different functionalities serving different types of people. So if — it would range from sort of students who want to learn and it would include all kinds of different people who, actually knowledge workers who want to complete their work and would sort of cover deep research, which allows people to very deep research into different topics.

Tencent’s management thinks that there are advantages to both developing Tencent’s own foundation models and using 3rd-party models

By investing in our own foundation models, we are able to fully leverage our proprietary data to tailor solutions to meet customized internal and customer needs, while at the same time, making use of external models allowed us to benefit from innovations across the industry.

Tencent’s management has been accelerating AI integration into Tencent’s cloud businesses, including its infrastructure as a service business, its platform as a service business, and its software as a service business; the AI-powered transcription and meeting summarisation functions in Tencent Meeting saw a year-on-year doubling in monthly active users to 15 million

We have been accelerating AI integration into our cloud business across our infrastructure, platform and Software as a Service solutions.

Through our Infrastructure as a Service solutions, enterprise customers can achieve high-performance AI training and inference capabilities at scale and developers can access and deploy mainstream foundation models.

For Platform as a Service, PaaS, our TI platform supports model fine-tuning and inference demands with flexibility, will provide powerful solutions supporting enterprise customers in customizing AI assistants using their own proprietary data and developers in generating mini programs and mobile applications through natural language prompts.

Our SaaS products increasingly benefit from AI-powered tools. Real-time transcription and meeting summarization functions in Tencent Meeting gained significant popularity resulting in monthly active users for these AI functions doubling year-on-year to 15 million. Tencent Docs also enhanced the user productivity and content generation and processing.

Tencent’s AI cloud revenue doubled in in 2024, despite management having limited the availability of GPUs for cloud services in preference for internal use-cases for ad tech, foundation model training, and inference for Yuanbao and Weixin; management has stepped up the purchase of GPUs in 2024 Q4 and expects the revenue growth of cloud services to accelerate as the new GPUs are deployed for external use cases; Tencent’s capital expenditures in 2024 Q4 increased more than 3x to US$10.7 billion from a year ago because of the higher purchases of GPUs; management believes the step-up in capex in 2024 Q4 is to a new higher steady state

In 2024, our AI cloud revenue approximately doubled year-on-year. Increased allocation of GPUs for internal use cases initially for ad tech and foundation model training and more recently on AI inference for Yuanbao and Weixin has limited our provision of GPUs to external clients and thus constrained our cloud services revenue growth. For external workloads, we have prioritized available GPUs towards high-value use cases and clients. Since the fourth quarter of 2024, we have stepped up our purchase of GPUs. And as we deploy these GPUs, we expect to accelerate the revenue growth of our overall cloud services…

…As the capabilities and benefits of AI become clearer, we have stepped up our AI investments to meet our internal business needs, train foundation models and support searching demand for inference we’re experiencing from our users. To consolidate our resources around this all important AI effort, we have reorganized our AI teams to sharpen focus on both fast product innovation and deep model research. Matching our stepped-up execution momentum and decision-making velocity, we increased annual CapEx more than threefold to USD 10.7 billion in 2024, equivalent to approximately 12% of our revenue with a notable uplift in fourth quarter of the year as we bought more GPUs for both inference needs as well as for our cloud services…

…We did step up CapEx to a new sort of higher steady state in the fourth quarter of last year…

…Part of the reason why you see such a big step up in terms of the CapEx in the fourth quarter is because we have a bunch of rush orders for GPUs for both inference as well as for our cloud service. And we would only be able to capture the large increase in terms of IaaS service demand when we actually install these GPUs into the data center, which would take some time. So I would say we probably have not really captured a lot of that during the first quarter. But over time, we will capture quite a bit of it with the arrival and installation of the GPUs.

Tencent’s management already sees positive returns for Tencent from their investment in AI; the positive returns come in 3 areas, namely, in advertising, in games, and in video and music services; in advertising, Tencent has been using AI to approve ad content more efficiently, improve ad targeting, streamline the ad creative process for advertisers, and deliver higher return on investment for advertisers; Tencent’s marketing services experienced revenue growth of 20% in 2024 because of AI integration, despite a challenging macro environment; in games, Tencent is using AI to improve content production efficiency and build in-game chat bots, among other uses; in video and music services, Tencent is using AI to improve productivity in content creation and effectively boost content discovery

We believe our investment in AI has already been generating positive returns for us…

…For advertising, we enhanced our advertising system with neural network AI capabilities since 2015. We rebuilt ad tech platform using large model capabilities since 2020, enabling long sequence user behavior analysis across multiple properties which resulted in increased user engagement and higher click-through rates. Since 2023, we have been adding large language model capabilities to facilitate more efficient approvals of ad content, to better understand merchandise categories and users commercial intent for more precise ad targeting and to provide generative AI tools for advertisers to streamline the ad creative process, leveraging AI-powered ad targeting capabilities and generative AI ad creative solutions. Our marketing services business is already a clear beneficiary of AI integration with revenue growth of 20% in 2024 amid challenging macro environment.

In games, we adopted machine learning technology in our PvP games since 2017. We leveraged AI in games to optimize matching experience, improve game balance and facilitate AI coaching for new players, empowering our evergreen games strategy. Our games business is now integrating large language model capabilities, enhanced 3D content production efficiency and to empower in-game chatbots.

For our video and music services, we’re leveraging AI to improve productivity in animation, live action video and music content creation. Our content recommendation algorithms are powered by AI and are proven effective in boosting content discovery. These initiatives enables us to better unlock the potential of our great content platforms…

…Across pretty much every industry we monitor the AI enhancements we’re deploying and delivering superior return on investment for advertisers versus what they previously enjoyed and versus what’s available elsewhere.

Tencent’s management expects to further increase capital expenditure in 2025 and for capital expenditure to be a low-teens percentage of revenue for the year; while capital expenditure in 2025 is expected to increase, the rate of growth has slowed down significantly

We intend to further increase our capital expenditures in 2025 and expect our CapEx to account for low teens percentage of our revenue…

…[Question] You guided a CapEx to revenue ratio of low-teens for 2025, which is a similar ratio as for ’24. So basically, this guidance implies a significant slowdown of CapEx growth.

Tencent’s management sees several nuances on the impact to Tencent’s profit margins from the higher AI capital expenditures expected, but they are optimistic that Tencent will be able to protect its margins; the AI capital expenditures go into 4 main buckets, namely, (1) ad tech and games, (2) large language model training, (3) renting out GPUs in the cloud business, and (4) consumer-facing inference; management sees good margins in the 1st bucket, decent margins in the 3rd bucket, and potentially some margin-pressure in the 4th bucket; but in the 4th bucket, management sees (1) the potential to monetise consumer-facing inference through a combination of advertising revenue and value-added services, and (2) avenues to reduce unit costs through software and better algorithms

[Question] As we step up the CapEx on AI, our margin will be inevitably dragged by additional depreciation and R&D expenses. So over the past few years, we have seen meaningful increase in margin as we focus on high-quality growth. So going forward, how should we balance between growth and profitability improvement?

[Answer] It’s worth digging into exactly where that CapEx is going to understand whether the depreciation becomes a margin pressure or not. So the most immediate use of the CapEx is GPUs to support our ad tech and to a lesser extent, our games businesses. And you can see from our results, you can hear from what Martin talked about, that, that CapEx actually generates good margins, high returns.

A second use of CapEx was GPUs for large language model training…

…Third, there’s CapEx related to our cloud business, which, we buy this GPU servers, we rent them out to customers, we generate a return. It may not be the highest return business in our portfolio but nonetheless, it’s a positive return. It covers the cost of the GPUs and therefore, the attendant depreciation.

And then finally, where I think there is potentially the short-term pressure is the CapEx for 2C [to-consumers] inference. And it — that is an additional cost pressure but we believe it’s a manageable cost pressure because that CapEx is a subset of the total CapEx. And we’re also optimistic that over time those — the 2C inference activity that we’re generating, just like previous activity within different Tencent platforms will be monetizing through a combination of advertising revenue and value-added services. So overall, while we understand that you have questions around the step-up in CapEx and how that translates into profitability over time, we’re actually quite optimistic that we can continue to grow the business while protecting margins…

…In the inference for consumer-facing product. There’s actually a lot of avenues through which we can actually reduce the unit cost by technical means, by software and by better algorithms. So I think that’s also sort of a factor to keep in mind.

Tencent’s management believes that the AI industry is now getting much higher productivity on large language model training from existing GPUs without needing to add additional GPUs at the pace previously expected, as a result of DeepSeek’s breakthroughs; previously, the belief was that each new generation of large language models would require an order of magnitude more GPUs; Tencent’s AI-related capital expenditure is the largest amongst Chinese technology companies; management thinks that Chinese technology companies are spending less on capital expenditure as a percentage of revenue than Western peers because Chinese companies have been prioritizing efficient utilization of GPUs without impairing the ultimate effectiveness of the AI technology developed

There was a period of time last year when there was a belief that every new generation of large language model required an order of magnitude more GPUs. That period of time ended with the breakthroughs that DeepSeek demonstrated. And now the industry and we within the industry are getting much higher productivity on a large language model training from existing GPUs without needing to add additional GPUs at the pace previously expected…

…There was a period last year when people asked us if our CapEx was big enough relative to our China peers, relative to our global peers. And now out of the listed companies, I think we had the largest CapEx of any China tech company in the fourth quarter. So we’re at the forefront among our China peers. In general, the China tech companies are spending less on CapEx as a percentage of revenue than some of their Western peers. But we believe for some time that’s because the Chinese companies are generally prioritizing efficiency and utilization — efficient utilization of the GPU servers. And that doesn’t necessarily impair the ultimate effectiveness of the technology that’s being developed. And I think DeepSeek’s success really sort of symbolized and solidified, demonstrated that, that reality.

Tencent’s management thinks AI can benefit Tencent’s games business in 3 ways, namely, (1) a direct, more short-term benefit in helping game developers be more productive, (2) an indirect, more long-term benefit in terms of games becoming an important element of human expression in an AI-dominated world, and (3) allow evergreen games to be more evergreen

We do believe that games benefit in a direct and potentially a less direct way from AI technology enhancements. The direct way is the game developers using AI to assist them in creating more content more quickly and serving more users more effectively. And then the indirect way, which may be more of a multi-decade rather than the second half of this year story is that as humanity uses AI more broadly, then we think there’ll be more time and also more desire for high agency activities among people who are now empowered by AI. And so one of the best ways for them to express themselves in a high agency way rather than a passive way is through interactive entertainment, which is games…

…We actually felt AI would allow evergreen games to be more evergreen. And we are already seeing sort of how AI can help us to execute and magnify our evergreen strategy. And part of it is within production, right, you can actually produce great content now within a shorter period of time so that you can keep updating the games with higher frequency of high-quality content. And with the PvE experience, when you have smarter box, right, you actually sort of make the game more exciting and more like PvP. And within PvP, a lot of the matching and balancing and coaching of new users can actually sort of be done in a much better way when you apply AI.

Tencent’s management sees strong competitive advantages that Tencent has when it comes to AI agents because of the large user base of Tencent’s products and the huge variety of activities that happen within Tencent’s products

We would be able to build stand-alone AI agents by leveraging models that are of great quality and at the same time by leveraging the fact that we have a lot of consumers on our different software platforms like our browser, like Yuanbao over time. But at the same time, right, even within Weixin and within QQ, we can have AI agents. And the AI agents can actually leverage the ecosystem within the apps and provide really great service to our users by completing complex tasks, right? If you look at Weixin, for example, Weixin has got a lot of users, a very long user time per day as well as high frequency of users opening up the app, that’s 1 advantage. The second advantage is that if you look at the activities within Weixin is actually very, very diversified, right? It’s not just sort of entertainment, it’s not just transactions, it’s actually sort of social communication and content and a lot of people will conduct their work within Weixin, a lot of people conduct their learning within Weixin and there are a lot of transactions that go through Weixin. And there’s a multitude of Mini Programs, which actually allowed all sorts of different activities to be carried out, right? So if you look at the Mini program ecosystem, we can easily build an agent based on a model that actually can connect to a lot of the different Mini Programs and have activities and complex tasks completed for our users. So I think those are all very distinctive advantages that we have.

Tencent’s management believes that AI search will eventually replace traditional search

At a high level, if we look at the history of web search subsuming web directory, if we look at our own behavior, with AI prompts vis-a-vis traditional search, I think it’s possible that AI search will subsume traditional search because ultimately, web directory traditional search, AI prompt, all represent mechanisms for accessing the Internet’s knowledge graph.

Tencent’s management believes that in China, AI chatbots will be monetised first through performance advertising followed by value-added services, as opposed to in the West, where AI chatbots have been monetised first through subscription models followed by performance advertising

In terms of how the AI prompt will be monetized, time will tell but I think that we can already see in the Western world, the first monetization is through subscription models and then over time, performance advertising will follow. I think in China, it will start with performance advertising and then value-added services will follow.

Veeva Systems (NYSE: VEEV)

Veeva’s management’s AI strategy for Veeva is to have its Commercial Cloud, Development Cloud, and Quality Cloud be the life sciences industry’s standard core systems of record; management is making Veeva’s data readily available for the development of AI applications by Veeva and 3rd parties through the Direct Data API released in 2024; management is seeing good uptake on the Direct Data API; the Direct Data API will be free to all of Veeva’s customers because management wants people to be building on the API; management found a way to offer the Direct Data API with lesser compute resources than originally planned for; Veeva is already using the Direct Data API internally, and 10 customers are already using it; it takes time for developers to be used to the Direct Data API, because it’s a fundamentally new type of API, but it’s a great API; management believes that Direct Data API will enable the life sciences industry to leverage their core data through AI faster than any other industry

We also executed well on our AI strategy. Commercial Cloud, Development Cloud, and Quality Cloud are becoming the industry’s standard core systems of record. With significant technology innovation including the Direct Data API released this year, we are making the data from our applications readily available for the development of relevant, timely AI solutions built by Veeva, our customers, and partners…

…We are seeing good uptake of the Direct Data API. And we — yes, as you mentioned, we recently announced that that’s going to be free to all of our customers. And — the reason there is we want everybody building on that pack of API. It’s just a much better, faster API for many use cases, and we found a way to do it where it was not going to consume as many compute resources as we thought it was…

…We are using it internally, for example, for connecting different parts of our clinical suite, different parts of our safety suite together and our partners are starting to do it. We have more than 10 customers that are already doing it. Some of them are large customers. it takes some time because it’s a different paradigm for integration. People have been using a hammer for a long time. And now you’re giving them a Jack Hammer and they got to learn how to use it. But we are super enthused. It’s a fundamental new type of API where you can get like all of the data out of your Vault, super quickly…

…I’m really pleased about what we’re doing for the life sciences industry because many of our core systems are Veeva and now their core systems are going to be enabled with this fundamental new API that’s going to allow them to leverage their core data faster than any other industry.

Reminder from management that Veeva recently announced 3 new AI solutions, namely, Vault CRM Bot, CRM Voice Control, and MLR Bot; management has more AI solutions in the pipeline, but the timing for release is still unclear; management wants to invest more in AI solutions and they think the company has strong leadership in that area

We announced new Veeva AI Solutions including Vault CRM Bot, CRM Voice Control, and MLR Bot…

……Now with CRM Voice Control that we’ll be bringing out this year, and also CRM Bot and the MLR Bot, medical legal regulatory review, and we have quite a few others in the plan, too. We don’t know exactly which ones we’ll bring out when, but we have — we’re putting more investment in AI solutions. We centralized the group around that, so we can develop. I have a strong leader there and develop more core competency around around AI.

Veeva’s management was initially a little skeptical of AI because of the amount of money flowing in, and the amount of hype surrounding it

[Question] I want to start with the AI offerings that you’ve built out Peter, maybe if you were on the tape back a year, there was a little bit of a perception from the investment community. That you were coming off as maybe a little bit skeptical on AI, but now you’ve come out with a lot of these products. Maybe can you walk us through kind of what’s driven the desire or the momentum to push out these products kind of quickly?

[Answer] AI is certainly captivating technology, right? So much money going into it, so much progress, and so much hype.

Veeva’s management thinks AI is shaking out the way they expected to, which is the existence of many large language models; management also thinks the development of AI has become more stable

If we just stay at that level, I’m really pleased that things are starting to shake out roughly how we thought they were going to take out. There’s not going to be one large language model, there are going to be multiple. There’s not going to be 50, but there’s going to be a good handful and they’re going to specialize in different areas. And it’s not so unstable anymore, where you wake up and everything changes, right? DeepSeek came out, came out, Yes, well, guess what? The world keeps turning. NVIDIA is going to have their own model? That’s okay, and the world keeps turning. So I think it’s starting to settle out.

Veeva’s management sees the infrastructure layer of AI as being really valuable, but they also see a lot of value in building specific use cases on top of the infrastructure layer, and that is where they want Veeva to play in

So it’s settling out that these core large language models are going to be at the platform level and that’s super valuable, right? That’s not where companies like Veeva play in, that core infrastructure level. It’s very valuable. But there’s a lot of great value on specific use cases on top that can be used in the workflow. So that’s what we’re doing now, focusing on our AI solutions.

Veeva’s management is using AI internally but it’s still early days and it has yet to contribute improvements to Veeva’s margin; Veeva’s expected margin improvements for 2025 (FY2026) is not related to AI usage

[Question] Going back to the topic of AI… how you’re leaning into kind of internal utilization too, if we think about kind of some of the margin strength you’re delivering throughout the business?

[Answer] Around the internal use of AI and the extent to which that was contributing to margins, I think. And I think the short answer there is it’s an area that we’re really excited about internally as well. We’re building strategies around, but it’s not a major contributor to the margin expansion that we saw in Q4 or in the coming year. So it’s something we’re looking into. We’re building strategies around. It’s not something we’re counting on, though, to deliver on this year’s guidance.

In 2023 and 2024, Veeva’s management was seeing customers get distracted from core technology spending because the customers were chasing AI; management is no longer seeing the AI distraction at play

I believe we called it before AI disruption, maybe that was 18 months or so a year ago. I think that’s largely behind us. Our customers have settled into what AI is and what it does. They’re still doing some innovation projects, but it’s not consuming them or distracting from the core work. So I think we’re largely through that [ 3 ] of AI distraction now.

Veeva’s management thinks that Veeva is the fastest path to AI for a life sciences industry CRM because any AI features will have to be embedded in the workflow of a life sciences company

It turns out Veeva is the fastest path to AI that you can use in CRM because it has to be done in the workflow of what you’re doing. This is not some generic AI. This is AI for pre-call planning for compliance, for how to — for the things that a pharmaceutical rep does in a compliant way based on the data sources that are needed in CRM. So Veeva is the fastest path to AI.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Adobe, Meituan, MongoDB, Okta, Tencent, and Veeva Systems. Holdings are subject to change at any time.

What We’re Reading (Week Ending 23 March 2025)

The best articles we’ve read in recent times on a wide range of topics, including investing, business, and the world in general.

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 23 March 2025:

1. Investing Politics: Globalization Backlash and Government Disruption! – Aswath Damodaran

Globalization has taken different forms through the ages, with some violent and toxic variants, but the current version of globalization kicked into high gear in the 1980s, transforming every aspect of our lives…

…The biggest winner from globalization has been China, which has seen its economic and political power surge over the last four decades. Note that the rise has not been all happenstance, and China deserves credit for taking advantage of the opportunities offered by globalization, making itself first the hub for global manufacturing and then using its increasing wealth to build its infrastructure and institutions…

…China’s share of global GDP increased ten-fold between 1980 and 2023…

…Between 2010 and 2023, China accounted for almost 38% of global economic growth, with only the United States having a larger share…

…Consumers have benefited from globalization in many ways, starting with more products to choose from and often at lower prices than in pre-globalization days. From being able to eat whatever we want to, anytime of the year, to wearing apparel that has become so cheap that it has become disposable, many of us, at least on the surface, have more buying power…

…Over the last few decades, not only have more companies been able to list themselves on financial markets, but these markets has become more central to public policy. In many cases, the market reaction to spending, tax or economic proposals has become the determinant on whether they get adopted or continued. As financial markets have risen in value and importance, the cities (New York, London, Frankfurt, Shanghai, Tokyo and Mumbai) where these markets are centered have gained in importance and wealth, if not in livability, at the expense of the rest of the world…

…The rise of China from globalization also illustrates the fading of Japan and Europe over the period, with the former declining from 17.8% of global GDP in 1995 to 3.96% in 2023 and the latter seeing its share dropping from 25.69% of global GDP in 1990 to 14.86%…

…I listed consumers as winners from globalization, and they were, on the dimensions of choice and cost, but they also lost in terms of control of where their products were made, and by whom. To provide a simplistic example, the shift from buying your vegetables, fish and meat from local farmers, fishermen and butchers to factory farmers and supermarkets may have made the food more affordable, but it has come at a cost…

…While there are a host of other factors that have also contributed to the decline of small businesses, globalization has been a major contributor, as smaller businesses now find themselves competing against companies who make their products thousands of miles away, often with very different cost structures and rules restricting them. Larger businesses not only had more power to adapt to the challenges of globalization, but have found ways to benefit from it, by moving their production to the cheapest and least restrictive locales. In one of my data updates for this year, I pointed to the disappearance of the small firm effect, where small firms historically have earned higher returns than large cap companies, and globalization is a contributing factor…

…The flip side of the rise of China and other countries as manufacturing hubs, with lower costs of operation, has been the loss of manufacturing clout and jobs for the West…

…In the United States, the number of manufacturing jobs peaked at close to 20 million in 1979 and dropped to about 13 million in 2024, and manufacturing wages have lagged wage growth in other sectors for much of that period…

…I believe that globalization has been a net plus for the global economy, but one reason it is in retreat is because of a refusal on the part of its advocates to acknowledge its costs and the dismissal of opposition to any aspect of globalization as nativist and ignorant…

…Trump, a real estate developer with multiple international properties, is an imperfect spokesperson of the anti-globalization movement, but it is undeniable that he has tapped into, and benefited from, its anger. While he was restrained by norms and tradition in his first term, those constraints seem to have loosened in this second go around, and he has wielded tariffs as a weapon and is open about his contempt for global organizations. While economists are aghast at the spectacle, and the economic consequences are likely to be damaging, it is not surprising that a portion of the public, perhaps even a majority, are cheering Trump on.

To those who are nostalgic for a return to the old times, I don’t believe that the globalization genie can go back into the bottle, as it has permeated not only every aspect of business, but also significant portions of our personal lives. The world that will prevail, if a trade war plays out, will be very different than the one that existed before globalization took off…

…On the revenue growth front, companies that derive most or all of their revenues domestically will benefit and companies that are dependent on foreign sales will be hurt by tariff wars…

…Collectively, about 28% of the revenues, in 2023, of the companies in the S&P 500 came from foreign markets, but technology companies are most exposed (with 59% of revenues coming from outside the country) and utilities least exposed (just 2%) to foreign revenue exposure. It is also worth noting that the larger market cap companies of the S&P 500 have a higher foreign market revenue exposure than smaller market cap companies…

…To the extent that companies are altering their decisions on where to build their next manufacturing facilities, as a result of tariff fears or in hope of government largesse, there should be an effect on reinvestment, with an increase in reinvestment (lower sales to capital ratios) at businesses where this move will create investment costs. Looking across businesses, this effect is likely to be more intense at manufacturing companies, where moving production is more expensive and difficult to do, than at technology or service firms…

…While it is easy to blame market uncertainty on Trump, tariffs and trade wars for the moment, the truth is that the forces that have led us here have been building for years, both in our political and economic arenas. In short, even if the tariffs cease to be front page news, and the fears of an immediate trade war ease, the underlying forces of anti-globalization that gave rise to them will continue to play out in global commerce and markets. For investors, that will require a shift away from the large cap technology companies that have been the market leaders in the last two decades back to smaller cap companies with a more domestic focus.

2. Big Retailers’ Hardball Tariff Playbook: Haggle, Diversify, Raise Prices – Hannah Miao and Sarah Nassauer

Some suppliers say Walmart, Home Depot and other retailers are pushing a variation of the same demand: Make a price concession or shift production out of China. Otherwise, the suppliers risk losing some business…

…Some of the requests have raised the ire of Chinese officials. Authorities in China summoned Walmart for a meeting in recent days after some suppliers complained the largest U.S. retailer by annual revenue was pressuring them to cut prices and absorb the tariff cost…

…Some pricing negotiations are hitting an impasse because many of these manufacturers are often already operating on razor-thin margins, according to suppliers. And retailers don’t want to raise prices for shoppers so they can continue to compete for market share…

…After the first 10% tariff on Chinese goods in February, Home Depot asked one of its U.S. suppliers of lighting and home decor to absorb the cost, according to an executive at the supplier. The supplier agreed to a two-month, 10% discount, part of which would be covered by its Chinese manufacturer.

After the second 10% tariff in March, the supplier declined another request from Home Depot to lower prices again. Instead, the supplier is moving production to Southeast Asia so it can eventually charge the home-improvement retailer the original price, the executive said…

…The tariff planning is especially complicated because companies have little sense of which tariff threats will materialize and where new ones could emerge, retailers and suppliers say…

…In some cases, retailers and manufacturers have decided it is worth it to keep production in China to maintain quality. Costco, the warehouse chain, plans to continue selling patio furniture made in China—even at an elevated price—because it is higher quality than versions made in other countries, Costco executives said. Costco and its supplier will absorb some of the cost increase and pass some on to shoppers, they said.

3. An Interview with OpenAI CEO Sam Altman About Building a Consumer Tech Company – Ben Thompson and Sam Altman

What’s going to be more valuable in five years? A 1-billion daily active user destination site that doesn’t have to do customer acquisition, or the state-of-the-art model?

SA: The 1-billion user site I think.

Is that the case regardless, or is that augmented by the fact that it seems, at least at the GPT-4 level, I mean, I don’t know if you saw today LG just released a new model. There’s going to be a lot of, I don’t know, no comments about how good it is or not, but there’s a lot of state-of-the-art models.

SA: My favorite historical analog is the transistor for what AGI is going to be like. There’s going to be a lot of it, it’s going to diffuse into everything, it’s going to be cheap, it’s an emerging property of physics and it on its own will not be a differentiator.

What will be the differentiator?

SA: Where I think there’s strategic edges, there’s building the giant Internet company. I think that should be a combination of several different key services. There’s probably three or four things on the order of ChatGPT, and you’ll want to buy one bundled subscription of all of those. You’ll want to be able to sign in with your personal AI that’s gotten to know you over your life, over your years to other services and use it there. There will be, I think, amazing new kinds of devices that are optimized for how you use an AGI. There will be new kinds of web browsers, there’ll be that whole cluster, someone is just going to build the valuable products around AI. So that’s one thing.

There’s another thing, which is the inference stack, so how you make the cheapest, most abundant inference. Chips, data centers, energy, there’ll be some interesting financial engineering to do, there’s all of that.

And then the third thing is there will be just actually doing the best research and producing the best models. I think that is the triumvirate of value, but most models except the very, very leading edge, I think will commoditize pretty quickly.

So when Satya Nadella said models are getting commoditized, that OpenAI is a product company, that’s still a friendly statement, we’re still on the same team there?

SA: Yeah, I don’t know if it came across as a compliment to most listeners, I think he meant that as a compliment to us…

…It doesn’t. But to that point, you just released a big API update, including access to the same computer use model that undergirds Operator, a selling point for GPT Pro. You also released the Responses API and I thought the most interesting part about the Responses API is you’re saying, “Look, we think this is much better than the Chat Completions API, but of course we’ll maintain that, because lots of people have built on that”. It’s sort of become the industry standard, everyone copied your API. At what point is this API stuff and having to maintain old ones and pushing out your features to the new ones turn into a distraction and a waste of resources when you have a Facebook-level opportunity in front of you?

SA: I really believe in this product suite thing I was just saying. I think that if we execute really well, five years from now, we have a handful of multi-billion user products, small handful and then we have this idea that you sign in with your OpenAI account to anybody else that wants to integrate the API, and you can take your bundle of credits and your customized model and everything else anywhere you want to go. And I think that’s a key part of us really being a great platform.

Well, but this is the tension Facebook ran into. It’s hard to be a platform and an Aggregator, to use my terms. I think mobile was great for Facebook because it forced them to give up on pretensions of being a platform. You couldn’t be a platform, you had to just embrace being a content network with ads. And ads are just more content and it actually forced them into a better strategic position.

SA: I don’t think we’ll be a platform in a way that an operating system is a platform. But I think in the same way that Google is not really a platform, but people use sign in with Google and people take their Google stuff around the web and that’s part of the Google experience, I think we’ll be a platform in that way…

…From my perspective, when you talk about serving billions of users and being a consumer tech company. This means advertising. Do you disagree?

SA: I hope not. I’m not opposed. If there is a good reason to do it, I’m not dogmatic about this. But we have a great business selling subscriptions.

There’s still a long road to being profitable and making back all your money. And then the thing with advertising is it increases the breadth of your addressable market and increases the depth because you can increase your revenue per user and the advertiser foots the bill. You’re not running into any price elasticity issues, people just use it more.

SA: Currently, I am more excited to figure out how we can charge people a lot of money for a really great automated software engineer or other kind of agent than I am making some number of dimes with an advertising based model…

…Especially Deep Research, it’s amazing. But I am maybe more skeptical about people’s willingness to go out and pay for something, even if the math is obvious, even if it makes them that much more productive. And meanwhile, I look at this bit where you’re talking about building memory. Part of what made the Google advertising model so brilliant is they didn’t actually need to understand users that much because people typed into the search bar what they were looking for. People are typing a tremendous amount of things into your chatbot. And even if you served the dumbest advertising ever, in many respects, and even if you can’t track conversions, your targeting capability is going to be out of this world. And, by the way, you don’t have an existing business model to worry about undercutting. My sense is this is so counter to what everyone at OpenAI signed up for, that’s the biggest hurdle. But to me, from a business analyst, this seems super obvious and you’re already late.

SA: The kind of thing I’d be much more excited to try than traditional ads is a lot of people use Deep Research for e-commerce, for example, and is there a way that we could come up with some sort of new model, which is we’re never going to take money to change placement or whatever, but if you buy something through Deep Research that you found, we’re going to charge like a 2% affiliate fee or something. That would be cool, I’d have no problem with that. And maybe there’s a tasteful way we can do ads, but I don’t know. I kind of just don’t like ads that much…

…SA: Totally. I think DeepSeek was — they made a great team and they made a great model, but the model capability was, I think, not the thing there that really got them the viral moment. But it was a lesson for us about when we leave a feature hidden, we left chains of thought hidden, we had good reasons for doing it, but it does mean we leave space for somebody else to have a viral moment. And I think in that way it was a good wake-up call. And also, I don’t know, it convinced me to really think differently about what we put in the free tier and now the free tier is going to get GPT-5 and that’s cool.

Ooo, ChatGPT-5 hint. Well, I’ll ask you more about that later…

…In your recent proposal about the AI Action Plan, OpenAI expressed concern about companies building on DeepSeek’s models, which are, in one of the phrases about them, “freely available”. Isn’t the solution, if that’s a real concern, to make your models freely available?

SA: Yeah, I think we should do that.

So when-

SA: I don’t have a launch to announce, but directionally, I think we should do that.

You said before, the one billion destination site is more valuable than the model. Should that flow all the way through to your release strategy and your thoughts about open sourcing?

SA: Stay tuned.

Okay, I’ll stay tuned. Fair enough.

SA: I’m not front-running, but stay tuned…

…Is there a bit where isn’t hallucination good? You released a sample of a writing model, and it sort of tied into one of my longstanding takes that everyone is working really hard to make these probabilistic models behave like deterministic computing, and almost missing the magic, which is they’re actually making stuff up. That’s actually pretty incredible.

SA: 100%. If you want something deterministic, you should use a database. The cool thing here is that it can be creative and sometimes it doesn’t create quite the thing you wanted. And that’s okay, you click it again.

Is that an AI lab problem that they’re trying to do this? Or is that a user expectation problem? How can we get everyone to love hallucinations?

SA: Well, you want it to hallucinate when you want and not hallucinate when you don’t want. If you’re asking, “Tell me this fact about science,” you’d like that not to be a hallucination. If you’re like, “Write me a creative story,” you want some hallucination. And I think the problem, the interesting problem is how do you get models to hallucinate only when it benefits the user?…

…I think some skeptics, including me, have framed some aspects of your calls for regulation as an attempt to pull up the ladder on would-be competitors. I’d ask a two-part question. Number one, is that unfair? And if the AI Action Plan did nothing other than institute a ban on state level AI restrictions and declare that training on copyright materials fair use, would that be sufficient?

SA: First of all, most of the regulation that we’ve ever called for has been just say on the very frontier models, whatever is the leading edge in the world, have some standard of safety testing for those models. Now, I think that’s good policy, but I sort of increasingly think the world, most of the world does not think that’s good policy, and I’m worried about regulatory capture. So obviously, I have my own beliefs, but it doesn’t look to me like we’re going to get that as policy in the world and I think that’s a little bit scary, but hopefully, we’ll find our way through as best as we can and probably it’ll be fine. Not that many people want to destroy the world.

But for sure, you don’t want to go put regulatory burden on the entire tech industry. Like we were calling for something that would have hit us and Google and a tiny number of other people. And again, I don’t think the world’s going to go that way and we’ll play on the field in front of us. But yes, I think saying that fair use is fair use and that states are not going to have this crazy complex set of differing regulations, those would be very, very good.

You are supporting export controls or by you, I mean, OpenAI in this policy paper. You talked about the whole stack, that triumvirate. Do you worry about a world where the US is dependent on Taiwan and China is not?

SA: I am worried about the Taiwan dependency, yes…

…Okay, sure. Intel needs a customer. That’s what they need more than anything, a customer that is not Intel. Get OpenAI, become the leading customer for the Gaudi architecture, commit to buying a gazillion chips and that will help them. That will pull them through. There’s your answer.

SA: If we were making a chip with a partner that was working with Intel and a process that was compatible and we had, I think, a sufficiently high belief in their ability to deliver, we could do something like that. Again, I want to do something. So I’m not trying to dodge…

…So Dario and Kevin Weil, I think, have both said or in various aspects that 99% of code authorship will be automated by sort of end of the year, a very fast timeframe. What do you think that fraction is today? When do you think we’ll pass 50% or have we already?

SA: I think in many companies, it’s probably past 50% now. But the big thing I think will come with agentic coding, which no one’s doing for real yet.

What’s the hangup there?

SA: Oh, we just need a little longer.

Is it a product problem or is it a model problem?

SA: Model problem.

Should you still be hiring software engineers? I think you have a lot of job listings.

SA: I mean, my basic assumption is that each software engineer will just do much, much more for a while. And then at some point, yeah, maybe we do need less software engineers…

…What is AGI? And there’s a lot of definitions from you. There’s a lot of definitions in OpenAI. What is your current, what’s the state-of-the-art definition of AGI?

SA: I think what you just said is the key point, which is it’s a fuzzy boundary of a lot of stuff and it’s a term that I think has become almost completely devalued. Some people, by many people’s definition, we’d be there already, particularly if you could go like transport someone from 2020 to 2025 and show them what we’ve got.

Well, this was AI for many, many years. AI was always what we couldn’t do. As soon as we could do it, it’s machine learning. And as soon as you didn’t notice it, it was an algorithm.

SA: Right. I think for a lot of people, it’s something about like a fraction of the economic value. For a lot of people, it’s something about a general purpose thing. I think they can do a lot of things really well. For some people, it’s about something that doesn’t make any silly mistakes. For some people, it’s about something that’s capable of self-improvement, all those things. It’s just there’s not good alignment there.

What about an agent? What is an agent?

SA: Something that can like go autonomously, do a real chunk of work for you.

To me, that’s the AGI thing. That is employee replacement level.

SA: But what if it’s only good at like some class of tasks and can’t do others? I mean, some employees are like that too…

…Given that, does that make you more optimistic, less optimistic? Do you see this bifurcation that I think there’s going to be between agentic people? This is a different agentic word, but see where we’re going. We need to invent more words here. We’ll ask ChatGPT to hallucinate one for us. People who will go and use the API and the whole Microsoft Copilot idea is you have someone accompanying you and it’s a lot of high talk, “Oh, it’s not going to replace jobs, it’s going to make people more productive”. And I agree that will happen for some people who go out to use it. But you look back, say, at PC history. The first wave of PCs were people who really wanted to use PCs. PCs, a lot of people didn’t. They had one put on their desk and they had to use it for a specific task. And really, you needed a generational change for people to just default to using that. Is AI, is that the real limiting factor here?

SA: Maybe, but that’s okay. Like as you mentioned, that’s kind of standard for other tech evolutions.

But you go back to the PC example, actually, the first wave of IT was like the mainframe, wiped out whole back rooms. And because actually, it turned out the first wave is the job replacement wave because it’s just easier to do a top-down implementation.

SA: My instinct is this one doesn’t quite go like that, but I think it’s always like super hard to predict.

What’s your instinct?

SA: That it kind of just seeps through the economy and mostly kind of like eats things little by little and then faster and faster.

You talk a lot about scientific breakthroughs as a reason to invest in AI, Dwarkesh Patel recently raised the point that there haven’t been any yet. Why not? Can AI actually create or discover something new? Are we over-indexing on models that just aren’t that good and that’s the real issue?

SA: Yeah, I think the models just aren’t smart enough yet. I don’t know. You hear people with Deep Research say like, “Okay, the model is not independently discovering new science, but it is helping me discover new science much faster.” And that, to me, is like pretty much as good.

Do you think a transformer-based architecture can ever truly create new things or is it just spitting out the median level of the Internet?

SA: Yes.

Well, what’s going to be the breakthrough there?

SA: I mean, I think we’re on the path. I think we just need to keep doing our thing. I think we’re like on the path…

…Do humans have innate creativity or is it just recombining knowledge in different sorts of ways?

SA: One of my favorite books is The Beginning of Infinity by David Deutsch, and early on in that book, there’s a beautiful few pages about how creativity is just taking something you saw before and modifying it a little bit. And then if something good comes out of it, someone else modifies it a little bit and someone else modifies it a little bit. And I can sort of believe that. And if that’s the case, then AI is good at modifying things a little bit.

To what extent is the view that you could believe that grounded in your long-standing beliefs versus what you’ve observed, because I think this is a very interesting — not to get all sort of high-level metaphysical or feel, like I said, theological almost — but there does seem to be a bit where one’s base assumptions fuel one’s assumptions about AI’s possibilities. And then, most of Silicon Valley is materialistic, atheistic, however you want to put it. And so of course, we’ll figure it out, it’s just a biological function, we can recreate it in computers. If it turns out we never actually do create new things, but we augment humans creating new things, would that change your core belief system?

SA: It’s definitely part of my core belief system from before. None of this is anything new, but no, I would assume we just didn’t figure out the right AI architecture yet and at some point, we will.

4. The Last Decision by the World’s Leading Thinker on Decisions – Jason Zweig

Kahneman was widely mourned nearly a year ago when his death was announced. Only close friends and family knew, though, that it transpired at an assisted-suicide facility in Switzerland. Some are still struggling to come to terms with his decision…

…But I never got to say goodbye to Danny and don’t fully understand why he felt he had to go. His death raises profound questions: How did the world’s leading authority on decision-making make the ultimate decision? How closely did he follow his own precepts on how to make good choices? How does his decision fit into the growing debate over the downsides of extreme longevity? How much control do we, and should we, have over our own death?…

…I think Danny wanted, above all, to avoid a long decline, to go out on his terms, to own his own death. Maybe the principles of good decision-making that he had so long espoused—rely on data, don’t trust most intuitions, view the evidence in the broadest possible perspective—had little to do with his decision.

His friends and family say that Kahneman’s choice was purely personal; he didn’t endorse assisted suicide for anyone else and never wished to be viewed as advocating it for others.

Some of Kahneman’s friends think what he did was consistent with his own research. “Right to the end, he was a lot smarter than most of us,” says Philip Tetlock, a psychologist at the University of Pennsylvania. “But I am no mind reader. My best guess is he felt he was falling apart, cognitively and physically. And he really wanted to enjoy life and expected life to become decreasingly enjoyable. I suspect he worked out a hedonic calculus of when the burdens of life would begin to outweigh the benefits—and he probably foresaw a very steep decline in his early 90s.”

Tetlock adds, “I have never seen a better-planned death than the one Danny designed.”…

…As I wrote in a column about Kahneman last year: “I once showed him a letter I’d gotten from a reader telling me—correctly but rudely—that I was wrong about something. ‘Do you have any idea how lucky you are to have thousands of people who can tell you you’re wrong?’ Danny said.”…

…Kahneman knew the psychological importance of happy endings. In repeated experiments, he had demonstrated what he called the peak-end rule: Whether we remember an experience as pleasurable or painful doesn’t depend on how long it felt good or bad, but rather on the peak and ending intensity of those emotions.

“It was a matter of some consternation to Danny’s friends and family that he seemed to be enjoying life so much at the end,” says a friend. “‘Why stop now?’ we begged him. And though I still wish he had given us more time, it is the case that in following this carefully thought-out plan, Danny was able to create a happy ending to a 90-year life, in keeping with his peak-end rule. He could not have achieved this if he had let nature take its course.”

Did turning 90 play a role in his decision? Kahneman and Tversky’s early research showed that when people are uncertain, they will estimate numbers by “anchoring,” or seizing on any figure that happens to be handy, regardless of how relevant it is to the decision.

Another of Kahneman’s principles was the importance of taking what he called the outside view: Instead of regarding each decision as a special case, you should instead consider it as a member of a class of similar situations. Gather data on comparable examples from that reference class, then consider why your particular case might have better or worse prospects.

One possible approach: Kahneman could have gathered data to determine whether people who live to the age of 95 or beyond tend to regret not dying at the age of 90—adjusting for the difficulty of getting reliable reports from patients with dementia and other debilitating conditions. Perhaps he did something along those lines; I don’t know…

…As Danny’s final email continued:

I discovered after making the decision that I am not afraid of not existing, and that I think of death as going to sleep and not waking up. The last period has truly not been hard, except for witnessing the pain I caused others. So if you were inclined to be sorry for me, don’t be.

As death approaches, should we make the best of whatever time we have left with those we love the most? Or should we spare them, and ourselves, from as much as possible of our inevitable decline? Is our death ours alone to own?

Danny taught me the importance of saying “I don’t know.” And I don’t know the answers to those questions. I do know the final words of his final email sound right, yet somehow feel wrong:

Thank you for helping make my life a good one.

5. Dead of Winter – Doomberg

After a somewhat colder-than-average winter and the cessation of gas flows through the very Sudzha pipeline that Russian special forces just snaked through for their surprise assault, European natural gas storage levels are at dangerously low levels for this time of year:…

…The last energy crisis began many months before Russia’s military crossed into Ukraine. Europe’s vulnerability—driven by a foolish decision to forgo topping off its reserves in the summer of 2021—almost certainly convinced Russian President Vladimir Putin that he held sufficient leverage to risk war in early 2022. Three years later, with the conflict seemingly entering its final stages, surely the continent isn’t repeating the mistakes of the recent past? Perhaps it is:

“As the first proper winter in Europe in three years is drawing to an end, the continent faces a race against time—and prices—to restock with natural gas for next winter…

Europe could need as many as 250 additional [liquefied natural gas] LNG cargoes to arrive in the summer to refill its inventories back up to 90% by November 1, as the current EU regulation stipulates, per Reuters calculations reported by columnist Ron Bousso… The LNG market appears to be tightening, with supply not rising fast enough in early 2025 to meet demand.”…

…Europe’s vulnerability is now measurably higher compared to three years ago. Russian natural gas no longer flows through the Nord Stream and Yamal pipelines, nor various connections through Ukraine, eliminating access to a total capacity of 18 billion cubic feet per day (bcf/d). Only the two pipelines entering Turkey via the Black Sea—TurkStream and Blue Stream—are still pumping gas. The balance of European demand will need to be met by expensive LNG imports, primarily from the US and Qatar.

Unfortunately for Europe, the LNG market has been facing challenges just as the continent appears poised to rely on it more than ever…

…Despite these many delays, some relief is in sight for 2025 with two major LNG expansions activating in the US. Cheniere’s Corpus Christi Stage 3 expansion produced its first cargo in February, adding 1.3 bcf/d of capacity. The first phase of Plaquemines LNG—built by the controversial firm Venture Global and itself a 1.3bcf/d facility—is in the commissioning process, a milestone celebrated last week by Chris Wright, Trump’s new Secretary of Energy…

…The one event that could significantly disrupt energy markets and pose a serious challenge to Brussels would be a major terrorist attack on European infrastructure. For example, if either of the large pipelines passing through Turkey were taken offline, prices would likely spike sharply. The loss of a large LNG import terminal, such as those in Spain or France, would also create severe strain. When operating on the edge, even small disturbances can knock the system out of equilibrium.

While Europe is playing an extremely risky game, absent the edge cases, it is likely to muddle through the next year without a full-blown disaster.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have a vested interest in Alphabet (parent of Google), Costco, Meta Platforms (parent of Facebook), and Microsoft. Holdings are subject to change at any time.

Is There Something Wrong With Pinduoduo’s Asset-Light Business Model?

The company’s gross property, plant, and equipment is barely sufficient to buy a laptop for each employee

Chinese e-commerce company Pinduoduo has experienced explosive growth in the past few years, with its revenue growing at a breath-taking pace of 142% per year from RMB 505 million in 2016 to RMB 247 billion in 2023. Profitability has also increased markedly over the same period, rising from a loss of RMB 322 million to a profit of RMB 60 billion. What’s even more impressive is that Pinduoduo has achieved this while being asset-light. The company ended 2023 with total assets of merely RMB 348 billion, which equates to a remarkable return on assets (profit as a percentage of total assets) of 17%. 

But I noticed two odd things about Pinduoduo’s asset-light nature as I dug deeper into the numbers. Firstly, Pinduoduo’s gross property, plant, and equipment per employee in 2023 was miles ahead of other large Chinese technology companies such as Alibaba, Meituan, and Tencent. This is shown in Table 1.

Table 1; Source: Company annual reports

Secondly – and the more important oddity here – was that Pinduoduo’s gross property, plant, and equipment per employee in 2017 was merely RMB 10,591, or RMB 0.01 million (gross property, plant, and equipment of RMB 12.3 million, and total employees of 1,159). According to ChatGPT, a professional laptop cost at least RMB 8,000 in China back in 2017, meaning to say, Pinduoduo’s gross property, plant, and equipment in that year was barely sufficient to purchase just a professional laptop for each employee. 

I’m not saying that something nefarious is definitely happening at Pinduoduo. But with the numbers above, I wonder if there’s something wrong with Pinduoduo’s purportedly asset-light business model. 

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Meituan and Tencent. Holdings are subject to change at any time.

What We’re Reading (Week Ending 16 March 2025)

The best articles we’ve read in recent times on a wide range of topics, including investing, business, and the world in general.

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 16 March 2025:

1. Want even tinier chips? Use a particle accelerator – The Economist

A more radical solution is to use a free-electron laser (FEL), where electrons travelling near the speed of light are manipulated to emit EUV radiation. The FEL process begins with a powerful electron gun that injects a beam of the particles into a miniature racetrack. The electrons then pass through a linear accelerator, which propels them to nearly the speed of light. Once accelerated, they enter a roughly 200-metre-long structure called an undulator, where a series of magnets generate a field whose polarity flips periodically. This wiggles the electrons, causing them to emit a beam of EUV photons with a specific wavelength.

Nicholas Kelez, the boss of xLight, a Silicon Valley startup developing FEL-based lithography, described the technology as a more powerful and tuneable “new light bulb” that he believes can be swapped into existing optical lithography machines. xLight expects to deliver the first commercial system within four years…

…Generating light using a FEL has some advantages over using lasers. The first is power: a lithography machine based on a FEL-based light source can be around six times more energy-efficient than a laser-plasma tool. Dispensing with molten-tin droplets also reduces the risk of contamination. Tuning such a machine for smaller wavelengths is also, at least theoretically, much easier: all that needs doing is tweaking the settings on the electron gun and the undulator. It would also be cheaper. A single FEL system can be repurposed to provide light for multiple lithography machines, allowing its operator to distribute the fixed costs across multiple chip-etching tools. Nakamura Norio from KEK estimates that the construction cost is around half that of a laser-based EUV tool and the running costs are around a fifteenth.

2. Is Manus a technological breakthrough or a marketing bubble? – Yuan Gan and Robert Wu

So, is Manus an empty marketing frenzy or a technological breakthrough comparable to DeepSeek? To answer this question, I reviewed every official example provided by Manus, the GAIA benchmark paper with its 467 questions, compared Manus’ performance with competitors on GAIA, and looked at the code of the “open-source versions of Manus.” Here are my findings:

  1. Is Manus an empty marketing frenzy or a technological breakthrough comparable to DeepSeek? Neither. Manus is not a marketing gimmick nor a fundamental technological revolution; it represents a breakthrough at the product level. Unlike DeepSeek, which focuses on a fundamental breakthrough in foundational model capabilities, Manus has made significant achievements in the direction of AI agents—reaching SOTA (State of the Art) levels on the authoritative evaluation metric GAIA, significantly ahead of peer products.
  2. Can the various open-source alternatives emerging these past few days replace Manus? No. The current open-source alternatives have a clear gap compared to Manus. Actual testing and data metrics show that Manus’ success rate in executing complex tasks far exceeds that of various open-source versions, by several times. Moreover, Manus has been specifically optimized for specific application scenarios, a fine-tuning that simple open-source replication cannot achieve.
  3. Is Manus a mature universal agent? No, Manus has not yet become a truly universal agent. To achieve this goal, it still needs to overcome three “major mountains”: a fundamental improvement in foundational model capabilities, a rich and diverse ecosystem of partners, and a solid and scalable engineering infrastructure…

…Whether it’s from the test experiences of netizens over the past few days or from the scores Manus has obtained on the GAIA benchmark, we can see that Manus and other universal AI agents are not yet mature.

So, how far away is a universal agent from being mature and commercially available?

I believe that to be mature and usable, it still needs to overcome three major challenges: foundational model capabilities, partner ecosystems, and engineering infrastructure.Currently, universal agents still rely on foundational large language models for task decomposition and execution, especially in the execution phase. In this phase, large language models face significant challenges in utilizing web information and computer operations…

…In the actual experience of OpenAI Operator, a significant issue is the restricted interaction between agents and external services. For example, when Operator accesses Reddit, GitHub, or other websites to complete tasks, it is often identified as abnormal traffic and blocked.

Currently, most agents access network services anonymously or with a generic identity, lacking a clear identity marker, leading to:

  • Being identified and blocked by websites’ anti-crawling mechanisms, including search engines like Google.
  • Inability to access services that require account login, such as obtaining information from Twitter and Facebook.
  • Inability to access personalized content and services, such as letting the agent view one’s own email…

…Unlike traditional internet services, which can often abstract services into “calling a microservice” as an instantaneous stateless operation, agent services are almost all long-duration, multi-state conversational interactions. After the product capabilities reach maturity, how to efficiently provide agent services to millions, or even tens of millions, of users is a significant engineering challenge.

3. ‘Project Osprey:’ How Nvidia Seeded CoreWeave’s Rise – By Cory Weinberg and Anissa Gardizy

Then Nvidia made a curious move: It agreed to spend $1.3 billion over four years to rent its own chips from CoreWeave, an upstart cloud computing firm trying to take on the giants. The deal, which CoreWeave dubbed “Project Osprey,” made Nvidia its second-largest customer last year after Microsoft, documents show, a detail not disclosed in CoreWeave’s filing last week for an initial public offering.

The deal shows how crucial Nvidia’s support has been to CoreWeave. The chip giant invested $100 million in the startup in early 2023. It funneled hundreds of thousands of high-end graphics processing units to CoreWeave. And it agreed to rent back its chips from CoreWeave through August 2027. Revenue from the deal represented 15% of CoreWeave’s sales last year…

…Prospective investors have been grappling with how long they can count on CoreWeave to sustain such big customers. Contracts with Microsoft and Nvidia, which together made up more than three-quarters of the company’s sales last year, expire between 2027 and 2029.

CoreWeave, meanwhile, has fueled its expansion with $8 billion in debt and $15 billion of long-term leases for data centers and office buildings, making it essential to attract new customers to supplement its original deals…

…It’s unclear exactly why Nvidia wanted to rent its own chips, but the company had several reasons to be interested in CoreWeave. Nvidia rents servers from cloud providers—including from CoreWeave—for its internal research and development teams.

Nvidia also rents servers from cloud providers for its DGX cloud computing service, which re-rents the servers to Nvidia cloud computing customers. CoreWeave has told investors it supports DGX…

…CoreWeave, in turn, purchased $380 million in chips and other hardware from Nvidia in 2023, documents shared with investors last year show. That spending total likely grew significantly last year. The company wrote in the prospectus that it purchases all of its chips from Nvidia, without specifying the total…

…The money Nvidia paid CoreWeave last year—in excess of $25 million a month, according to sales projections—was far more than what some of the other customers CoreWeave displayed in its IPO prospectus were spending. For example, CoreWeave cited model fine-tuning startup Replicate and quant trading firm Jane Street, which expected to spend hundreds of thousands of dollars a month with CoreWeave last year, internal materials show. 

4. Twitter thread on Diamond Online’s interview of Tatsuro Kiyohara – Japan Stock Feed

Interviewer: The Japanese stock market has been experiencing sharp declines. What’s your outlook for the next year?

Kiyohara: First and foremost, let me be clear—I don’t claim to have an accurate read on the market. The only time my predictions have been right is when investors lose their composure and panic. Over the past five years, I’ve only “gotten it right” twice—once during the COVID-19 crash in March 2020 and again on August 5, 2024. Those were the only times I felt truly confident and decided to aggressively buy Japanese stocks. Given how often I’ve been wrong, asking me for a market outlook is, quite frankly, insane…

…Interviewer: I see… but despite that, many individual investors would still love to hear your insights. After all, you’re the legendary salaryman investor who built an ¥80 billion fortune. Could you at least offer some guidance?

Kiyohara: If I told you, “The Nikkei will hit 42,000 by year-end,” what meaning would that have? That kind of prediction is pointless. But since that’s not a very helpful answer, let’s take a different approach. If I were still actively managing money today, what kind of portfolio would I hold? Let’s imagine what my positioning would look like…

…Kiyohara: If I were managing ¥10 billion ($66 million) today, my portfolio would be structured as follows:

▶️80% Long (¥8 billion in buy positions)

– ¥5 billion in small-cap stocks

– ¥2 billion in large-cap or mid-cap stocks

– ¥1 billion in REITs (Real Estate Investment Trusts)

▶️20% Short (¥2 billion in sell positions)

– Short positions in four highly liquid large-cap stocks

– (At the time of this interview, the market is already declining, so these aren’t necessarily the best picks, but as an example, I’d consider names like Advantest, Fujikura, Sanrio, and Mitsubishi Heavy Industries.)…

….Kiyohara: If you buy during a crash and think, “What if the market drops even further?”—you’re doing it wrong. “But that’s easier said than done,” you might say. Fair enough. So here’s how I train my mindset during downturns: “If I don’t buy today, then when?”…

…Interviewer: For more advanced investors, is an 80% long / 20% short portfolio a bullish position?

Kiyohara: Not necessarily.

– My core holdings are in net-cash-rich small-cap stocks, which means they are largely immune to overall market movements

– Higher interest rates are actually a positive for these companies, since they have strong balance sheets.

– In my view, this portfolio is neutral, reflecting a stance of “There could be a market drop, but it’s not a certainty.”…

…Kiyohara: I’ve said before that I’m terrible at predicting markets, but if we zoom out, I’m bullish on Japan. For years, Japan’s stock market was trapped in a box. Foreigners would buy, the market would rise. They’d sell, and the market would drop. It was a simple cycle. It was like a lottery—you had to ‘buy a ticket’ (stocks) to have a chance, but ultimately, you were at the mercy of foreign investors’ flows…

…For the first time, Japanese companies truly belong to their shareholders. That’s a massive structural shift—a revolution. Yes, risks remain. But this governance transformation is so significant that it outweighs them. That’s why, even if a crash comes, I focus on making money from the rebound rather than betting on the decline. Put simply: If share buybacks and dividend hikes continue, stock prices will rise.

5. What it will really take to feed the world – Bill Gates

Many discussions about feeding the world focus on increasing agricultural productivity through improved seeds, healthier soils, better farming practices, and more productive livestock (all priorities for the Gates Foundation). Vaclav, however, insists we already produce more than enough food to feed the world. The real challenge, he says, is what happens after the food is grown…

…Some of the world’s biggest food producers have the highest rates of undernourishment. Globally, we produce around 3,000 calories per person per day—more than enough to feed everyone—but a staggering one-third of all food is wasted. (In some rich countries, that figure climbs to 45 percent.) Distribution systems fail, economic policies backfire, and food doesn’t always go where it’s needed.

I’ve seen this firsthand through the Gates Foundation’s work in sub-Saharan Africa, where food insecurity is driven by low agricultural productivity and weak infrastructure. Yields in the region remain far lower than in Asia or Latin America, in part because farmers rely on rain-fed agriculture rather than irrigation and have limited access to fertilizers, quality seeds, and digital farming tools. But even when food is grown, getting it to market is another challenge. Poor roads drive up transport costs, inadequate storage leads to food going bad, and weak trade networks make nutritious food unaffordable for many families…

…While severe hunger has declined globally, micronutrient deficiencies remain stubbornly common, even in wealthy countries. One of the most effective solutions has been around for nearly a century: food fortification. In the U.S., flour has been fortified with iron and vitamin B since the 1940s. This simple step has helped prevent conditions like anemia and neural tube defects and improve public health at scale—close to vaccines in terms of lives improved per dollar spent…

…CRISPR gene editing, for instance, could help develop crops that are more resilient to drought, disease, and pests—critical for farmers facing the pressures of climate change. Vaclav warns that we can’t count on technological miracles alone, and I agree. But I also believe that breakthroughs like CRISPR could be game-changing, just as the Green Revolution once was…

…And some of these solutions aren’t about producing more food at all—they’re about wasting less of what we already have. Better storage and packaging, smarter supply chains, and flexible pricing models could significantly reduce spoilage and excess inventory. In a conversation we had about the book, Vaclav pointed out that Costco (which might seem like the pinnacle of U.S. consumption) stocks fewer than 4,000 items, compared to 40,000-plus in a typical North American supermarket.

That kind of efficiency—focusing on fewer, high-turnover products—reduces waste, lowers costs, and ultimately eases pressure on global food supply, helping make food more affordable where it is needed most.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have a vested interest in Costco and Microsoft. Holdings are subject to change at any time.

Company Notes Series (#7): iCAD

Editor’s note: This is the latest edition in the “Company Notes Series”, where we periodically share our notes on companies we’ve studied in the recent past but currently have no vested interest in (we may invest in or sell shares in the companies mentioned at any time). The notes are raw and not updated, and the “as of” date for the data is given at the start of the notes. The first six editions in the series can be found here, here, here, here, here, and here. Please give us your thoughts on the series through the “Contact Us” page; your feedback will determine if we continue with it. Thanks in advance!

Start of notes

Data as of 26 October 2024

Details on company

  • Stock ticker: ICAD
  • Listing exchange: NASDAQ
  • HQ: Nashua, New Hampshire, USA
  • Employee count at end-2023: 69 (67 full-time; 24 in sales and marketing; 16 in R&D)

iCAD’s business

  • iCAD is a global leader in AI-powered cancer detection whose laser focus is to create a world where cancer can’t hide.
  • iCAD provides the ProFound Breast Health Suite, which is a  solution for breast cancer detection, breast density assessment, evaluation of one- or two-year breast cancer risk, and evaluation of cardiovascular risk based on breast arterial calcifications. The ProFound Breast Health Suite is cleared by the US Food & Drug Administration (FDA), has received CE mark (for European regulations), and has Health Canada licensing. It is available in over 50 countries and iCAD estimates that it has been used for more than 40 million mammograms worldwide in the five years ended 2023. The ProFound Breast Health Suite contains four solutions: (1) ProFound Detection, (2) ProFound Density, (3) ProFound Risk, and (4) ProFound Heart Health.
  • ProFound Detection is clinically proven to improve breast cancer detection and radiologist performance. Currently at V3.0, ProFound Detection is built with the latest in deep-learning and 3rd generation artificial intelligence. It is used for 2D and 3D mammography screening. ProFound Detection rapidly and accurately analyzes each individual image or slice to identify potentially malignant lesions. Analyzing for masses, distortion, calcifications, and asymmetry, it localizes, segments and classifies lesions giving them a score and a case score for the overall exam. ProFound Detection is FDA cleared, CE marked, and Health Canada licensed. V4.0 of the solution is under review with the FDA.
  • ProFound Density provides an objective and consistent breast density assessment, helping clinics align to the new FDA-MQSA (Mammography Quality Standards Act) notification requirement to patients, which took effect in September 2024. Using mammographic images, ProFound Density analyzes a woman’s breast anatomy, measuring the adipose and fibroglandular tissue dispersion and texture, and categorizes her breast density within the appropriate BI-RADS 5th edition density category. ProFound Density gives clinicians an integrated workflow for identifying and reporting breast density, allowing for personalized patient planning with supplemental screening and customized schedules when needed. ProFound Density is FDA cleared, CE marked, and Health Canada licensed. The newest ProFound Density, V4.0, is under review by the FDA.
  • ProFound Risk reads 2D or 3D mammogram images to provide one-to-two year risk assessments for breast cancer. It is the first tool of its kind. ProFound Risk uses a new model for predicting breast cancer during an annual mammogram screening that has been found to be 2.4X more accurate compared to traditional life-time models based on family and medical history. By evaluating several data points in a patient’s scanned image, ProFound Risk calculates a more accurate short-term Risk Score for developing cancer in one or two years. ProFound Risk is CE marked, Health Canada licensed, and available for investigational use only in the US; ProFound Risk is under review by the FDA.
  • ProFound Heart Health measures the presence and extent of breast arterial calcifications from the same mammogram used to identify breast cancer, breast density and breast cancer risk. From one mammogram, clinicians assess the patient’s risk of heart disease and recommend further surveillance or review by other care teams. Breast arterial calcification assessment is pending regulatory licensing and available for investigational use only. iCAD’s Heart Health solution is under review by the FDA. Starting in October 2022, iCAD and Solis Mammography have been collaborating to develop and commercialize AI to evaluate cardiovascular disease based on breast arterial calcifications. Multiple studies have shown a correlation between the amount of breast arterial calcifications, which are visibly detectable on mammograms, to cardiovascular risk. iCAD and Solis are working together to use AI to quantify the amount of breast arterial calcification in mammograms, correlated to the risk of disease, and define meaningful clinical pathways for high-risk women.
  • iCAD has historically offered its solution as perpetually licensed software, primarily pre-installed on and sold with an iCAD configured, off-the-shelf computer, capable of optimally running the software. In 2022, iCAD began offering its full suite of breast AI solutions as a software package, without customers needing to purchase hardware; in the same year (in the first quarter of the year to be precise) iCAD also launched an on-premise, subscription model for its software. In the first quarter of 2024, iCAD began offering its solutions under a cloud-based, software-as-a-service (SaaS) model; iCAD’s SaaS product is powered by Google Cloud’s cloud computing infrastructure. Management’s current focus is on transitioning iCAD’s business model from a perpetual model to a recurring revenue model (this includes on-premise subscriptions and cloud-based SaaS).
  • iCAD is a US-centric business, with the USA making up 87% of total revenue in 2023. The remaining came from Europe (10%) and other geographies (3%).
  • In North America, iCAD sells its ProFound Breast Health Suite solutions solutions through a direct regional sales force and channel partners including OEMs, Radiology Picture Archiving and Communication System (PACS) vendors, AI Platform vendors, and distributors. In Europe and the Middle East, iCAD’s solutions are sold through a direct sales force and 22 reseller relationships with regional distributors:
    • iCAD’s OEM partners include GE Healthcare (focused on the manufacture and distribution of diagnostic imaging equipment), Fujifilm Medical Systems (a subsidiary of Fuji focused on the manufacture and distribution of X-rays and other imaging equipment) and Siemens Medical Systems.
    • iCAD’s PACS partnerships include Change Healthcare (a leading independent healthcare technology company focused on insights, innovation and accelerating the transformation of the US healthcare system) and Sectra (an international medical imaging IT solutions and cybersecurity company).
    • iCAD’s AI Platform vendors include Ferrum Health (works with global leaders of AI applications to provide a robust catalog of AI applications on a single, secure platform serving clinical service lines across healthcare enterprises), and Blackford (a wholly owned subsidiary of Bayer AG that provides a platform for integration of multiple disparate AI applications and algorithms with existing systems). 
  • iCAD has customer concentration risk. GE Healthcare is a major OEM (original equipment manufacturer) customer of iCAD, accounting for 22% of iCAD’s total revenues in 2022 and 2023. iCAD’s cancer detection products are also sold through OEM partners other than GE Healthcare; in total, all of iCAD’s OEM partners accounted for 32% of iCAD’s total revenue in 2023 and 29% of total revenue in 2022. iCAD also had one major direct customer that accounted for 8% of total revenue in 2023 and 4% in 2022. iCAD’s end-users appear to be radiologists working in facilities that conduct mammograms.
  • Prior to the third quarter of 2023, iCAD had two reporting segments: Detection and Therapy. The Detection segment houses the ProFound Breast Health Suite while the Therapy segment housed the Xoft business. On 23 October 2023, iCAD completed the sale of Xoft. Xoft was sold for US$5.76 million, and it is an electronic brachytherapy platform designed to deliver isotope-free (non-radioactive) radiation treatment in virtually any clinical setting without the limitations of radionuclides. So iCAD is currently left with only one reporting segment (Detection) and one product suite (ProFound Breast Health Suite).

Market opportunity

  • Incidences of breast cancer are growing. According to the World Health Organization, breast cancer is the most common cancer worldwide, with 2.26 million new cases diagnosed worldwide in 2020. One in eight women will get breast cancer in her lifetime, and every 14 seconds, a woman is diagnosed with breast cancer world-wide. When diagnosing breast cancer, early detection matters, as the cancer is more likely to respond to treatment and can result in greater survival rates.
  • According to the American Cancer Society, the relative 5-year survival rate from breast cancer is 99% when it is detected early, when the cancer is localized (no spread outside of the breast). When detected later, the 5-year survival rates drop, to 86% when the cancer has regional spread (to nearby areas outside the breast such as the lymph nodes), and to just 31% when it has distant spread (to areas such as the lungs, liver, and/or bones). iCAD calculates if diagnoses were shifted one stage earlier for 20% of the 280,000 women in the US diagnosed with breast cancer each year, there can be savings of approximately US$3.7 billion across 2-years of patient treatment and healthcare costs. But the problems with early-detection are:
    • 59% of women in the US miss their recommended screening mammograms, and for those who regularly screen for breast cancer, 20%-40% of cancers are missed in mammogram screenings with up to 50% missed in women with dense breast tissue.
    • Traditional risk assessment models have also relied on family history of the disease as a leading risk factor when in fact, and most surprising, 89% of women diagnosed with breast cancer have no direct family history of the disease and 90-95% are not related to inherited gene mutation.
  • AI can help radiologists spot cancer faster, with greater accuracy and save more lives. With the continuing migration from 2D FFDM (full-field digital mammography) reading systems to 3D DBT (digital breast tomosynthesis) systems, radiologists are spending twice the amount of time reading hundreds more images per 3D case compared to the four images captured with 2D (two images per breast). As a result, 50% of radiologists are overworked – and burnout is reported to be 49%. Simultaneously, false-positives and unnecessary recalls for suspected cancers have continued at similar rates while hard-to-detect interval cancers are being missed or diagnoses are delayed, leading to poor patient-experience.    
  • 40.5 million annual mammograms are conducted in the US across 8,834 certified facilities, as measured by the FDA Mammography Quality Standards Act (MQSA). Yet, only 37% of facilities are using a CAD (computer-aided diagnosis) or advanced AI mammography solution, according to Research & Markets United States Mammography and Breast Imaging Market Outlook Report 2022-2025. Of the 3,268 facilities using AI, iCAD has an active customer base of 1,488, or approximately 46% of the AI market, and approximately 17% of the total US market. Moreover, if more countries adopt the practice of each mammography exam being read by a single radiologist using AI, rather than the current practice in most countries of having two radiologists read each exam, this could be a tailwind for iCAD. Furthermore, some western European countries have already implemented, or are planning to implement, mammography screening programs, which may increase the number of screening mammograms performed in those countries.
  • iCAD’s large enterprise customers include Solis, Radiology Partners, SimonMed, Ascension and Cleveland Clinic, who collectively serve about 15% of the US mammography screening market. Some of them are in the early stages of rolling out iCAD’s solutions and continue to expand into more sites and markets each month. 
  • Globally, more than 31,000 mammography systems serve approximately 250 million women in the age range recommended for annual mammograms. Expanding to the 63% of the market that is not using AI, plus additional wins in the segment using AI but not ProFound, results in significant opportunity for new business for iCAD. The company has strengthened its sales team to take advantage of these opportunities.  
  • Data is the key to training robust machine learning and AI. In this regard iCAD is well positioned.  iCAD’s algorithms for ProFound Detection are trained on over 6 million images, including one of the largest 3D image datasets gathered from over 100 sites from around the globe. The training data includes the highest amount of sourcing from the US, providing diverse data that is ethnically, racially, and age representative of the US population. 
  • As it stands, iCAD’s ProFound Detection already performs better than the competition as shown in Figure 1. What’s missing in Figure 1 is that ProFound Detection also positively finds or predicts more cancers up to 2-3 years earlier by circling and scoring suspicious lesions for radiologists. As mentioned earlier, the latest version of ProFound Detection, V4.0, is under review with the FDA. V4.0 is built on the newest deep-learning neural network AI for breast cancer, density and risk. Per regulatory test data, iCAD has observed V4.0 will deliver significant improvements in specificity, sensitivity, and the highest AUC (area under the curve) for Specificity and Sensitivity for breast cancer detection at 92.5%.
Figure 1; Source: iCAD 2024 September investor presentation
  • iCAD appears to have strong partnerships. It has been partnering GE Healthcare for more than 20 years; in 2017 iCAD became the only integrated AI breast cancer detection solution in GE Healthcare’s systems; in November 2023, ProFound Detection and other iCAD solutions were integrated in GE Healthcare’s new MyBreastAI Suite offering. In November 2022, iCAD announced a strategic development and commercialization agreement with Google Health to integrate Google’s AI into iCAD’s breast imaging portfolio; iCAD expanded its partnership with Google Health in 2023 and signed a 20-year research and development agreement for the co-development, testing, and integration of Google’s AI technology with the ProFound Breast Health Suite for 2D mammography for worldwide commercialization to potentially ease radiologist workload and reduce healthcare disparities for women. iCAD is also working with Duke University, Indiana University, University of Pennsylvania and Karolinska Institute on artificial intelligence advancements and clinical testing. 
  • iCAD’s solutions are also interoperable with more than 50 PACS (Radiology Picture Archiving and Communication System) solutions worldwide. 
  • Early performance results from the first 30,000 ProFound cloud cases delivered an impressive processing time that’s over 50% faster compared to many traditional on-premise deployment solutions.

Management

  • Dana Brown, 59, was appointed president and CEO on 6 March 2023. Brown has been Chair of iCAD’s board since January 2023, and a director of iCAD since January 2022. Prior to being iCAD CEO, Brown was Strategic Advisor to Susan G. Komen, the world’s leading nonprofit breast cancer organization, and served as the organization’s Senior Vice President, Chief Strategy and Operations Officer from November 2018 to April 2022. Prior to Susan G. Komen, Brown served as Senior Vice President and Chief Digital Officer at United Way Worldwide, a global network of over 1,800 nonprofit fundraising affiliates. Brown was a founding member of several successful ventures; she co-founded and served as Chief Marketing Officer for MetaSolv Software (acquired by Oracle), and served as Chief Executive Officer of Ipsum Networks. Her background makes her well-qualified to lead iCAD’s transition from selling perpetual software licenses to on-premise subscriptions and cloud-based SaaS:
    • MetaSolv was an ERP (enterprise resource planning) system for telecom, media and internet providers. Brown helped MetaSolv become a category killer, consistently winning deals against many of the larger players of the time, including Lucent and Telcordia. MetaSolv was initially developed for a mainframe computing environment, and later became client server, and then SaaS. 
    • After MetaSolv, Brown became CEO of Ipsum Networks, a network performance management software company credited with pioneering route analytics technology. Ipsum Networks went on to be acquired by Cisco in its early stages.
    • After Ipsum Networks, Brown was on the turnaround teams of technology companies in the content delivery and mobile space, until 2011 or so, when she shifted to non-profit work by joining United Way Worldwide, then a US$5 billion global nonprofit. At United Way, Brown was its first Chief Digital Officer. She architected a co-development partnership between United Way and Salesforce to build a new cloud called Philanthropy Cloud, a platform that empowers corporations to put their values into action. Philanthropy Cloud extends each company’s reach by engaging its customers and employees in philanthropic endeavors, enhancing brand reputation and awareness, attracting and retaining top talent and delivering greater impact.
    • As of early-2023, Susan G. Komen has invested more than US$3.3 billion in groundbreaking breast cancer research, community health outreach, advocacy and programs in more than 60 countries. At Susan G. Komen, Brown led efforts to create MyKomen Health, the first patient care team and researcher engagement platform designed to guide patients through the end-to-end breast cancer journey with access to experts, resources and aid, as well as ShareForCures, the first ever patient-powered breast cancer research registry, storing information ranging from determinants of health, risk profiles, health status, including documentation of care deliveries such as diagnosis and interventions, all gathered from digital and nondigital EMR, EHR data as well as genomic data.
  • Brown has successfully led iCAD on a 3-phase transformation plan since becoming the CEO:
    • Phase 1 was to realign iCAD’s base. iCAD reduced cash burn which helped it to end the year in a strong cash position of US$21.7 million with no debt. In addition, iCAD made progress in its transition to a subscription-based recurring revenue business model.
    • Phase 2 was to strengthen iCAD’s foundation. iCAD’s brand changed from being product-focused into patient-focused, and the new branding was introduced at the annual, global Radiological Society of North America meeting in November 2023. Other parts of the Phase 2 transformation included iCAD’s aforementioned deals with Google Health and GE Healthcare in 2023, and the sale of Xoft.
    • Phase 3 is to invest in iCAD’s growth opportunities, where iCAD has a three-part, overlapping approach. The first part is to expand existing accounts; the second is to grow both direct and indirect channels; and the third is to enter new markets with new solutions, with an example being the ProFound Heart Health solution, which is currently available only for investigational use and is under review by the FDA.
  • Under Brown’s leadership, iCAD’s Total ARR (Total ARR stands for total annual recurring revenue and it consists of Maintenance Services ARR, Subscription ARR, and Cloud ARR) has steadily grown as shown in Table 1, especially Subscription ARR and Cloud ARR. The earliest data goes back to the third quarter of 2023, when iCAD started breaking down Total ARR.
  • Brown has also steadily grown the number of deals iCAD has signed for the subscription and cloud models for the ProFound Breast Health Suite, as shown in Table 2.
Table 1; Source: iCAD earnings call transcripts
Table 2; Source: iCAD earnings call transcripts
  • A major downside to Brown’s leadership of iCAD is lack of skin in the game. As of 22 April 2024, Brown controlled merely 238,334 shares of iCAD (and this includes 198,334 exercisable options under Brown’s name), which is worth just US$0.47 million at the 26 October 2024 stock price of US$1.96.
  • Another downside to Brown’s leadership is her compensation structure. It consists of (1) an annual base salary of US$400,000, (2) a cash and/or equity-based annual bonus of up to 30% of the annual base salary, and (3) options to acquire 250,000 shares of common stock, subject to a 3-year vesting schedule and 10-year expiration period. Brown is not incentivized on the long-term business results of iCAD.
  • The other senior members of iCAD’s management team are:
    • Chief Technology Officer, Jonathan Go. He joined iCAD in 2006 and became CTO in February 2019.
    • Chief Financial Officer, Eric Lonnqvist. He joined iCAD in February 2021 as Vice President of Financial Planning and Analysis, and became CFO in April 2023.

Financials

  • Given iCAD’s sale of Xoft in October 2023 and its ongoing shift to a recurring revenue business model from a perpetual license model, the company’s historical financials are not important. What is important is the growth of iCAD’s Total ARR, in particular, its Subscription ARR and Cloud ARR. As mentioned earlier, iCAD’s Total ARR and Subscription ARR have grown steadily in recent quarters, with Subscription ARR being a particular standout. It’s still early days for Cloud ARR with only one quarter’s worth of data, but management’s commentary in the 2024 second quarter earnings call is positive:

“While early days, the adoption of our SaaS offering has been going better than planned…

…There was conversations in Q1 even before cloud was readily available. But that being said, I do think the deals close quicker and the performance that Dana mentioned of the product has helped. Some of the customers want to test the product in the environment, and they did, and the results were positive. So some of the bigger deals were sped up and closed quicker.

And then other deals in that 10 count, those were some migrations that we want the customers that weren’t even thinking about cloud and their subscription — or their service contract ended and they go, “Well, this sounds interesting.” It moved kind of quickly some customers that wanted to get rid of hardware that they’re tired of buying these boxes every 3 years and they’re ready to move the technology forward and it just kind of clicked. But some of the bigger ones have been — there’ve been talks in Q1 also before these came together.

But going forward, I think that just because of all the positives Dana mentioned in the opening remarks, there’s going to be a natural push and it’s with how the product is performing and the excitement in customers when you talk to it and just the ease to use it compared to the perpetual model. We’re feeling a lot of pressure that it’s moving quicker in this direction…

…Yes. I do believe that cloud is going to grow faster than subscription. At what point in time its growth rate overtakes subscription is still a little bit TBD since we’ve only had it commercially available for one quarter. But the early indicators are positive. So yes, we do see that as where the future is, right, for iCAD and for our customers.”

  • iCAD’s balance sheet is strong; as of 30 June 2024, there was cash of US$20.4 million, zero debt, and total lease liabilities of just US$0.27 million. Management believes this level of cash is  sufficient to fund iCAD’s operations with no need to raise additional funding.
  • The company’s cash burn has also stabilised materially as shown in Figure 2.
Figure 2; Source: iCAD 2024 September investor presentation

Valuation (as of 26 October 2024)

  • 26 October 2024 stock price of US$1.96.
  • As of 30 June 2024, iCAD had issued shares of 26.540 million, and stock options issued of 3.048 million, for a fully diluted share count of 29.588 million. At the 26 October 2024 stock price, this gives a fully diluted market capitalisation of US$57.99 million.
  • iCAD’s Total ARR as of 30 June 2024 is US$9.2 million. This gives a price-to-ARR ratio of just 6.3. Assuming a 15% net profit margin at maturity, we’re looking at an implied price-to-earnings ratio of 42. Is this high or low? Total ARR has been growing pretty rapidly, and should grow even faster in the coming years with the very recent introduction of the cloud model. 

Risks

The important risks are, in no order of merit:

  • Customer concentration (discussed earlier in the “iCAD’s business” section of the notes)
  • Competition: Many competitors have significantly greater financial, technical, and human resources than iCAD and are well-established in the healthcare market. iCAD currently faces direct competition in its cancer detection and breast density assessment businesses from Hologic, Inc. (Marlborough, MA), Volpara Solutions Limited (Rochester, NY), ScreenPoint Medical (Nijmegen, Netherlands), Densitas Inc. (Halifax, Nova Scotia, Canada), Therapixel (Paris, France), and Lunit (Seoul, South Korea). I couldn’t find first-hand information, but recent research on iCAD done by an investment writer stated that GE Healthcare and Hologic are the two most important companies manufacturing DBT (digital breast tomosynthesis) machines, and that iCAD’s software is compatible with Hologic’s machines. The research also stated that Hologic accounts for nearly 70% of installed mammography units, so if Hologic removes iCAD’s compatibility, iCAD’s business could take a major hit. 
  • Near-term decline in financials: iCAD’s shift to a recurring revenue model will create temporarily lower GAAP revenue and negative cash flow because (1) revenue from subscription-based licenses are recognised ratably, and (2) cash is collected ratably compared to all up front in the perpetual license model. If iCAD’s financials decline at a time when financial conditions are tight in general, it could cause distress for the company. 

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have no vested interest in any company mentioned. Holdings are subject to change at any time.

What We’re Reading (Week Ending 09 March 2025)

The best articles we’ve read in recent times on a wide range of topics, including investing, business, and the world in general.

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 09 March 2025:

1. The Troubled Energy Transition – Daniel Yergin, Peter Orszag, and Atul Arya

The fundamental objective of the energy transition is to replace most of today’s energy system with a completely different system. Yet throughout history, no energy source, including traditional biomass of wood and waste, has declined globally in absolute terms over an extended period.

The first energy transition began in 1709, when a metalworker named Abraham Darby figured out that coal provided “a more effective means of iron production” than wood. And the ensuing “transition” took place over at least a century. Although the nineteenth century has been called “the century of coal,” the energy scholar Vaclav Smil has observed that coal did not overtake traditional biomass energy sources (such as wood and crop residues) until the beginning of the twentieth century. Oil, discovered in western Pennsylvania in 1859, would overtake coal as the world’s top energy source in the 1960s. Yet that did not mean that the absolute amount of coal used globally was falling—in 2024, it was three times what it had been in the 1960s.

The same pattern is playing out today. About 30 percent of the world’s population still depends on traditional biomass for cooking, and demand for hydrocarbons has yet to peak or even plateau. The portion of total energy usage represented by hydrocarbons has changed little since 1990, even with the massive growth in renewables. (In the same period, overall energy use has increased by 70 percent.) And the global population is expected to grow by approximately two billion in the coming decades, with much of that growth taking place in the global South. In Africa—a demographically young continent whose population has been projected to increase from 18 percent of the global population today to 25 percent by 2050—almost 600 million people live without electricity, and roughly one billion lack access to clean cooking fuel. Traditional biomass energy still fuels almost half the continent’s total energy consumption…

…Technological, policy, and geopolitical uncertainty makes it challenging to estimate the costs associated with achieving net zero by 2050. But one thing is certain: the costs will be substantial.

The most recent estimate comes from the Independent High-­Level Expert Group on Climate Finance, whose numbers provided a framework for the COP29 meeting—the UN’s annual forum on climate change—in Azerbaijan. It projected that the investment requirement globally for climate action will be $6.3 to $6.7 trillion per year by 2030, rising to as much as $8 trillion by 2035. It further estimated that the global South countries will account for almost 45 percent of the average incremental investment needs from now to 2030, and they have already been falling behind in meeting their financing needs, especially in sub-Saharan Africa.

Based on such estimates, the magnitude of energy-transition costs would average about five percent a year of global GDP between now and 2050. If global South countries are largely exempted from these financial burdens, global North countries would have to spend roughly ten percent of annual GDP—for the United States, over three times the share of GDP represented by defense spending and roughly equal to what the U.S. government spends on Medicare, Medicaid, and Social Security combined…

…In other words, achieving net zero will also require an unprecedented reorganization of capital flows from the global North to the global South, which will necessitate substantial investments in renewable-energy infrastructure at a time when, according to the International Monetary Fund, 56 percent of low-income countries are “at high levels of debt distress.” While innovative financing mechanisms (such as debt-for-climate and debt-for-nature swaps) will help, low sovereign-debt ratings throughout the developing world present a major obstacle to outside investment and raise capital costs. As a result, the bulk of the financial burden will be borne by advanced economies. But even there, debt has risen considerably—average public debt today is over 100 percent of GDP, a level not seen since World War II and a major constraint on governments’ ability to finance the transition through public spending…

…At the moment, almost half the population of the developing world—three billion people—annually uses less electricity per capita than the average American refrigerator does. As energy use grows, “carbonizing” will precede “decarbonizing.” Natural gas is a readily available option, and it’s a better alternative to coal, as well as to traditional biomass fuels that produce harmful indoor air pollution. Although global oil demand seems slated to plateau in the early 2030s, natural gas consumption is expected to continue to increase well into the 2040s. Production of liquefied natural gas is on track to increase by 65 percent by 2040, meeting energy security needs in Europe, replacing coal in Asia, and driving economic growth in the global South…

…The clash of priorities between the North and the South is especially striking when it comes to carbon tariffs. Many global North governments have, as part of their efforts to reduce emissions, put up barriers preventing other countries from taking the same carbon-based economic development path that they took to achieve prosperity. The European Union has launched the first phase of its Carbon Border Adjustment Mechanism. The CBAM is intended to support European climate objectives globally by initially imposing import tariffs on products such as steel, cement, aluminum, and fertilizer based on the carbon emissions embedded in their production and then expanding to more imports. Critics in the global North have argued that such measures would be ineffective because of the enormous complexity of supply chains and the associated difficulty of tracking embedded carbon in imports. Critics in the global South see the CBAM as a barrier to their economic growth. Ajay Seth, India’s economic affairs secretary, has argued that CBAM would force higher costs on the Indian economy: “With income levels which are one-twentieth of the income levels in Europe, can we afford a higher price? No, we can’t.”…

…The International Energy Agency has projected that global demand for the minerals needed for “clean energy technologies” will quadruple by 2040. At the top of the list are such critical minerals as lithium, cobalt, nickel, and graphite, as well as copper. Between 2017 and 2023 alone, demand for lithium increased by 266 percent; demand for cobalt rose by 83 percent; and demand for nickel jumped by 46 percent. Between 2023 and 2035, S&P expects the demand for lithium to increase by another 286 percent; cobalt, by 96 percent; and nickel, by 91 percent. Electric vehicles require two and a half to three times more copper than an internal combustion engine car; battery storage, offshore and onshore wind systems, solar panels, and data centers all require significant amounts of copper. S&P’s analysis of future copper demand found that global copper supply will have to double by the middle of the 2030s to meet current policy ambitions for net-zero emissions by 2050. This is extremely unlikely, considering that, based on S&P data that tracked 127 mines that have come online globally since 2002, it takes more than 20 years to develop a major new mine; in the United States, it takes an average of 29 years…

…China already has a dominant position in mining and a predominant position in the processing of minerals into metals essential for renewable energy infrastructure. It accounts for over 60 percent of the world’s rare-earth mining production (compared with nine percent for the United States) and more than 90 percent of the processing and refining of rare earths. It produces 77 percent of the world’s graphite, processes 98 percent of it, and processes over 70 percent of the world’s lithium and cobalt and almost half the copper.

Beijing aims to extend this dominance to what it calls the “global new energy industrial chain,” with its commanding position in batteries, solar panels, and electric vehicles, as well as in deploying massive amounts of capital toward energy infrastructure in the developing world. With China’s huge scale and low costs, Beijing describes this effort as an extensive and integrated approach to developing and dominating the renewable energy sector. From 2000 to 2022, it issued $225 billion in loans for energy projects in 65 strategically significant nations, with about 75 percent of that directed toward coal, oil, and gas development. Between 2016 and 2022, China provided more energy project financing around the world than any major Western-backed multilateral development bank, including the World Bank…

…Electrification trends suggest that power demand in the United States will double between now and 2050. Electricity consumption is already outpacing recent demand forecasts. PJM, which manages electricity transmission from Illinois to New Jersey, almost doubled its growth projection between 2022 and 2023 and is warning of the danger of shortfalls in electricity before the end of the decade…

…Today’s energy transition is meant to be fundamentally distinct from every previous energy transition: it is meant to be transformative rather than an additive. But so far it is “addition,” not replacement. The scale and variety of the challenges associated with the transition mean that it will not proceed as many expect or in a linear way: it will be multidimensional, proceeding at different rates with a different mix of technologies and different priorities in different regions. That reflects the complexities of the energy system at the foundation of today’s global economy. It also makes clear that the process will unfold over a long period and that continuing investment in conventional energy will be a necessary part of the energy transition.

2. An Interview with Benedict Evans About AI Unknowns – Ben Thompson and Benedict Evans

Well, you wrote about Deep Research a couple of weeks ago and you were pretty disappointed in the output. They used a smartphone report as the demo and it’s interesting, because the Deep Research case that convinced me was actually interview prep, and the key thing about it was, it was a lot of qualitative information that was helpful, and I wasn’t looking for quantitative information. Does that ring true of your experience?

BE: It does, yes. There’s a lot of different things one can say about this, and most of what I said was, it’s kind of interesting and puzzling rather than just, “It’s crap”. It’s very easy to say, “This is amazing and it changes the world, and it’s the fifth industrial revolution”, and it’s very easy to say, “This is all a bunch of crap and it’s the biggest waste of time and money since NFTs, please subscribe to my Substack”, and leave it at that. But what I struggle with is, it’s actually much more interesting and more complicated.

There’s a simple statement which is, “These things are good at things that don’t have wrong answers”. The quote someone used a couple of years ago was, “They tend to be good at things that computers are bad at and bad at things that computers are good at”, they’re very bad at precise specific information retrieval, which is what computers begin with. But on the other hand, you can ask them a question like, “What would be a good thing to take on a picnic?”, and that’s a question that a computer just couldn’t answer, that’s not a SQL query, and an LLM can answer that.

I think a lot of the product challenge and use case challenge around these things is trying to work out how you translate that into something that you’re actually trying to do. You gave the example of interview prep, which is — actually I don’t do interviews, but that would be something where, yeah, I can see that would be useful. The funny thing here is that OpenAI, I wasn’t testing it myself, I went and looked at OpenAI’s own product page, this is their test where they’re telling me this is what it’s for and this is what it’s good at, and proceed to show it to doing precise information retrieval, which of course, it can’t do. So just for the people who haven’t looked into OpenAI’s product page, it suggests some use cases, and one of them is, “Make a table of a bunch of countries with smartphone adoption by operating system””, and also stuff like, “Who wants to learn languages”.

The wrong report for the wrong guy.

BE: Yeah. The problem is, as many people may know, I used to be a telecoms analyst, so I looked at this and thought, “Okay, well let me have a look”. Problem one is, it used Statista, which is basically an SEO spam house that aggregates other people’s data. Saying “Source: Statista” at best is kind of saying “Source: a Google Search”, they’re not actually telling me what the source is and secondly, StatCounter, which tells you traffic. And I looked at this and I thought — I won’t monologue too long, I promise — I looked at this and I thought-

No, this is great. I completely agree with where you’re going.

BE: -there’s two questions here. The first is, is this model accurately working out what sources it should use? And then, is it getting the correct data from those sources?

And the question that OpenAI have posed is smartphone adoption. Well, what does that mean exactly? Are you asking me about unit sales? Are you asking about the install base? Are you asking me about usage? Are you asking me about outdoor usage? Because a use case that they propose actually was something like a translation app. Adoption isn’t any of those, it might be any of those, depending on context.

ChatGPT has come up with, number one, StatCounter, which is a metric of usage, not the install base and then, it’s come up with Statista, which is actually going to Kantar, which I think is doing install base, but I’m not sure, and those two come up with two different numbers, and the number that’s in ChatGPT, the Deep Research report, is a third number.

The thing that I thought about this was, you’ve asked this a probabilistic question, not a deterministic question. You’ve asked it a “What should I take on a picnic?”-type question, you haven’t asked it a precise database-y question where a computer would know the answer. You’ve asked it to work out what you want, and then you’ve asked it to work out where to get it from, and then you’ve asked it to do the database retrieval and actually report what’s on those pages. It’s kind of done okay at the first two, or it’s done what I would expect an intern to do on the first two, and as I wrote, if I had an intern, I would’ve said, “This is why you wouldn’t use either of those two”.

Yeah.

BE: But an intern wouldn’t know, and then it’s copied the number down wrong, which is where you smack the intern on the back of the head and tell them to go back and do it again. There’s a lot of different ways you can talk about this. Are you using these things for precise information retrieval? Are you using them for things that don’t really have wrong answers? Are you using them for qual or for quant? Are you using them for brainstorming? How do you work out what your problem is and whether it would be good at that, and how it would map against that?

But at the end of the day, I went and asked it to do a thing and it told me it had done the thing, and it’s wrong. In fact, it’s worse than that. OpenAI asked it to do the thing and it did the thing, and it was wrong.

And then put it up as a demo!

BE: There’s this really profoundly important thing here, which I had this feeling looking at the new Claude model today as well, or yesterday, is people talk about these models getting better a lot, but if you’ve given me a table with 20 entries and some of them are wrong, what do you mean when you say the model’s better? Do you mean that all the odd entries are right? Or do you mean that the entries are more likely to be right? Those are very different things. The idea that we would get to these models to the point that they always would be right and you would know that in a way that you would know a database would be right, we have no idea if that’s possible. That’s a profound scientific question in the field. What do you do with something that can do amazing things but you don’t know if it’s right?…

…You made this point, this tool’s the most helpful if you are already an expert because it saves you time and you can identify the errors. But if you’re not an expert, it’s incredibly dangerous. I wrote about this a couple of weeks ago, I asked for this report on an industry I happen to know a lot about, and Deep Research completely missed a big player because it was a private company, there was no listings about it even though anyone in the field would have known about it. You now have an unknown known, there’s something that you think you know, but actually, you don’t know. You’ve been convinced to be more ignorant.

BE: Yeah, listening to you talk, I’m reminded, actually, of one of these very old fallacies from engineers, which is to say, “The users have to learn how it works”, which is the thing you see with open source over and over again or, “The users will have to learn”. Of course you can say that, but the users won’t learn and it’s not the user’s job to learn how it works. The more that you force people to have to understand how it works, the more you limit the potential adoption of it.

Zooming back a little bit, something that I have in draft at the moment is that, if we think about where we’ve come in the last whatever it is, two and a bit years since GPT 3.5 launched, at the beginning of 2023, say, you could think there was a cluster of questions that would determine what was going to happen. How much will these scale? How big will the models get? How expensive will it be? What will happen to the error rates? What will reasoning be? Are there barriers to entry? Are there winner-takes-all effects? Is there enough data? You can make a list of a dozen questions, they all kind of link together.

Out of that there’s one possible output which is there’s one computer that runs the whole world, and the other extreme is it ends up being like databases or indeed like machine learning in which if you were to say today, “How many databases are there?”, that’s just a meaningless question. What are you talking about?

Since then, none of those questions have really been answered, except for the extent that it seems clear right now that these things are going to be commodities, although still quite expensive commodities. But anyone who’s got a billion dollars can have one, you don’t need a hundred billion dollars, and there’s not a lot of winner-takes-all effect the way there was with smartphones. Anyone who’s got $500 million or $100 million dollars, or something, pick a number, can have an outlet, can have a frontier model. But all the other questions, we don’t know. We don’t know what will happen to the error rate. We don’t know how big the models will get or how long the scaling works.

One of the things that kind of came out of that was, there’s a path that says, you can just go to ChatGPT and say, “I want to move to Taiwan, how do I do that?”, “I need to file my taxes in New York, London, and California, do it for me”. And the model can go and read the right websites and ask you for a photograph of your bank statement and just make you the PDF and do it for you…

…BE: And suddenly, there’s this explosion of complexity in the data centers, and we have to know about it. You have to know chips, and there are all these papers. But I did this deck, I do this annual presentation, and I ended my presentation, the section that talked about scaling, with a quote from the Godfather that says, “If history teaches you anything, it’s that you can kill anybody”, and I crossed out, “Kill anybody” and said, “Commodity computing tends to get cheap”. You’ve got all of this complexity in the creation of the models and the data center and everything else, and yet, I don’t know, I look at Grok and I think, okay, in less than two years, you managed to produce a state-of-the-art model. Is that really, really good or really, really bad?

That’s bearish for model creation.

BE: That is not as positive as you think. Yes, they’re a great team, and well done for building a 100,000 GPU cluster, but what this tell us is it’s a commodity…

…BE: There’s no difference, that’s what that means. But yeah, I think you can extend this to the doomers, where it was clear that the people who were thinking this stuff is going to take over the world in the next three months just had no conception of how the world worked outside of their shared group house in the Berkeley Hills. The puzzle and the analogy I always used to give, looking at, going back to talking about use cases, is imagine the first people seeing VisiCalc, the first spreadsheet in the late ’70s.

Yep.

BE: So if you saw this and you were an accountant, it blew your mind because the line is like you change the interest rate here and all the other numbers on the spreadsheet change.

Yep. [John] Gruber and I were just talking about this the other day, and you could watch it change!

BE: You say that now and people now are like, “Yes…?”, but back then you did spreadsheets on paper with a pencil and so if you’re an accountant, you have to have this. Certainly, you can look up the pricing of the Apple II that you needed to run VisiCalc, the full setup with a floppy drive and a printer and a monitor was like 15 grand adjusted for inflation. But if you were a lawyer and you see it, you think, “Well, that’s great, my accountant should see this, but that’s not what I do all day”.

Yep.

BE: Now, show me a word processor that can do word counts and footnotes and line numbers. That, I will pay for, that solves a problem that I have. And the challenge of the text box and the blinking cursor is either you really know you’ve got a problem that it solves, which is coding and marketing, or you’re the kind of person that’s instinctively looking for tools to solve things in their company, which is the bottom-up IT adoption and it’s no coding and everything else, but that’s a very small portion of the population.

And then it’s everybody else who didn’t see that they had that problem until an enterprise SaaS company came and showed them that they were spending three hours a week on it and sold them something for 10 grand a seat to fix it. Otherwise, you’ve got this prompt. What do I do with it?

I completely agree with you and this is where one of the analogies I’ve been thinking about is going back to the first — arguably the greatest direct job displacement in IT was actually the first mainframes, where it’s just like, “Okay, we don’t need an entire backroom of bookkeepers, we don’t need an entire backroom of ERP trackers, we can just do it on a computer”, and it was like a one-to-one replacement. What’s interesting about this is right now, AI is a bottoms-up phenomenon because you need so much agency to go out and find a way to use this, and because the model doesn’t learn, you have to learn how to use the model. It’s like people wanting to bring PCs to work in the late ’70s, and it’s like, “What are you trying to do here?”.

BE: And if you look at what these people were doing, all the books and magazines at the time were, “You should learn to code. Or at a minimum, you need to buy a database software program”. So it wasn’t you buy Quicken, it’s you buy a database plus software program and you make your own Quicken by yourself.

That’s such a great point, it’s the same thing it was with code. That was the thing at the beginning and you can do something yourself, it’s an excellent point. But it does make me think that if you’re a top-down decision maker, you can, number one, decide that the cost-benefit is worth actually removing an entire category or entire department. And number two, you can tolerate the error rate because you’ll do a cost-benefit analysis that says, “Okay, I’m at X percent error rate, this is going to cost me Y amount of money. How does that balance versus this collection of humans who are also going to make X amount of errors and cost me Y amount of money?” And you go in and you just shift people out and that is actually going to be a more meaningful change or where it’s going to start. It’s going to be less about getting people to fundamentally change how they work and more about, “You know what? We can do it good enough with AI for this whole category of work. Sorry, you guys are laid off”.

BE: Yeah, I think it depends. And if you go back and think about how the last waves of automated enterprise software worked, whether it’s SaaS or on-prem or whatever, there’s two or three things here. So one of them is you don’t just replace a whole department with one piece of software.

No, not now. I’m talking about back in the ’70s.

BE: Yeah. But the other part of that is that didn’t result in fewer white-collar workers or fewer accountants. Excel didn’t result in fewer accountants. I mean, my joke was always that, young people won’t believe this, but before Excel, investment bankers worked really long hours. Now they can get their work finished at lunchtime on Fridays and go home to The Hamptons. And of course, that’s not what happened.

Well, why is it that that isn’t what happened? I think, and it comes back to what I was saying earlier, the base case here is that you work in invoice processing and now you have a new piece of software that’s better at resolving failed invoice payments, and that is worth X-hundred thousand dollars a year to your company, and so they come out and they buy this piece of software. Or it’s a feature that gets added to their existing software or it plugs into SAP or Salesforce or Oracle or whatever it is, and it’s that piece of software and today, the typical big company has 400 or 500 SaaS applications, maybe more. And the HR team, account department has 45 different applications, there’s the German tax planning thing, and there’s the thing that manages stock options, and there’s the thing that manages training for new recruits, and the thing that makes sure that everybody’s taking their compliance exam, and the other thing that makes sure everyone’s done their compliance exam, and the training thing, and you just keep adding these all up. And they all find value and they all find budget and they all move costs from one place to another and the big company has 400 or 500 other new things.

The base case is that generative AI will be another 400 or 500 of these, and it will replace half of them, and it will double the value of the other half. 250 of them will get killed and 250 of them will get a bunch of new features and there’ll be another 250 of them on top, and now you’ll have 750 or 1,000 new applications in this company and there will be bits of LLM scattered all the way through them just as there’s bits of machine learning scattered all the way through them, just as there’s databases scattered all the way through them. It will just be kind of a building block in making software. Does that mean that there’s less net employment or does that just mean that a bunch of jobs go away and get replaced by new jobs?…

…BE: If you’re in the accounting industry, this changes the whole nature of your industry, so that’s one observation.

The second observation would be — yes, you don’t know what those changes will be, you don’t know how it is, what all the new things it will be, you start by doing what you already know you need to do. And then over time, you realize there’s new things you can do with this, which is your point about feeds, and you could say the same thing about the emergence of Instagram and online dating, all the stuff that happened that wasn’t obvious at the time.

However, I think there’s a completely opposite point, which is equally interesting, about how new this is or how different this is, which is that if you’re looking at the Internet in 1995, you kind of knew how fast computers were and how fast they’d be in like five years, and you knew how many people had broadband, and you had a pretty good sense — you could plot a line on a chart of how many people are going to have broadband and how fast, on a three to five-year view and you kind of knew how much PCs cost, and you knew what annual PC sales were, and you could make a guess that, okay, this is going to mean that PC sales quadruple and everyone will go out and buy a PC, which is more or less what happened, and you knew how many middle-class households there were in America, so you kind of knew what the PC market could, in principle, do. The same thing with smartphones, you knew how fast 3G was, you knew how fast 4G was, you knew how fast the chips were.

What I was getting at is, with LLMs, we don’t know that. We don’t know what that roadmap is for the fundamental technical capabilities of the thing, which is different to anything from the web to flight or cars. We weren’t looking at a smartphone and saying, “Well, this is where we are today, but maybe in two or three years, it will be able to unroll and fill the whole wall, or maybe it’ll have a year’s battery life”. You kind of knew what the really basic core physics constraints were, and we don’t know that for this stuff.

Well, especially with this accuracy point. Daniel Gross made this point a few weeks ago too, I think it’s really profound, that there’s just a really stark fundamental difference between 100% accuracy and 99% accuracy.

BE: Well, this is a problem with saying “better” models. What do you mean “better”? Do you mean it was 82 and now it’s 83? Or do you mean it was 80 and now it’s 100 and it will always be 100? That’s a completely different thing.

3. Reality has a surprising amount of detail – John Salvatier

It’s tempting to think ‘So what?’ and dismiss these details as incidental or specific to stair carpentry. And they are specific to stair carpentry; that’s what makes them details. But the existence of a surprising number of meaningful details is not specific to stairs. Surprising detail is a near universal property of getting up close and personal with reality.

You can see this everywhere if you look. For example, you’ve probably had the experience of doing something for the first time, maybe growing vegetables or using a Haskell package for the first time, and being frustrated by how many annoying snags there were. Then you got more practice and then you told yourself ‘man, it was so simple all along, I don’t know why I had so much trouble’. We run into a fundamental property of the universe and mistake it for a personal failing.

If you’re a programmer, you might think that the fiddliness of programming is a special feature of programming, but really it’s that everything is fiddly, but you only notice the fiddliness when you’re new, and in programming you do new things more often.

You might think the fiddly detailiness of things is limited to human centric domains, and that physics itself is simple and elegant. That’s true in some sense – the physical laws themselves tend to be quite simple – but the manifestation of those laws is often complex and counterintuitive…

…The more difficult your mission, the more details there will be that are critical to understand for success.

You might hope that these surprising details are irrelevant to your mission, but not so. Some of them will end up being key. Wood’s tendency to warp means it’s more accurate to trace a cut than to calculate its length and angle. The possibility of superheating liquids means it’s important to use a packed bed when boiling liquids in industrial processes lest your process be highly inefficient and unpredictable. The massive difference in weight between a rocket full of fuel and an empty one means that a reusable rocket can’t hover if it can’t throttle down to a very small fraction of its original thrust, which in turn means it must plan its trajectory very precisely to achieve 0 velocity at exactly the moment it reaches the ground.

You might also hope that the important details will be obvious when you run into them, but not so. Such details aren’t automatically visible, even when you’re directly running up against them. Things can just seem messy and noisy instead. ‘Spirit’ thermometers, made using brandy and other liquors, were in common use in the early days of thermometry. They were even considered as a potential standard fluid for thermometers. It wasn’t until the careful work of Swiss physicist Jean-André De Luc in the 18th century that physicists realized that alcohol thermometers are highly nonlinear and highly variable depending on concentration, which is in turn hard to measure.

You’ve probably also had experiences where you were trying to do something and growing increasingly frustrated because it wasn’t working, and then finally, after some time you realize that your solution method can’t possibly work.

Another way to see that noticing the right details is hard, is that different people end up noticing different details…

…Before you’ve noticed important details they are, of course, basically invisible. It’s hard to put your attention on them because you don’t even know what you’re looking for. But after you see them they quickly become so integrated into your intuitive models of the world that they become essentially transparent. Do you remember the insights that were crucial in learning to ride a bike or drive? How about the details and insights you have that led you to be good at the things you’re good at?

This means it’s really easy to get stuck. Stuck in your current way of seeing and thinking about things. Frames are made out of the details that seem important to you. The important details you haven’t noticed are invisible to you, and the details you have noticed seem completely obvious and you see right through them. This all makes makes it difficult to imagine how you could be missing something important…

…If you’re trying to do impossible things, this effect should chill you to your bones. It means you could be intellectually stuck right at this very moment, with the evidence right in front of your face and you just can’t see it.

This problem is not easy to fix, but it’s not impossible either. I’ve mostly fixed it for myself. The direction for improvement is clear: seek detail you would not normally notice about the world. When you go for a walk, notice the unexpected detail in a flower or what the seams in the road imply about how the road was built. When you talk to someone who is smart but just seems so wrong, figure out what details seem important to them and why. In your work, notice how that meeting actually wouldn’t have accomplished much if Sarah hadn’t pointed out that one thing. As you learn, notice which details actually change how you think.

If you wish to not get stuck, seek to perceive what you have not yet perceived.

4. America’s Growing Trade Deficit Is Selling the Nation Out From Under Us. Here’s a Way to Fix the Problem – And We Need to Do It Now – Warren Buffett

Take a fanciful trip with me to two isolated, side-by-side islands of equal size, Squanderville and Thriftville. Land is the only capital asset on these islands, and their communities are primitive, needing only food and producing only food. Working eight hours a day, in fact, each inhabitant can produce enough food to sustain himself or herself. And for a long time that’s how things go along…

…Eventually, though, the industrious citizens of Thriftville decide to do some serious saving and investing, and they start to work 16 hours a day. In this mode, they continue to live off the food they produce in the eight hours of work but begin exporting an equal amount to their one and only trading outlet, Squanderville.

The citizens of Squanderville are ecstatic about this turn of events, since they can now live their lives free from toil but eat as well as ever. Oh, yes, there’s a quid pro quo – but to the Squanders, it seems harmless: All that the Thrifts want in exchange for their food is Squanderbonds (which are denominated, naturally, in Squanderbucks).

Over time Thriftville accumulates an enormous amount of these bonds, which at their core represent claim checks on the future output of Squanderville. A few pundits in Squanderville smell trouble coming. They foresee that for the Squanders both to eat and pay off – or simply service – the debt they’re piling up will eventually require them to work more than eight hours a day…

…Meanwhile, the citizens of Thriftville begin to get nervous. Just how good, they ask, are the IOUs of a shiftless island? So the Thrifts change strategy: Though they continue to hold some bonds, they sell most of them to Squanderville residents for Squanderbucks and use the proceeds to buy Squanderville land. And eventually the Thrifts own all of Squanderville.

At that point, the Squanders are forced to deal with an ugly equation: They must now not only return to working eight hours a day in order to eat—they have nothing left to trade—but must also work additional hours to service their debt and pay Thriftville rent on the land so imprudently sold. In effect, Squanderville has been colonized by purchase rather than conquest.

It can be argued, of course, that the present value of the future production that Squanderville must forever ship to Thriftville only equates to the production Thriftville initially gave up and that therefore both have received a fair deal. But since one generation of Squanders gets the free ride and future generations pay in perpetuity for it, there are—in economist talk—some pretty dramatic “intergenerational inequities.”…

…Sooner or later the Squanderville government, facing ever greater payments to service debt, would decide to embrace highly inflationary policies—that is, issue more Squanderbucks to dilute the value of each. After all, the government would reason, those irritating Squanderbonds are simply claims on specific numbers of Squanderbucks, not on bucks of specific value. In short, making Squanderbucks less valuable would ease the island’s fiscal pain.

That prospect is why I, were I a resident of Thriftville, would opt for direct ownership of Squanderville land rather than bonds of the island’s government. Most governments find it much harder morally to seize foreign-owned property than they do to dilute the purchasing power of claim checks foreigners hold. Theft by stealth is preferred to theft by force…

…The time to halt this trading of assets for consumables is now, and I have a plan to suggest for getting it done. My remedy may sound gimmicky, and in truth it is a tariff called by another name. But this is a tariff that retains most free-market virtues, neither protecting specific industries nor punishing specific countries nor encouraging trade wars. This plan would increase our exports and might well lead to increased overall world trade. And it would balance our books without there being a significant decline in the value of the dollar, which I believe is otherwise almost certain to occur.

We would achieve this balance by issuing what I will call Import Certificates (ICs) to all U.S. exporters in an amount equal to the dollar value of their exports. Each exporter would, in turn, sell the ICs to parties—either exporters abroad or importers here—wanting to get goods into the U.S. To import $1 million of goods, for example, an importer would need ICs that were the byproduct of $1 million of exports. The inevitable result: trade balance.

Because our exports total about $80 billion a month, ICs would be issued in huge, equivalent quantities—that is, 80 billion certificates a month—and would surely trade in an exceptionally liquid market. Competition would then determine who among those parties wanting to sell to us would buy the certificates and how much they would pay. (I visualize that the certificates would be issued with a short life, possibly of six months, so that speculators would be discouraged from accumulating them.)

For illustrative purposes, let’s postulate that each IC would sell for 10 cents—that is, 10 cents per dollar of exports behind them. Other things being equal, this amount would mean a U.S. producer could realize 10% more by selling his goods in the export market than by selling them domestically, with the extra 10% coming from his sales of ICs.

In my opinion, many exporters would view this as a reduction in cost, one that would let them cut the prices of their products in international markets. Commodity-type products would particularly encourage this kind of behavior. If aluminum, for example, was selling for 66 cents per pound domestically and ICs were worth 10%, domestic aluminum producers could sell for about 60 cents per pound (plus transportation costs) in foreign markets and still earn normal margins. In this scenario, the output of the U.S. would become significantly more competitive and exports would expand. Along the way, the number of jobs would grow…

…To see what would happen to imports, let’s look at a car now entering the U.S. at a cost to the importer of $20,000. Under the new plan and the assumption that ICs sell for 10%, the importer’s cost would rise to $22,000. If demand for the car was exceptionally strong, the importer might manage to pass all of this on to the American consumer. In the usual case, however, competitive forces would take hold, requiring the foreign manufacturer to absorb some, if not all, of the $2,000 IC cost.

There is no free lunch in the IC plan: It would have certain serious negative consequences for U.S. citizens. Prices of most imported products would increase, and so would the prices of certain competitive products manufactured domestically. The cost of the ICs, either in whole or in part, would therefore typically act as a tax on consumers.

That is a serious drawback. But there would be drawbacks also to the dollar continuing to lose value or to our increasing tariffs on specific products or instituting quotas on them—courses of action that in my opinion offer a smaller chance of success. Above all, the pain of higher prices on goods imported today dims beside the pain we will eventually suffer if we drift along and trade away ever larger portions of our country’s net worth.

I believe that ICs would produce, rather promptly, a U.S. trade equilibrium well above present export levels but below present import levels. The certificates would moderately aid all our industries in world competition, even as the free market determined which of them ultimately met the test of “comparative advantage.”

This plan would not be copied by nations that are net exporters, because their ICs would be valueless. Would major exporting countries retaliate in other ways? Would this start another Smoot-Hawley tariff war? Hardly. At the time of Smoot-Hawley we ran an unreasonable trade surplus that we wished to maintain. We now run a damaging deficit that the whole world knows we must correct…

…The likely outcome of an IC plan is that the exporting nations—after some initial posturing—will turn their ingenuity to encouraging imports from us. Take the position of China, which today sells us about $140 billion of goods and services annually while purchasing only $25 billion. Were ICs to exist, one course for China would be simply to fill the gap by buying 115 billion certificates annually. But it could alternatively reduce its need for ICs by cutting its exports to the U.S. or by increasing its purchases from us. This last choice would probably be the most palatable for China, and we should wish it to be so.

If our exports were to increase and the supply of ICs were therefore to be enlarged, their market price would be driven down. Indeed, if our exports expanded sufficiently, ICs would be rendered valueless and the entire plan made moot. Presented with the power to make this happen, important exporting countries might quickly eliminate the mechanisms they now use to inhibit exports from us.

5. The hidden cost of AI: Trading long-term resilience for short-term efficiency – Eric Markowitz

AI is the latest in a long lineage of efficiency-maximizing tools. It promises to make research instantaneous, strip away uncertainty, and optimize everything from hiring to investment analysis. But for all the gains in speed and precision, we rarely stop to ask: What are we losing in return? Because there is no such thing as a free lunch…

…AI makes the world feel more scientific than ever. It can generate business strategies, write persuasive emails, and surface patterns invisible to human analysis. But the most important decisions — the ones that lead to breakthroughs, revolutions, and paradigm shifts — are rarely the result of pure data analysis.

Some of the best ideas in history looked irrational at first. They required deep research, yes — but more importantly, they required taste. (And perhaps a bit of luck.)

Taste is an underrated concept in a world obsessed with efficiency. It’s the ability to recognize something valuable before the numbers prove it. The ability to see beyond spreadsheets and sentiment analysis and understand how an idea actually fits into the world. If everyone has access to the same AI-generated insights, the only thing that remains scarce is independent thinking. And that is precisely where the edge lies.

None of this is an argument against AI. It’s an argument for knowing what not to outsource to our robot overlords. AI is a tool. A powerful one. But it is not a substitute for intuition, nor a replacement for deep thinking. The institutions and ideas that endure the longest are those that understand what to hold onto even as the world around them changes. 

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have a vested interest in Alphabet (parent of Google). Holdings are subject to change at any time.

The Latest Thoughts From American Technology Companies On AI (2024 Q4) – Part 2

A collection of quotes on artificial intelligence, or AI, from the management teams of US-listed technology companies in the 2024 Q4 earnings season.

The way I see it, artificial intelligence (or AI), really leapt into the zeitgeist in late-2022 or early-2023 with the public introduction of DALL-E2 and ChatGPT. Both are provided by OpenAI and are software products that use AI to generate art and writing, respectively (and often at astounding quality). Since then, developments in AI have progressed at a breathtaking pace.

With the latest earnings season for the US stock market – for the fourth quarter of 2024 – coming to its tail-end, I thought it would be useful to collate some of the interesting commentary I’ve come across in earnings conference calls, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. This is an ongoing series. For the older commentary:

I’ve split the latest commentary into two parts for the sake of brevity. This is Part 2, and you can Part 1 here. With that, I’ll let the management teams take the stand… 

Microsoft (NASDAQ: MSFT)

Microsoft’s management is seeing enterprises move to enterprise-wide AI deployments 

Enterprises are beginning to move from proof of concepts to enterprise-wide deployments to unlock the full ROI of AI. 

Microsoft’s AI business has surpassed an annual revenue run rate of $13 billion, up 175% year-on-year; Microsoft’s AI business did better than expected because of Azure, Microsoft Copilot (within Copilot, price per seat was a strength and still retains good signal for value)

Our AI business has now surpassed an annual revenue run rate of $13 billion, up 175% year-over-year…

…[Question] Can you give more color on what drove the far larger-than-expected Microsoft AI revenue? We talked a bit about the Azure AI component of it. But can you give more color on that? And our estimates are that the Copilot was much bigger than we had expected and growing much faster. Any more details on the breakdown of what that Microsoft AI beat would be great.

[Answer] A couple of pieces to that, which you correctly identified, number one is the Azure component we just talked about. And the second piece, you’re right, Microsoft Copilot was better. And what was important about that, we saw strength both in seats, both new seats and expansion seats, as Satya talked about. And usage doesn’t directly impact revenue, but of course, indirectly does as people get more and more value added. And also price per seat was actually quite good. We still have a good signal for value.

Microsoft’s management is seeing AI scaling laws continue to show up in both pre-training and inference-time compute, and both phenomena have been observed internally at Microsoft for years; management has seen gains of 2x in price performance for each new hardware generation, and 10x for each new model generation

AI scaling laws continue to compound across both pretraining and inference time compute. We ourselves have been seeing significant efficiency gains in both training and inference for years now. On inference, we have typically seen more than 2x price performance gain for every hardware generation and more than 10x for every model generation due to software optimizations. 

Microsoft’s management is balancing across training and inference in the buildout of Microsoft’s AI capacity; the buildout going forward will be governed by revenue growth and capability growth; Microsoft’s Azure data center capacity is expanding in line with both near-term and long-term demand signals; Azure has more than doubled its capacity in the last 3 years, and added a record amount of capacity in 2024; Microsoft’s data centres uses both in-house as well as 3rd-party chips

Much as we have done with the commercial cloud, we are focused on continuously scaling our fleet globally and maintaining the right balance across training and inference as well as geo distribution. From now on, it’s a more continuous cycle governed by both revenue growth and capability growth thanks to the compounding effects of software-driven AI scaling laws and Moore’s Law…

…Azure is the infrastructure layer for AI. We continue to expand our data center capacity in line with both near-term and long-term demand signals. We have more than doubled our overall data center capacity in the last 3 years, and we have added more capacity last year than any other year in our history. Our data centers, networks, racks and silicon are all coming together as a complete system to drive new efficiencies to power both the cloud workloads of today and the next-generation AI workloads. We continue to take advantage of Moore’s Law and refresh our fleet as evidenced by our support of the latest from AMD, Intel, NVIDIA, as well as our first-party silicon innovation from Maia, Cobalt, Boost and HSM.

Microsoft’s management is seeing growth in raw storage, database services, and app platform services as AI apps scale, with an example being Azure OpenAI apps that run on Azure databases and Azure App Services

We are seeing new AI-driven data patterns emerge. If you look underneath ChatGPT or Copilot or enterprise AI apps, you see the growth of raw storage, database services and app platform services as these workloads scale. The number of Azure OpenAI apps running on Azure databases and Azure App Services more than doubled year-over-year, driving significant growth in adoption across SQL, Hyperscale and Cosmos DB.

OpenAI has made a new large Azure commitment; OpenAI’s APIs run exclusively on Azure; management is still very happy with the OpenAI partnership; Microsoft has ROFR (right of first refusal) on OpenAI’s Stargate project

As we shared last week, we are thrilled OpenAI has made a new large Azure commitment…

… And with OpenAI’s APIs exclusively running on Azure, customers can count on us to get access to the world’s leading models…

…[Question] I wanted to ask you about the Stargate news and the announced changes in the OpenAI relationship last week. It seems that most of your investors have interpreted this as Microsoft, for sure, remaining very committed to OpenAI’s success, but electing to take more of a backseat in terms of funding OpenAI’s future training CapEx needs. I was hoping you might frame your strategic decision here around Stargate.

[Answer] We remain very happy with the partnership with OpenAI. And as you saw, they have committed in a big way to Azure. And even in the bookings, what we recognized is just the first tranche of it. And so you’ll see, given the ROFR we have, more benefits of that even into the future. 

Microsoft’s management thinks Azure AI Foundry has best-in-class tooling run times for users to build AI agents and access thousands of AI models; Azure AI Foundry already has 200,000 monthly active users after just 2 months; the models available on Azure AI Foundry include DeepSeek’s R1 model, and more than 30 industry-specific models from partners; Microsoft’s Phi family of SLMs (small language model) has over 20 million downloads

Azure AI Foundry features best-in-class tooling run times to build agents, multi-agent apps, AIOps, API access to thousands of models. Two months in, we already have more than 200,000 monthly active users, and we are well positioned with our support of both OpenAI’s leading models and the best selection of open source models and SLMs. DeepSeek’s R1 launched today via the model catalog on Foundry and GitHub with automated red teaming, content safety integration and security scanning. Our Phi family of SLM has now been downloaded over 20 million times. And we also have more than 30 models from partners like Bayer, PAYG AI, Rockwell Automation, Siemens to address industry-specific use cases.

Microsoft’s management thinks Microsoft 365 Copilot is the UI (user interface) for AI; management is seeing accelerated adoption of Microsoft 365 Copilot across all deal sizes; majority of Microsoft 365 Copilot customers purchase more seats over time; daily users of Copilot more than doubled sequentially in 2024 Q4, while usage intensity grew 60% sequentially; more than 160,000 organisations have used Copilot Studio, creating more than 400,000 custom agents in 2024 Q4, uo 2x sequentially; Microsoft’s data cloud drives Copilot as the UI for AI; management is seeing Copilot plus AI agents disrupting business applications; the initial seats for Copilot were for departments that could see immediate productivity benefits, but the use of Copilot then spreads across the enterprise

Microsoft 365 Copilot is the UI for AI. It helps supercharge employee productivity and provides access to a swarm of intelligent agents to streamline employee workflow. We are seeing accelerated customer adoption across all deal sizes as we win new Microsoft 365 Copilot customers and see the majority of existing enterprise customers come back to purchase more seats. When you look at customers who purchased Copilot during the first quarter of availability, they have expanded their seat collectively by more than 10x over the past 18 months. To share just one example, Novartis has added thousands of seats each quarter over the past year and now have 40,000 seats. Barclays, Carrier Group, Pearson and University of Miami all purchased 10,000 or more seats this quarter. And overall, the number of people who use Copilot daily, again, more than doubled quarter-over-quarter. Employees are also engaging with Copilot more than ever. Usage intensity increased more than 60% quarter-over-quarter and we are expanding our TAM with Copilot Chat, which was announced earlier this month. Copilot Chat, along with Copilot Studio, is now available to every employee to start using agents right in the flow of work…

…More than 160,000 organizations have already used for Copilot Studio, and they collectively created more than 400,000 custom agents in the last 3 months alone, up over 2x quarter-over-quarter…

…What is driving Copilot as the UI for AI as well as our momentum with agents is our rich data cloud, which is the world’s largest source of organizational knowledge. Billions of e-mails, documents and chats, hundreds of millions of Teams meetings and millions of SharePoint sites are added each day. This is the enterprise knowledge cloud. It is growing fast, up over 25% year-over-year…

…What we are seeing is Copilot plus agents disrupting business applications, and we are leaning into this. With Dynamics 365, we took share as organizations like Ecolab, Lenovo, RTX, TotalEnergies and Wyzant switched to our AI-powered apps from legacy providers…

…[Question] Great to hear about the strength you’re seeing in Copilot… Would love to get some color on just the common use cases that you’re seeing that give you that confidence that, that will ramp into monetization later.

[Answer] I think the initial sort of set of seats were for places where there’s more belief in immediate productivity, a sales team, in finance or in supply chain where there is a lot of, like, for example, SharePoint grounded data that you want to be able to use in conjunction with web data and have it produce results that are beneficial. But then what’s happening very much like what we have seen in these previous generation productivity things is that people collaborate across functions, across roles, right? For example, even in my own daily habit, it’s I go to chat, I use Work tab and get results, and then I immediately share using Pages with colleagues. I sort of call it think with AI and work with people. And that pattern then requires you to make it more of a standard issue across the enterprise. And so that’s what we’re seeing.

Azure grew revenue by 31% in 2024 Q4 (was 33% in 2024 Q3), with 13 points of growth from AI services (was 12 points in 2024 Q3); Azure AI services was up 157% year-on-year, with demand continuing to be higher than capacity;  Azure’s non-AI business had weaker-than-expected growth because of go-to-market execution challenges

Azure and other cloud services revenue grew 31%. Azure growth included 13 points from AI services, which grew 157% year-over-year, and was ahead of expectations even as demand continued to be higher than our available capacity. Growth in our non-AI services was slightly lower than expected due to go-to-market execution challenges, particularly with our customers that we primarily reach through our scale motions as we balance driving near-term non-AI consumption with AI growth.

For Azure’s expected growth of 31%-32% in 2025 Q1 (FY2025 Q3), management expects  contribution from AI services to grow from increased AI capacity coming online; management expects Azure’s non-AI services to still post healthy growth, but there are still impacts from execution challenges; management expects Azure to no longer be capacity-constrained by the end of FY2025 (2025 Q2); Azure’s capacity constraint has been in power and space

In Azure, we expect Q3 revenue growth to be between 31% and 32% in constant currency driven by strong demand for our portfolio of services. As we shared in October, the contribution from our AI services will grow from increased AI capacity coming online. In non-AI services, healthy growth continues, although we expect ongoing impact through H2 as we work to address the execution challenges noted earlier. And while we expect to be AI capacity constrained in Q3, by the end of FY ’25, we should be roughly in line with near-term demand given our significant capital investments…

…When I talk about being capacity constrained, it takes two things. You have to have space, which I generally call long-lived assets, right? That’s the infrastructure and the land and then you have to have kits. We’re continuing, and you’ve seen that’s why our spend has pivoted this way, to be in the long-lived investment. We have been short power and space. And so as you see those investments land that we’ve made over the past 3 years, we get closer to that balance by the end of this year.

More than half of Microsoft’s cloud and AI-related capex in 2024 Q4 (FY2025 Q2) are for long-lived assets that will support monetisation over the next 15 years and more, while the other half are for CPUs and GPUs; management expects Microsoft’s capex in 2025 Q1 (FY2025 Q3) and 2025 Q2 (FY2025 Q4) to be at similar levels as 2024 Q4 (FY2025 Q2); FY2026’s capex will grow at a lower rate than in FY2025; the mix of spend in FY2026 will shift to short-lived assets in FY2026; Microsoft’s long-lived infrastructure investments are fungible; the long-lived assets are land; the presence of Moore’s Law means that management does not want to invest too much capex in any one year because the hardware and software will become much better in just 1 year; management thinks Microsoft’s AI infrastructure should be continuously upgraded to take advantage of Moore’s Law; Microsoft’s AI capex growth going forward will be tagged to customer contract delivery; the fungibility of Microsoft’s AI infrastructure investments relates to not just inference (which is the primary use case), but also training, post training, and running the commercial cloud business

More than half of our cloud and AI-related spend was on long-lived assets that will support monetization over the next 15 years and beyond. The remaining cloud and AI spend was primarily for servers, both CPUs and GPUs, to serve customers based on demand signals, including our customer contracted backlog…

…Next, capital expenditures. We expect quarterly spend in Q3 and Q4 to remain at similar levels as our Q2 spend. In FY ’26, we expect to continue investing against strong demand signals, including customer contracted backlog we need to deliver against across the entirety of our Microsoft Cloud. However, the growth rate will be lower than FY ’25 and the mix of spend will begin to shift back to short-lived assets, which are more correlated to revenue growth. As a reminder, our long-lived infrastructure investments are fungible, enabling us to remain agile as we meet customer demand globally across our Microsoft Cloud, including AI workloads…

…When I talk about being capacity constrained, it takes two things. You have to have space, which I generally call long-lived assets, right? That’s the infrastructure and the land and then you have to have kits. We’re continuing, and you’ve seen that’s why our spend has pivoted this way, to be in the long-lived investment. We have been short power and space. And so as you see those investments land that we’ve made over the past 3 years, we get closer to that balance by the end of this year…

…You don’t want to buy too much of anything at one time because, in Moore’s Law, every year is going to give you 2x, your optimization is going to give you 10x. You want to continuously upgrade the fleet, modernize the fleet, age the fleet and, at the end of the day, have the right ratio of monetization and demand-driven monetization to what you think of as the training expense…

…I do think the way I want everyone to internalize it is that the CapEx growth is going through that cycle pivot, which is far more correlated to customer contract delivery, no matter who the end customer is…

…  the other thing that’s sometimes missing is when we say fungible, we mean not just the primary use, which we’ve always talked about, which is inference. But there is some training, post training, which is a key component. And then they’re just running the commercial cloud, which at every layer and every modern AI app that’s going to be built will be required. It will be required to be distributed, and it will be required to be global. And all of those things are really important because it then means you’re the most efficient. And so the investment you see us make in CapEx, you’re right, the front end has been this sort of infrastructure build that lets us really catch up not just on the AI infrastructure we needed, but think about that as the building itself, data centers, but also some of the catch-up we need to do on the commercial cloud side. And then you’ll see the pivot to more CPU and GPU. 

Microsoft’s management thinks DeepSeek had real innovations, but those are going to be commoditized and become broadly used; management thinks that innovations in AI that reduce the cost of inference will drive more consumption and more apps being developed, and make AI more ubiquitous, which are all positive forces for Microsoft

I think DeepSeek has had some real innovations. And that is some of the things that even OpenAI found in ’01. And so we are going to — obviously, now that all gets commoditized and it’s going to get broadly used. And the big beneficiaries of any software cycle like that is the customers, right? Because at the end of the day, if you think about it, right, what was the big lesson learned from client server to cloud? More people bought servers, except it was called cloud. And so when token prices fall, inference computing prices fall, that means people can consume more, and there will be more apps written. And it’s interesting to see that when I referenced these models that are pretty powerful, it’s unimaginable to think that here we are in sort of beginning of ’25, where on the PC, you can run a model that required pretty massive cloud infrastructure. So that type of optimization means AI will be much more ubiquitous. And so therefore, for a hyperscaler like us, a PC platform provider like us, this is all good news as far as I’m concerned.

Microsoft has been reducing prices of GPT models over the years through inference optimizations

We are working super hard on all the software optimizations, right? I mean, just not the software optimizations that come because of what DeepSeek has done, but all the work we have done to, for example, reduce the prices of GPT models over the years in partnership with OpenAI. In fact, we did a lot of the work on the inference optimizations on it, and that’s been key to driving, right?

Microsoft’s management is aware that launching a frontier AI model that is too expensive to serve is useless

One of the key things to note in AI is you just don’t launch the frontier model, but if it’s too expensive to serve, it’s no good, right? It won’t generate any demand.

Microsoft’s management is seeing many different AI models being used for any one application; management thinks that there will always be a combination of different models used in any one application

What you’re seeing is effectively lots of models that get used in any application, right? When you look underneath even a Copilot or a GitHub Copilot or what have you, you already see lots of many different models. You build models. You fine-tune models. You distill models. Some of them are models that you distill into an open source model. So there’s going to be a combination…

….There’s a temporality to it, right? What you start with as a given COGS profile doesn’t need to be the end because you continuously optimize for latency and COGS and putting in different models.

NVIDIA (NASDAQ: NVDA)

NVIDIA’s Data Center revenue again had incredibly strong growth in 2024 Q4, driven by demand for the Hopper GPU computing platform and the ramping of the Blackwell GPU platform 

In the fourth quarter, Data Center revenue of $35.6 billion was a record, up 16% sequentially and 93% year-on-year, as the Blackwell ramp commenced and Hopper 200 continued sequential growth. 

Blackwell’s sales exceeded management’s expectations and is the fastest product ramp in NVIDIA’s history; it is common for Blackwell clusters to start with 100,000 GPUs or more and NVIDIA has started shipping for multiple such clusters; management architected Blackwell for inference; Blackwell has 25x higher token throughput and 20x lower cost for AI reasoning models compared to the Hopper 100; Blackwell has a NVLink domain that handles the growing complexity of inference at scale; management is seeing great demand for Blackwell for inference, with many of the early GB200 (GB200 is based on the Blackwell family of GPUs) deployments earmarked for inference; management expects NVIDIA’s gross margin to decline slightly initially as the Blackwell family ramps, before rebounding; management expects a significant ramp of Blackwell in 2025 Q1; the Blackwell Ultra, the next generation of GPUs within the Blackwell family, is slated for introduction in 2025 H2; the system architecture between Blackwell and Blackwell Ultra is exactly the same

In Q4, Blackwell sales exceeded our expectations. We delivered $11 billion of Blackwell revenue to meet strong demand. This is the fastest product ramp in our company’s history, unprecedented in its speed and scale…

…With Blackwell, it will be common for these clusters to start with 100,000 GPUs or more. Shipments have already started for multiple infrastructures of this size…

…Blackwell was architected for reasoning AI inference. Blackwell supercharges reasoning AI models with up to 25x higher token throughput and 20x lower cost versus Hopper 100. Its revolutionary transformer engine is built for LLM and mixture of experts inference. And its NVLink domain delivers 14x the throughput of PCIe Gen 5, ensuring the response time, throughput and cost efficiency needed to tackle the growing complexity of inference at scale…

…Blackwell has great demand for inference. Many of the early GB200 deployments are earmarked for inference, a first for a new architecture…

…As Blackwell ramps, we expect gross margins to be in the low 70s. Initially, we are focused on expediting the manufacturing of Blackwell systems to meet strong customer demand as they race to build out Blackwell infrastructure. When fully ramped, we have many opportunities to improve the cost and gross margin will improve and return to the mid-70s, late this fiscal year…

…Continuing with its strong demand, we expect a significant ramp of Blackwell in Q1…

…Blackwell Ultra is second half…

…The next train is on an annual rhythm and Blackwell Ultra with new networking, new memories and of course, new processors, and all of that is coming online…

…This time between Blackwell and Blackwell Ultra, the system architecture is exactly the same. It’s a lot harder going from Hopper to Blackwell because we went from an NVLink 8 system to a NVLink 72-based system. So the chassis, the architecture of the system, the hardware, the power delivery, all of that had to change. This was quite a challenging transition. But the next transition will slot right in. Blackwell Ultra will slot right in.

NVIDIA’s management sees post-training and model customisation has demanding orders of magnitude more compute than pre-training

The scale of post-training and model customization is massive and can collectively demand orders of magnitude, more compute than pretraining.

NVIDIA’s management is seeing accelerating demand for NVIDIA GPUs for inference, driven by test-time scaling and new reasoning models; management thinks reasoning models require 100x more compute per task than one-shot inference models; management is hopeful that future generation of reasoning models will require millions of times more compute; management is seeing that the vast majority of NVIDIA’s compute today is inference

Our inference demand is accelerating, driven by test-time scaling and new reasoning models like OpenAI o3, DeepSeek-R1 and Grok 3. Long thinking reasoning AI can require 100x more compute per task compared to one-shot inferences…

…. The amount of tokens generated, the amount of inference compute needed is already 100x more than the one-shot examples and the one-shot capabilities of large language models in the beginning. And that’s just the beginning. This is just the beginning. The idea that the next generation could have thousands times and even hopefully, extremely thoughtful and simulation-based and search-based models that could be hundreds of thousands, millions of times more compute than today is in our future…

……The vast majority of our compute today is actually inference and Blackwell takes all of that to a new level.

Companies such as ServiceNow, Perplexity, Microsoft, and Meta are using NVIDIA’s software and GPUs to achieve lower costs and/or better performance with their inference workloads

ServiceNow tripled inference throughput and cut costs by 66% using NVIDIA TensorRT for its screenshot feature. Perplexity sees 435 million monthly queries and reduced its inference costs 3x with NVIDIA Triton Inference Server and TensorRT-LLM. Microsoft Bing achieved a 5x speed up at major TCO savings for Visual Search across billions of images with NVIDIA TensorRT and acceleration libraries…

…Meta’s cutting-edge Andromeda advertising engine runs on NVIDIA’s Grace Hopper Superchip serving vast quantities of ads across Instagram, Facebook applications. Andromeda harnesses Grace Hopper’s fast interconnect and large memory to boost inference throughput by 3x, enhanced ad personalization and deliver meaningful jumps in monetization and ROI.

NVIDIA has driven a 200x reduction in inference costs in the last 2 years

We’re driven to a 200x reduction in inference costs in just the last 2 years.

Large cloud service providers (CSPs) were half of NVIDIA’s Data Centre revenue in 2024 Q4, and up nearly 2x year-on-year; large CSPs were the first to stand up Blackwell systems

In Q4, large CSPs represented about half of our data center revenue, and these sales increased nearly 2x year-on-year. Large CSPs were some of the first to stand up Blackwell with Azure, GCP, AWS and OCI bringing GB200 systems to cloud regions around the world to meet surging customer demand for AI. 

Regional clouds increased as a percentage of NVIDIA’s Data Center revenue in 2024 Q4, driven by AI data center build outs globally; management is seeing countries across the world building AI ecosystems

Regional cloud hosting NVIDIA GPUs increased as a percentage of data center revenue, reflecting continued AI factory build-outs globally and rapidly rising demand for AI reasoning models and agents where we’ve launched a 100,000 GB200 cluster-based incidents with NVLink Switch and Quantum-2 InfiniBand…

…Countries across the globe are building their AI ecosystems and demand for compute infrastructure is surging. France’s EUR 200 billion AI investment and the EU’s EUR 200 billion InvestAI initiatives offer a glimpse into the build-out to set redefined global AI infrastructure in the coming years.

NVIDIA’s revenue from consumer internet companies tripled year-on-year in 2024 Q4

Consumer Internet revenue grew 3x year-on-year, driven by an expanding set of generative AI and deep learning use cases. These include recommender systems, vision-language understanding, synthetic data generation, search and agentic AI.

NVIDIA’s revenue from enterprises nearly doubled year-on-year in 2024 Q4, partly with the help of agentic AI demand

Enterprise revenue increased nearly 2x year on accelerating demand for model fine-tuning, RAG and agentic AI workflows and GPU accelerated data processing.

NVIDIA’s management has introduced NIMs (NVIDIA Inference Microservices) focused on AI agents and leading AI agent platform providers are using these tools

We introduced NVIDIA Llama Nemotron model family NIMs to help developers create and deploy AI agents across a range of applications, including customer support, fraud detection and product supply chain and inventory management. Leading AI agent platform providers, including SAP and ServiceNow are among the first to use new models.

Healthcare companies are using NVIDIA’s AI products to power healthcare innovation

Health care leaders, IQVIA, Illumina and Mayo Clinic as well as ARC Institute are using NVIDIA AI to speed drug discovery, enhance genomic research and pioneer advanced health care services with generative and agentic AI.

Hyundai will be using NVIDIA’s technologies for the development of AVs (autonomous vehicles); NVIDIA’s automotive revenue had strong growth year-on-year and sequentially in 2024 Q4, driven by ramp in AVs; automotive companies such as Toyota, Aurora, and Continental are working with NVIDIA to deploy AV technologies; NVIDIA’s AV platform has passed 2 of the automotive industry’s foremost authorities for safety and cybersecurity

 At CES, Hyundai Motor Group announced it is adopting NVIDIA technologies to accelerate AV and robotics development and smart factory initiatives…

…Now moving to Automotive. Revenue was a record $570 million, up 27% sequentially and up 103% year-on-year…

…Strong growth was driven by the continued ramp in autonomous vehicles, including cars and robotaxis. At CES, we announced Toyota, the world’s largest auto maker will build its next-generation vehicles on NVIDIA Orin running the safety certified NVIDIA DriveOS. We announced Aurora and Continental will deploy driverless trucks at scale powered by NVIDIA DRIVE Thor. Finally, our end-to-end autonomous vehicle platform NVIDIA DRIVE Hyperion has passed industry safety assessments like TÜV SÜD and TÜV Rheinland, 2 of the industry’s foremost authorities for automotive-grade safety and cybersecurity. NVIDIA is the first AV platform to receive a comprehensive set of third-party assessments.

NVIDIA’s management has introduced the NVIDIA Cosmos World Foundation Model platform for the continued development of autonomous robots; Uber is one of the first major technology companies to adopt the NVIDIA Cosmos World Foundation Model platform

At CES, we announced the NVIDIA Cosmos World Foundation Model Platform. Just as language, foundation models have revolutionized language AI, Cosmos is a physical AI to revolutionize robotics. Leading robotics and automotive companies, including ridesharing giant Uber, are among the first to adopt the platform.

As a percentage of total Data Center revenue, NVIDIA’s Data Center revenue in China is well below levels seen prior to the US government’s export controls; management expects the Chinese market to be very competitive

Now as a percentage of total data center revenue, data center sales in China remained well below levels seen on the onset of export controls. Absent any change in regulations, we believe that China shipments will remain roughly at the current percentage. The market in China for data center solutions remains very competitive.

NVIDIA’s networking revenue declined sequentially in 2024 Q4, but the networking-attach-rate to GPUs remains robust at 75%; NVIDIA is transitioning to NVLink 72 with Spectrum-X (Spectrum-X is NVIDIA’s Ethernet networking solution); management expects networking revenue to resume growing in 2025 Q1; management sees AI requiring a new class of networking, for which the company’s NVLink, Quantum Infiniband, and Spectrum-X networking solutions are able to provide; large AI data centers, including OpenAI’s Stargate project, will be using Spectrum X

Networking revenue declined 3% sequentially. Our networking attached to GPU compute systems is robust at over 75%. We are transitioning from small NVLink 8 with InfiniBand to large NVLink 72 with Spectrum-X. Spectrum-X and NVLink Switch revenue increased and represents a major new growth vector. We expect networking to return to growth in Q1. AI requires a new class of networking. NVIDIA offers NVLink Switch systems for scale-up compute. For scale out, we offer Quantum InfiniBand for HPC supercomputers and Spectrum-X for Ethernet environments. Spectrum-X enhances the Ethernet for AI computing and has been a huge success. Microsoft Azure, OCI, CoreWeave and others are building large AI factories with Spectrum-X. The first Stargate data centers will use Spectrum-X. Yesterday, Cisco announced integrating Spectrum-X into their networking portfolio to help enterprises build AI infrastructure. With its large enterprise footprint and global reach, Cisco will bring NVIDIA Ethernet to every industry.

NVIDIA’s management is seeing 3 scaling laws at play in the development of AI models, namely pre-training scaling, post-training scaling, and test-time compute scaling

There are now multiple scaling laws. There’s the pre-training scaling law, and that’s going to continue to scale because we have multimodality, we have data that came from reasoning that are now used to do pretraining. And then the second is post-training scaling law, using reinforcement learning human feedback, reinforcement learning AI feedback, reinforcement learning, verifiable rewards. The amount of computation you use for post training is actually higher than pretraining. And it’s kind of sensible in the sense that you could, while you’re using reinforcement learning, generate an enormous amount of synthetic data or synthetically generated tokens. AI models are basically generating tokens to train AI models. And that’s post-training. And the third part, this is the part that you mentioned is test-time compute or reasoning, long thinking, inference scaling. They’re all basically the same ideas. And there you have a chain of thought, you’ve search.

NVIDIA’s management thinks the popularity of NVIDIA’s GPUs stems from its fungibility across all kinds of AI model architectures and use cases; NVIDIA’s management thinks that NVIDIA GPUs have an advantage over the ASIC (application-specific integrated circuit) AI chips developed by others because of (1) the general-purpose nature of NVIDIA GPUs, (2) NVIDIA’s rapid product development roadmap, (3) the software stack developed for NVIDIA GPUs that is incredibly hard to replicate

The question is how do you design such an architecture? Some of it — some of the models are auto regressive. Some of the models are diffusion-based. Some of it — some of the times you want your data center to have disaggregated inference. Sometimes it is compacted. And so it’s hard to figure out what is the best configuration of a data center, which is the reason why NVIDIA’s architecture is so popular. We run every model. We are great at training…

…When you have a data center that allows you to configure and use your data center based on are you doing more pretraining now, post training now or scaling out your inference, our architecture is fungible and easy to use in all of those different ways. And so we’re seeing, in fact, much, much more concentration of a unified architecture than ever before…

…[Question] We heard a lot about custom ASICs. Can you kind of speak to the balance between custom ASIC and merchant GPU?

[Answer] We build very different things than ASICs, in some ways, completely different in some areas we intercept. We’re different in several ways. One, NVIDIA’S architecture is general whether you’re — you’ve optimized for auto regressive models or diffusion-based models or vision-based models or multimodal models or text models. We’re great in all of it. We’re great at all of it because our software stack is so — our architecture is flexible, our software stack ecosystem is so rich that we’re the initial target of most exciting innovations and algorithms. And so by definition, we’re much, much more general than narrow…

…The third thing I would say is that our performance and our rhythm is so incredibly fast. Remember that these data centers are always fixed in size. They’re fixed in size or they’re fixed in power. And if our performance per watt is anywhere from 2x to 4x to 8x, which is not unusual, it translates directly to revenues. And so if you have a 100-megawatt data center, if the performance or the throughput in that 100-megawatt or the gigawatt data center is 4x or 8x higher, your revenues for that gigawatt data center is 8x higher. And the reason that is so different than data centers of the past is because AI factories are directly monetizable through its tokens generated. And so the token throughput of our architecture being so incredibly fast is just incredibly valuable to all of the companies that are building these things for revenue generation reasons and capturing the fast ROI…

…The last thing that I would say is the software stack is incredibly hard. Building an ASIC is no different than what we do. We build a new architecture. And the ecosystem that sits on top of our architecture is 10x more complex today than it was 2 years ago. And that’s fairly obvious because the amount of software that the world is building on top of architecture is growing exponentially and AI is advancing very quickly. So bringing that whole ecosystem on top of multiple chips is hard.

NVIDIA’s management thinks that only consumer AI and search currently have well-developed AI use cases, and the next wave will be agentic AI, robotics, and sovereign AI

We’ve really only tapped consumer AI and search and some amount of consumer generative AI, advertising, recommenders, kind of the early days of software. The next wave is coming, agentic AI for enterprise, physical AI for robotics and sovereign AI as different regions build out their AI for their own ecosystems. And so each one of these are barely off the ground, and we can see them.

NVIDIA’s management sees the upcoming Rubin family of GPUs as being a big step-up from the Blackwell family

The next transition will slot right in. Blackwell Ultra will slot right in. We’ve also already revealed and been working very closely with all of our partners on the click after that. And the click after that is called Vera Rubin and all of our partners are getting up to speed on the transition of that and so preparing for that transition. And again, we’re going to provide a big, huge step-up.

NVIDIA’s management sees AI as having the opportunity to address a larger part of the world’s GDP than any other technology has ever had

No technology has ever had the opportunity to address a larger part of the world’s GDP than AI. No software tool ever has. And so this is now a software tool that can address a much larger part of the world’s GDP more than any time in history.

NVIDIA’s management sees customers still actively using older families of NVIDIA GPUs because of the high level of programmability that CUDA has

People are still using Voltas and Pascals and Amperes. And the reason for that is because there are always things that — because CUDA is so programmable you could use it — one of the major use cases right now is data processing and data curation. You find a circumstance that an AI model is not very good at. You present that circumstance to a vision language model, let’s say, it’s a car. You present that circumstance to a vision language model. The vision language model actually looks at the circumstances and said, “This is what happened and I wasn’t very good at it.” You then take that response — the prompt and you go and prompt an AI model to go find in your whole lake of data, other circumstances like that, whatever that circumstance was. And then you use an AI to do domain randomization and generate a whole bunch of other examples. And then from that, you can go train the model. And so you could use the Amperes to go and do data processing and data curation and machine learning-based search. And then you create the training data set, which you then present to your Hopper systems for training. And so each one of these architectures are completely — they’re all CUDA-compatible and so everything runs on everything. But if you have infrastructure in place, then you can put the less intensive workloads onto the installed base of the past. All of our GPUs are very well employed.

Paycom Software (NYSE: PAYC)

Paycom’s management rolled out an AI agent six months ago to its service team, and has seen higher immediate response rates to clients and eliminated service tickets by 25% from a year ago; Paycom’s AI agent is driving internal efficiencies, higher client satisfaction, and higher Net Promoter Scores

Paycom’s AI agent, which was rolled out to our service team 6 months ago, utilizes our own knowledge-based semantic search model to provide faster responses and help our clients more quickly and consistently than ever before. As responses continuously improve over time, our client interactions become more valuable, and we connect them faster to the right solution. As a result, we are seeing improved immediate response rates and have eliminated service tickets by over 25% compared to a year ago…

…With automations like AI agent, we are realizing internal efficiencies, driving increasing client satisfaction and seeing higher Net Promoter Scores.

PayPal (NASDAQ: PYPL)

One of the focus areas for PayPal’s management in 2025 will be on raising efficiency with the help of AI 

Fourth is efficiency and effectiveness. In 2024, we reduced headcount by 10%. We made deliberate investments in AI and automation, which are critical to our future. This year, we are prioritizing the use of AI to improve the customer experience and drive efficiency and effectiveness within PayPal.

PayPal’s management sees AI being a huge opportunity for PayPal given the company’s volume of data; PayPal is using AI on its customer facing side to more efficiently process customer support cases and interactions with customers (PayPal Assistant has been rolled out and it has cut down phone calls and active events for PayPal); PayPal is using AI to personalise the commerce journey for consumers; PayPal is also using AI for back-office productivity and risk decisions

[Question] The ability to use AI for more operating efficiency. And are those initiatives that are requiring some incremental investment near term? Or are you already seeing sort of a positive ROI from that?

[Answer] AI is opening up a huge opportunity for us. First, at our scale, we saw 26 billion transactions on our platform last year. We have a massive data set that we are actively working and investing in to be able to drive our effectiveness and efficiency…

…First, on the customer-facing side, we’re leveraging AI to really become more efficient in our support cases and how we interact with our customers. We see tens of millions of support cases every year, and we’ve rolled out our PayPal Assistant, which is now really cutting down phone calls and active events that we have. 

We also are leveraging AI to personalize the commerce journey, and so working with our merchants to be able to understand and create this really magical experience for consumers. When they show up at a checkout, it’s not just a static button anymore. This really can become a dynamic, personalized button that starts to understand the profile of the consumer, the journey that they’ve been on, perhaps across merchants, and be able to enable a reward or a cash-back offer in the moment or even a Buy Now, Pay Later offer in a dynamic experience…

…In addition, we also are looking at our back office and ensuring that not just on the engineering and employee productivity side, but also in things like our risk decisions. We see billions and billions of risk decisions that often, to be honest, we’re very manual in the past. We’re now leveraging AI to be able to understand globally what are the nature of these risk decisions and how do we automate these across both risk models as well as even just ensuring that customers get the right response at the right time in an automated fashion.

Salesforce (NYSE: CRM)

Salesforce ended 2024 (FY2025) with $900 million in Data Cloud and AI ARR (annual recurring revenue), up 120% from a year ago; management has never seen products grow at this rate before, especially Agentforce

We ended this year with $900 million in Data Cloud and AI ARR. It grew 120% year-over-year. We’ve never seen products grow at these levels, especially Agentforce.

Salesforce’s management thinks building digital labour (AI agents) is a much bigger market than just building software

I’m sure you saw those ARK slides that got released over the weekend where she said that she thought this digital labor revolution, which is really like kind of what we’re in here now, this digital labor revolution, this looks like it’s anywhere from a few trillion to $12 trillion. I mean, I kind of agree with her. I think this is much, much bigger than software. I mean, for the last 25 years, we’ve been doing software to help our customers manage their data. That’s very exciting. I think building software that kind of prints and deploys digital workers is more exciting.

Salesforce’s unified platform, under one piece of code, combining customer data and an agentic platform, is what gives Agentforce its accuracy; Agentforce already has 3,000 paying customers just 90 days after going live; management thinks Agentforce is unique in the agentic capabilities it is delivering; Salesforce is Customer Zero for Agentforce; Agentforce has already resolved 380,000 service requests for Salesforce, with an 84% resolution rate, and just 2% of requests require human escalation; Agentforce has accelerated Salesforce’s sales-quoting cycles by 75% and increased AE (account executive) capacity by 7%; Agentforce is helping Salesforce engage more than 50 leads per day, freeing up the sales team for higher-value conversations; management wants every Salesforce customer to be using Agentforce; Data Cloud is at the heart of Agentforce; management is seeing customers across every industry deploying Agentforce; management thinks Salesforce’s agentic technology works better than many other providers, and that other providers are just whitewashing their technology with the “agent” label; Agentforce is driving growth across Salesforce’s portfolio; Salesfroce has prebuilt 170 specialised Agentforce industry skills; Agentforce’s 3,000 customers come from a diverse set of industries

Our formula now really for our customers is this idea that we have these incredible Customer 360 apps. We have this incredible Data Cloud, and this incredible agentic platform. These are the 3 layers. But that it is a deeply unified platform, it’s a deeply unified platform, it’s just one piece of code, that’s what makes it so unique in this market…

…It’s this idea that it’s a deeply unified platform with one piece of code all wrapped in a beautiful layer of trust. And that’s what gives Agentforce this incredible accuracy that we’re seeing…

…Just 90 days after it went live, we’ve already have 3,000 paying Agentforce customers who are experiencing unprecedented levels of productivity, efficiency and cost and cost savings. No one else is delivering at this level of capability…

…We’re seeing some amazing results on Salesforce as Customer Zero for Agentforce. Our digital labor force is resolving tens of thousands of customer service inquiries, freeing our human employees to focus on the most nuanced issues and customer relationships. We’re seeing tremendous momentum and success stories emerge as we execute our vision to make every company, every single company, every customer of ours, an Agentforce company, that is, we want every customer to be an Agentforce customer…

…We also continued phenomenal growth with Data Cloud this year, which is the heart of Agentforce. Data Cloud is the fuel that powers Agentforce and our customers are investing in it…

…We’re seeing customers deploy Agentforce across every industry…

…You got to be aware of the false agent because the false agent is out there where people can use the word agent or they kind of — they’re trying to whitewash all the agent, the thing, everywhere. But the reality is there is the real agents and there are the false agents, and we’re very fortunate to have the real stuff going on here. So we’ve got a lot more groundbreaking AI innovation coming…

…Today, we’re live on Agentforce across service and sales, our business technology organization, customer support and more. And the results are phenomenal. Since launching on our Salesforce help portal in October, Agentforce has autonomously handled 380,000 service requests, achieving an incredible 84% resolution rate and only 2% of the requests require human escalation. And we’re using Agentforce for quoting, accelerating our quoting cycles by more than 75%. In Q4, we increased our AE [account executive] capacity while still driving productivity up 7% year-over-year. Agentforce is transforming how we do outbound prospecting, already engaging more than 50 leads per day with personalized outreach and timely follow-ups, freeing up our teams to focus on high-value conversation. Our reps are participating in thousands of sales coaching training sessions each month…

…Agentforce is revolutionizing how our customers work by bringing AI-powered insights and actions directly into the workflows across the Customer 360 applications. This is driving strong growth across our portfolio. Sales Cloud and Service Cloud both achieved double-digit growth again in Q4. We’re seeing fantastic momentum with Slack, with customers like ZoomInfo, Remarkable and MIMIT Health using Agentforce and Slack to boost productivity…

…We’ve prebuilt over 170 specialized Agentforce industry skills and a team of 400 specialists, supporting transformations across sectors and geographies…

…We closed more than 3,000 paid Agentforce deals in the quarter. As customers continue to harness the value of AI deeply embedded across our unified platform, it is no surprise that these customers average nearly 4 clouds. And these customers came from a diverse set of industries with more than half in technology, manufacturing, financial services and HLS.

Lennar, the USA’s largest homebuilder, has been a Salesforce customer for 8 years and it is deploying Agentforce to fulfill their management’s vision of selling all kinds of new products; jewelry company, Pandora, an existing Salesforce customer, is deploying Agentforce with the aim of handling 30%-60% of its service cases with Agentforce; pharmaceutical giant Pfizer is using Agentforce to augment its sales teams; Singapore-based airline, Singapore Airlines, is now a customer of Agentforce and wants to deliver service through it; Goodyear is using Agentforce to automate and increase the effectiveness of its sales efforts; Accenture is using Agentforce to coach its sales team and expects to achieve higher win rates; Deloitte is using Agentforce and expects to achieve significant productivity gains

We’ve been working with Lennar, the nation’s largest homebuilder. And most of you know Lennar is really an incredible company, and they’ve been a customer of ours for about 8 years…

…You probably know Stuart Miller, Jon Jaffe, amazing CEOs. And those co-CEOs called me and said, “Listen, these guys have done a hackathon around Agentforce. We’ve got 5 use cases. We see incredible opportunities on our margin, incredible opportunities in our revenue. And do you have our back if we’re going to deploy this?” And we said, “Absolutely. We’ve deployed it ourselves,” which is the best evidence that this is real. And they are just incredible, their vision as a homebuilder providing 24/7 support, sales leads through all their digital channels. They’re able to sell all kinds of new products. I think they’re going to sell mortgages and insurance and all kinds of things to their customers. And the cool thing is they’re using our sales product, our service product, marketing, MuleSoft, Slack, Tableau, they use everything. But they are able to leverage it all together by realizing that just by turning it on, they get this incredible Agentforce capability…

…I don’t know how many of you know about Pandora. If you’ve been to a shopping center, you will see the Pandora store. You walk in, they have this gorgeous jewelry. They have these cool charm bracelets. They have amazing products. And if you know their CEO, Alex, he’s absolutely phenomenal…

…They’re in 100 countries. They employ 37,000 people worldwide. And Alex has this great vision to augment their employees with digital labor. And this idea that whether you’re on their website or in their store, or whatever it is, that they’re going to be able to do so much more with Agentforce. They already use — first of all, they already use Commerce Cloud. So if you’ve been to pandora.com and bought their products — and if you have it, by the way, it’s completely worthwhile. It’s great. And you can experience our Commerce Cloud, but it’s deeply integrated with our Service Cloud, with Data Cloud. It’s the one unified platform approach. And now they’re just flipping the switch, turning agents on, and they’re planning to deliver 30% to 60% of their service cases with Agentforce. That is awesome. And I really love Alex’s vision of what’s possible….

…The last customer I really want to hit on, which I’m so excited about, is Pfizer. And Albert is an incredible CEO. They are doing unbelievable things. They’ve been a tremendous customer. But now they’re really going all in on our Life Sciences Cloud…

…And with Agentforce, sales agents, for example, with Pfizer, that’s — they’ve got 20,000 customer-facing employees and customer-facing folks. That is just a radical extension for them with agents…

…I’m sure a lot of you — like, I have flown in Singapore Air. You know what? It’s a great airline. The CEO, Goh, is amazing. And he has a huge vision that also came out of Dreamforce, where — they’ve already delivered probably the best service of any airline in the world — they want to deliver it through agents. So whether you’re doing it with service or sales or marketing or commerce or all the different things that Singapore Air is doing with us, you’re going to be able to do this right on Singapore Air…

…Goodyear is partnering with us on their transformation, using Agentforce to automate and increase the effectiveness of their sales efforts. With Agentforce for Field Service, Goodyear will be able to reduce repair time by assisting technicians with answers to vehicle-related questions and autonomously scheduling field tech appointments…

…Accenture is using Agentforce Sales Coach, which provides personalized coaching and recommendations for sales teams, which is expected to lead to higher win rates. And Deloitte is projecting significant productivity gains and saved workforce hours as they roll out Agentforce over the next few years.

Salesforce’s management expects modest revenue contribution from Agentforce in 2025 (FY2026); contribution from Agentforce is expected to be more meaningful in 2026 (FY2027)

Starting with full fiscal year ’26. We expect revenue of $40.5 billion to $40.9 billion, growth of approximately 7% to 8% year-over-year in nominal and constant currency. And for subscription and support revenue, we expect growth of approximately 9% year-over-year in constant currency…

…On Agentforce, we are incredibly excited about the customer momentum we are seeing. However, the adoption cycle is still early as we focus on deployment with our customers. As a result, we are assuming a modest contribution to revenue in fiscal ’26. We expect the momentum to build throughout the year, driving a more meaningful contribution in fiscal ’27.

Salesforce has long had a mix of per-seat and consumption pricing models; for now, Agentforce is a consumption product, but management sees Agentforce evolving to a mix of per-seat and consumption pricing models; there was a customer that bought Agentforce in 2024 Q4 (FY2025 Q4) along with other Salesforce products and the customer signed a $7 million Agentforce contract and a $13 million contract for the other products; based on early days of engagement with Agentforce customers, management sees significant future upside to Salesforce’s pricing structure; Agentforce’s pricing will also take into account whether Agentforce will bring other human-based clouds into the customer; Agentforce is currently creating some halo around Salesforce’s other products

We’ve kind of started the company out with the per user pricing model, and that’s about humans. We price per human, so you’re kind of pricing per human. And then we have products, though, that are also in the consumption world as well. And of course, those started in the early days, things like our sandboxes, even things like our Commerce Cloud, even our e-mail marketing product, our Marketing Cloud. These are consumption-based products we’ve had for years…

…Now we have these kind of products that are for agents also, and agents are also a consumption model. So when we look at our Data Cloud, for example, that’s a consumption product. Agentforce is a consumption product. But it’s going to be a mix. It’s going to be a mix between what’s going on with our customers with how many humans do they have and then how many agents are they deploying…

…In the quarter, we did a large transaction with a large telecommunications company… we’re rebuilding this telecommunications company. So it’s Sales Cloud, it’s Service Cloud, it’s Marketing Cloud. It’s all of our core clouds, but then also it’s Agentforce. And the Agentforce component, I think, was maybe $7 million in the transaction. So she was buying $7 million of Agentforce. She bought $13 million in our products for humans, and I think that was about $20 million in total…

…We will probably move into the near future from conversations as we price most of our initial deals to universal credit. It will allow our customers far more flexibility in the way they transact with us. But we see this as a significant upside to our pricing structures going forward. And that’s what we’ve seen in the early days with our engagement with customers…

…Here’s a transaction that you’re doing, let’s say, a customer comes in, they’re very interested in building an agentic layer on their company, is that bringing other human-based clouds along with it?…

…[Question] Is Agentforce having a bit of a halo effect around some of your other products, meaning, as we are on the journey to get more monetization from Agentforce, are you seeing pickups or at least higher activity levels in some of your other products?

[Answer] That’s exactly right. And we’re seeing it in the way that our customers are using our technology, new ideas, new workflows, new engagements. We talked about Lennar as an example, their ability to handle leads after hours that they weren’t able to get back to or respond to in a quick time frame are now able to touch and engage with those leads. And that, of course, flows into their Salesforce automation system. And so we are seeing this halo effect with our core technology. It is making every single one of our core apps better as they deliver intelligence, underpinning these applications.

Salesforce’s management sees the combination of apps, data, and agents as the winning combination in an AI-world; management disputes Microsoft’s narrative that software apps will become a dumb database layer in an AI-dominated world, because it is the combination of apps, data, and agents that is important

I don’t know any company that’s 100% agents. I don’t know of any company that doesn’t need automation for its humans. I don’t know of any company that doesn’t need a data cloud where it needs a consistent common data repository for all of its agents to gain their intelligence. And I don’t know of any company that’s not going to need an agentic layer. And that idea of having apps, data and agents, I think, is going to be the winning combination…

…[Question] As part of that shift to agentic technology, there’s been a lot of debate about the SaaS technology and the business model. The SaaS tech stack that you built and pioneered, how does that fit into the agentic world? Is there a risk that SaaS just becomes a CRUD database?

[Answer] I’ve heard that Microsoft narrative, too. So I watched the podcast you watched, and that’s a very interesting idea. Here’s how I look at it, which is, I believe there is kind of a holy trinity here of AI CRM, which is the apps, the data and the agents. And these three things have to kind of work together. And I kind of put my money where our mouth is where we kind of built it and we delivered it. And you can see the 380,000 conversations that we had as point of evidence here in the last 90 days on our service and with a very high resolution rate of 84%. You can go to help.salesforce.com, and you can see that today.

Now Microsoft has had Copilot available for, I think, about 2 years or more than 2 years. And I know that they’re the reseller of OpenAI and they’ve invested, they kind of repackaged this ChatGPT, whatever. But where on their side are they delivering agents? Where in their company have they done this? Are they a best practice? Because I think that while they can say such a thing, do they have humans and agents working together to create customer success? Are they rebalancing their workforce with humans and agents? I think that it’s a very interesting point that, yes, the agentic layer is very important, but it doesn’t operate by itself. It operates with data, with a Data Cloud that has to be federated through your company, to all your data sources. And humans, we’re still here

Salesforce’s management is seeing Agentforce deliver tremendous efficiency in Salesforce’s customer support function that they may rebalance some customer-support roles into other roles; management is currently seeing AI coding tools improve the productivity of Salesforce’s engineering team by 30% and thinks even more productivity can be found; management will not be expanding Salesforce’s engineering team this year, but will grow the sales team

We really are seeing tremendous efficiency with help.salesforce.com. So we may see the opportunity to rebalance some of those folks into sales and marketing and other functions…

…We definitely have seen a lot of efficiency with engineering and with some of the new tools that I’ve seen, especially some of these high-performance coding tools. One of the key members of my staff who’s here in the room with us has just showed me one of his new examples of what we’re able to do with these coding tools, pretty awesome. And we’re not going to hire any new engineers this year. We’re seeing 30% productivity increase on engineering. And we’re going to really continue to ride that up…

…We’re going to grow sales pretty dramatically this year. Brian has got a big vision for how to grow the sales organization. probably another 10% to 20%, I hope, this year because we’re seeing incredible levels of demand.

Salesforce’s management thinks that AI agents is one of the catalysts to drive GDP growth

So if you want productivity to go up and you want GDP to grow up and you want growth, I think that digital labor is going to be one of the catalysts to make that happen.

Shopify (NASDAQ: SHOP)

Shopify launched its first AI-powered search integration with Perplexity in 2024

Last year, we… launched our first AI-powered search integration with Perplexity, enabling new ways for buyers to find merchants.

One of Shopify’s management’s focus areas in 2025 is to continue embracing AI by investing more in Sidekick and other AI capabilities that help merchants launch and grow faster; management wants to shift Shopify towards producing goal-oriented software; management believes Shopify is well-positioned as a leader for commerce in an AI-driven world

We will continue to embrace the transformative potential of AI. This technology is not just a part of the future, it is redefining it. We’ve anticipated this. So we’re already transforming Shopify into a platform where users and machines work seamlessly together. We plan to deepen our investment in Sidekick and other AI capabilities to help not just brand-new merchants to launch, but also to help larger merchants scale faster and drive greater productivity. Our efforts to shift towards more goal-oriented software will further help to streamline operations and improve decision-making. This focus on embracing new ways of thinking and working positions us not only as the platform of choice today, but also as a leader for commerce in the AI-driven era with a relentless focus on cutting-edge technology.

Shopify’s management believes Shopify will be one of the major net beneficiaries in the AI era as the company is leveraging AI really well, such as its partnerships with Perplexity and OpenAI

I actually think Shopify will very much be one of the major net beneficiaries in this new AI era. I think we are widely recognized as one of the best companies that foster long-term partnership. And so when it comes to partnership in AI, whether it’s Perplexity, where we’re now powering their search results with incredible product across the Shopify product catalog or OpenAI where we’re using — we have a direct set of their APIs to help us internally, we are really leveraging it as best as we can.

In terms of utilising AI, Shopify’s management sees 2 angles; the 1st angle is Shopify using AI to help merchants with mundane tasks and allow merchants to focus only on the things they excel at; the 2nd angle is Shopify using AI internally to make developers and customer-support teams more effective (with customer-support teams, Shopify is using AI to handle low-quality conversations with customers)

[Question] A question in regards to AI and the use of AI internally. Over the last year or so, you’ve made significant investments. Where are you seeing it operationally having the most impact? And then what has been the magnitude of productivity gains that you’ve seen?

[Answer] We think about it in sort of 2 ways. The first is from a merchant perspective, how can we make our merchants way more successful, get them to do things faster, more effectively. So things like Sidekick or media editor or Shopify Inbox, Semantic Search, Sidekick, these are things that now — that every merchant should want when they’re not just getting started, but also scaling their business. And those are things that are only available from Shopify. So we’re trying to make some of the more mundane tasks far more easy to do and get them to focus on things that only they can — only the merchants can do. And I think that’s an important aspect of what Shopify will bring…

…Internally, however, this is where it gets really interesting, because not only can we use it to make our developers more effective, but also, if you think about our support organization, now we can ensure that our support team is actually having very high-quality conversations with merchants, whereas a lot of low-quality conversations, things like configuring a domain or a C name or a user name and password issue, that can be handled really elegantly by AI.

Taiwan Semiconductor Manufacturing Company (NYSE: TSM)

TSMC’s AI accelerators revenue more than tripled in 2024 and was mid-teens percent of overall revenue in 2024, but management expects AI accelerators revenue to double in 2025; management sees really strong AI-related demand in 2025

Revenue from AI accelerators, which we now define as AI GPU, AI ASICs and HBM controller for AI training and inference in the data center, accounted for close to mid-teens percent of our total revenue in 2024. Even after more than tripling in 2024, we forecast our revenue from AI accelerator to double in 2025 as the strong surge in AI-related demand continues…

…[Question] Try to get a bit more clarity on the cloud growth for 2025. I think, longer term, without a doubt, the technology definitely has lots of potential for demand opportunities, but I think — if we look at 2025 and 2026, I think there could be increasing uncertainties coming from maybe [indiscernible] spending, macro or even some of the supply chain challenges. And so I understand the management just provided a pretty good guidance for this year for sales to double. And so if you look at that number, do you think there is still more upside than downside as we go through 2025?

[Answer] I certainly hope there is upside, but I hope I get — my team can supply enough capacity to support it. Does that give you enough hint? 

TSMC’s management saw a mixed year of recovery for the global semiconductor industry in 2024 with strong AI-related demand but mild recovery in other areas

2024 was a mixed year of recovery for the global semiconductor industry. AI-related demand was strong, while other applications saw only a very mild recovery as macroeconomics condition weighed on consumer sentiment and the end market demand. 

TSMC’s management expects mid-40% revenue CAGR from AI accelerators in the 5-years starting from 2024 (previous forecast was for 50% CAGR for the 5-years starting from 2024, but off a lower base); management expects AI accelerators to be the strongest growth driver for TSMC’s overall HPC  platform and overall revenue over the next few years

Underpinned by our technology leadership and broader customer base, we now forecast the revenue growth from AI accelerators to approach a mid-40% CAGR for the 5-year period starting off the already higher base of 2024. We expect AI accelerators to be the strongest driver of our HPC platform growth and the largest contributor in terms of our overall incremental revenue growth in the next several years.

TSMC’s management expects 20% revenue CAGR in USD terms in the 5-years starting from 2024, driven by growth across all its platforms; management thinks that in the next few years, TSMC’s smartphone and PC end-markets will have higher silicon content and faster replacement cycle, driven by AI-related demand, which will in turn drive robust demand for TSMC’s chip manufacturing service; the AI-related demand in the smartphone and PC end-markets are related to edge-AI

Looking ahead, as the world’s most reliable and effective capacity provider, TSMC is playing a critical and integral role in the global semiconductor industry. With our technology leadership, manufacturing excellence and customer trust, we are well positioned to address the growth from the industry megatrend of 5G, AI and HPC with our differentiated technologies. For the 5-year period starting from 2024, we expect our long-term revenue growth to approach a 20% CAGR in U.S. dollar term, fueled by all 4 of our growth platform, which are smartphone, HPC, IoT and automotive…

…[Question] I believe that 20% starting from a very — already very high base in 2024 is a really good long-term objective but just wondering that, aside from the strong AI demand, what’s your view on the traditionals, applications like PC and the smartphone, growth and particularly for this year.

[Answer] This year is still mild growth for PC and smartphone, but everything is AI related, all right, so you can start to see why we have confidence to give you a close to 20% CAGR in the next 5 years. AI: You look at a smartphone. They will put AI functionality inside, and not only that. So the silicon content will be increased. In addition to that, actually the replacement cycle will be shortened. And also they need to go into the very advanced technology because of, if you want to put a lot of functionality inside a small chip, you need a much more advanced technology to put those [indiscernible]. Put all together, that even smartphone, the unit growth is almost low single digit, but then the silicon and the replacement cycle and the technology migration, that give us more growth than just unit growth; similar reason for PC…

…On the edge AI, in our observation, we found out that our customers start to put up more neural processing inside. And so we estimated the 5% to 10% more silicon being used. [ Can it be ] every year 5% to 10%? Definitely it is no, right? So they will move to next node, the technology migration. That’s also to TSMC’s advantage. Not only that, I also say that, the replacement cycle, I think it will be shortened because of, when you have a new toy that — with AI functionality inside it, everybody want replacing, replace their smartphone, replace their PCs. And [ I count that one ] much more than the — a mere 5% increase.

TSMC’s upcoming A16 process technology is best suited for specific HPC (high-performance computing) products, which means it is best suited for AI-related workloads

We will also introduce A16 featuring Super Power Rail or SPR as separate offering. TSMC’s SPR is a innovative, best-in-class backside power delivery solution that is first in the industry to incorporate a novel backside metal scheme that preserves gate density and device width flexibility to maximize product benefit. Compared with N2P, A16 provide a further 8% to 10% speed improvement at the same power or 15% to 20% power improvement at the same speed, and additional 7% to 10% chip density gain. A16 is the best suitable for specific HPC product with a complex signal route and dense power delivery network. Volume production is scheduled for second half 2026.

TSMC’s management thinks that the US government’s latest AI restrictions will only have a minimal impact on the company’s business

[Question] Overnight, the U.S. seems to put a new framework on restricting China’s AI business, right? So I’m wondering whether that will create some business impact to your China business.

[Answer] We don’t have all analysis yet, but the first look is not significantly. It’s manageable. So that meaning that, my customers who are being restricted [ or something ], we are applying for the special permit for them. And we believe that we have confidence that they will get some permission, so long as they are not in the AI area, okay, especially automotive industry. Or even you talk about crypto mining, yes.

TSMC’s management does not want to reveal the level of demand for AI-related ASICs (application-specific integrated circuits) from the cloud hyperscalers, but they are confident that the demand is real, and that the cloud hyperscalers will be working with TSMC as they all need leading-edge technology for their AI-related ASICs

[Question] Broadcom’s CEO recently laid out a large SAM for AI hyperscalers building out custom silicon. I think he was talking about million clusters from each of the customers he has in the next 2 or 3 years. What’s TSMC’s perspective on all this? 

[Answer] I’m not going to answer the question of the specific number, but let me assure you that, whether it’s ASIC or it’s graphic, they all need a very leading-edge technology. And they’re all working with TSMC, okay, so — and the second one is, is the demand real. Was — is — as a number that my customers said. I will say that the demand is very strong.

AI makes up all of the current demand for CoWoS (chip on wafer on substrate) capacity that TSMC’s management is seeing, but they think non-AI-related demand for CoWoS will come in the near future from CPUs and servers; there are rumours of a cut in orders for CoWoS, but management is not seeing any cuts; it appears that HBM (high bandwidth memory) is the key constraint on AI demand, instead of CoWoS; CoWoS was over 8% of TSMC’s revenue in 2024 and will be over 10% in 2025; CoWoS gross margin is better than before, but still lower than the corporate average

[Question] When can we see non-AI application such as server, smartphone or anything else can be — can start to adopt CoWoS capacity in case there is any fluctuation in the AI demand?

[Answer] Today is all AI focused. And we have a very tight capacity and cannot even meet customers’ need, but whether other products will adopt this kind of CoWoS approach, they will. It’s coming and we know that it’s coming.

[Question] When?

[Answer] It’s coming… On the CPU and on the server chip. Let me give you a hint…

…[Question] About your CoWoS and SoIC capacity ramp. Can you give us more color this year? Because recently there seemed to be a lot of market noises. Some add orders. Some cut orders, so I would like to see your view on the CoWoS ramp.

[Answer] That’s a rumor. I assure you. We are working very hard to meet the requirement of my customers’ demand, so “cut the order,” that won’t happen. We actually continue to increase, so we are — again I will say that. We are working very hard to increase the capacity…

…[Question] A question on AI demand. Is there a scenario where HBM is more of a constraint on the demand, rather than CoWoS which seems to be the biggest constraint at the moment? 

[Answer] I don’t comment on other supplier, but I know that we have a very tight capacity to support the AI demand. I don’t want to say I’m the bottleneck. TSMC, always working very hard with customer to meet their requirement…

…[Question] So we have observed an increasing margin of advanced packaging. Could you remind us the CoWoS contribution of last year? And do you expect the margin to kind of approach the corporate average or even exceed it after the so-called — the value reflection this year?

[Answer] Overall speaking, advanced packaging accounted for over 8% of revenue last year. And it will account for over 10% this year. In terms of gross margins, it is better. It is better than before but still below the corporate average. 

AI makes up all of the current demand for SoIC (system on integrated chips) that TSMC’s management is seeing, but they think non-AI-related demand for SoIC will come in the future

Today, SoIC’s demand is still focused on AI applications, okay? For PC or for other area, it’s coming but not right now.

Tesla (NASDAQ: TSLA)

Tesla’s management thinks Tesla’s FSD (Full Self Driving) technology has grown up a lot in the past few years; management thinks that car-use can grow from 10 hours per week to 55 hours per week with autonomous vehicles; autonomous vehicles can be used for both cargo and people delivery; FSD currently works very well in the USA , and will soon work well everywhere else; the constraint Tesla is currently experiencing with autonomous vehicles is in battery packs; FSD makes traffic commuting safer; FSD is currently on Version 13, and management believes Version 14 will have a significant step-improvement; Tesla has launched the Cortex training cluster at Gigafactory Austin, and it has played a big role in advancing FSD; Tesla will launch unsupervised FSD in June 2025 in Austin; Tesla already has thousands of its cars driving autonomously daily in its factories in Fremont and Texas, and Tesla will soon do that in Austin and elsewhere in the world; Tesla’s solution for autonomous vehicles is a generalised AI solution which does not need high-precision maps; Tesla’s unsupervised FSD work outside of Austin even when it’s launched only in June 2025 in Austin, but management just wants to be cautious; management thinks Tesla will release unsupervised FSD in many parts of the USA by end-2025; management’s safety-standard for FSD is for it to be far, far, far superior to humans; management thinks Tesla will have unsupervised FSD in almost every market this year

For a lot of people, like their experience of Tesla autonomy is like if it’s even a year old, if it’s even 2 years old, it’s like meeting someone when they’re like a toddler and thinking that they’re going to be a toddler forever. But obviously not going to be a toddler forever. They grow up. But if their last experience was like, “Oh, FSD was a toddler.” It’s like, well, it’s grown up now. Have you seen it? It’s like walks and talks…

…My #1 recommendation for anyone who doubts is simply try it. Have you tried it? When is the last time you tried it? And the only people who are skeptical, the only people who are skeptical are those who have not tried it.

So a car goes — a passenger car typically has only about 10 hours of utility per week out of 168, a very small percentage. Once that car is autonomous, my rough estimate is that it is in use for at least 1/3 of the hours per week, so call it, 50, maybe 55 hours of the week. . And it can be used for both cargo delivery and people delivery…

…That same asset, the thing that — these things that already exist with no incremental cost change, just a software update, now have 5x or more the utility than they currently have. I think this will be the largest asset value increase in human history…

…So look, the reality of autonomy is upon us. And I repeat my advice, try driving the car or let it drive you. So now it works very well in the U.S., but of course, it will, over time, work just as well everywhere else…

…Our current constraint is battery packs this year but we’re working on addressing that constraint. And I think we will make progress in addressing that constraint…

…So a bit more on full self-driving. Our Q4 vehicle safety report shows continued year-over-year improvement in safety for vehicles. So the safety numbers, if somebody has supervised full self-driving turn on or not, the safety differences are gigantic…

…People have seen the immense improvement with version 13, and with incremental versions in version 13 and then version 14 is going to be yet another step beyond that, that is very significant. We launched the Cortex training cluster at Gigafactory Austin, which was a significant contributor to FSD advancement…

…We’re going to be launching unsupervised full self-driving as a paid service in Austin in June. So I talked to the team. We feel confident in being able to do an initial launch of unsupervised, no one in the car, full self-driving in Austin in June…

…We already have Teslas operating autonomously unsupervised full self-driving at our factory in Fremont, and we’ll soon be doing that at our factory in Texas. So thousands of cars every day are driving with no one in them at our Fremont factory in California, and we’ll soon be doing that in Austin and then elsewhere in the world with the rest of our factories, which is pretty cool. And the cars aren’t just driving to exactly the same spot because, obviously, it all — [ implied ] at the same spot. The cars are actually programmed with where — with what lane they need to park into to be picked up for delivery. So they drive from the factory end of line to their destination parking spot and to be picked up for delivery to customers and then doing this reliably every day, thousands of times a day. It’s pretty cool…

…Our solution is a generalized AI solution. It does not require high-precision maps of locality. So we just want to be cautious. It’s not that it doesn’t work beyond Austin. In fact, it does. We just want to be — put our toe in the water, make sure everything is okay, then put a few more toes in the water, then put a foot in the water with safety of the general public as and those in the car as our top priority…

…I think we will most likely release unsupervised FSD in many regions of the country of the U.S. by the end of this year…

…We’re looking for a safety level that is significantly above the average human driver. So it’s not anywhere like much safer, not like a little bit safer than human, way safer than human. So the standard has to be very high because the moment there’s any kind of accident with an autonomous car, that immediately gets worldwide headlines, even though about 40,000 people die every year in car accidents in the U.S., and most of them don’t even get mentioned anywhere. But if somebody [ scrapes a shed ] within autonomous car, it’s headline news…

…But I think we’ll have unsupervised FSD in almost every market this year, limited simply by regulatory issues, not technical capability. 

Tesla’s management thinks the compute needed for Optimus will be 10x that of autonomous vehicles, even though a humanoid robot has 1,000x more uses than an autonomous vehicle; management has seen the cost of training Optimus (or AI, in general) dropping dramatically over time; management thinks Optimus can produce $10 trillion in revenue, and that will still make the training needs of $500 billion in compute a good investment; management realises their revenue projections for Optimus sound insane, but they believe in it (sounds like a startup founder trying to get funding from VCs); it’s impossible for management to predict the exact timing for Optimus because everything about the robot has to be designed and built from the ground up by Tesla (nothing could be bought off-the-shelf by Tesla), but management thinks Tesla will build a few thousand Optimus robots by end-2025, and that these robots will be doing useful work in Tesla’s factories in the same timeframe; management’s goal is to ramp up Optimus production at a far faster rate than anything has ever been ramped; Optimus can even do delicate things such as play the piano; Optimus is still not design-locked for production; Tesla might be able to deliver Optimus to external clients by 2026 H2; management is confident that at scale, Optimus will be cheaper to produce than a car

The training needs for Optimus, our Optimus humanoid robot are probably at least ultimately 10x what is needed for the car, at least to get to the full range of useful role. You can say like how many different roles are there for a humanoid robot versus a car? A humanoid robot has probably 1,000x more uses and more complex things than in a car. That doesn’t mean the training scales by 1,000 but it’s probably 10x…

…It doesn’t mean like — or Tesla’s going to spend like $500 billion in training compute because we will obviously train Optimus to do enough tasks to match the output of Optimus robots. And obviously, the cost of training is dropping dramatically with time. But it is — it’s one of those things where I think long-term, Optimus will be — Optimus has the potential to be north of $10 trillion in revenue, like it’s really bananas. So that you can obviously afford a lot of training compute in that situation. In fact, even $500 billion training compute in that situation will be quite a good deal…

…With regard to Optimus, obviously, I’m making these revenue predictions that sound absolutely insane, I realize that. But they are — I think they will prove to be accurate…

…There’s a lot of uncertainty on the exact timing because it’s not like a train arriving at the station for Optimus. We are designing the train and the station and in real time while also building the tracks. And sort of like, why didn’t the train arrive exactly at 12:05? And like we’re literally designing the train and the tracks and the station in real-time while like how can we predict this thing with absolute precision? It’s impossible. The normal internal plan calls for roughly 10,000 Optimus robots to be built this year. Will we succeed in building 10,000 exactly by the end of December this year? Probably not. But will we succeed in making several thousand? Yes, I think we will. Will those several thousand Optimus robots be doing useful things by the end of the year? Yes, I’m confident they will do useful things…

…Our goal is to run Optimus production faster than maybe anything has ever been ramped, meaning like aspirationally in order of magnitude, ramp per year. Now if we aspire to an order of magnitude ramp per year, perhaps, we only end up with a half order of magnitude per year. But that’s the kind of growth that we’re talking about. It doesn’t take very many years before we’re making 100 million of these things a year, if you go up by let’s say, a factor by 5x per year…

…This is an entirely new supply chain, it’s entirely new technology. There’s nothing off the shelf to use. We tried desperately with Optimus to use any existing motors, any actuators, sensors. Nothing worked for a humanoid robot at any price. We had to design everything from physics-first principles to work for humanoid robot and with the most sophisticated hand that has ever been made before by far. Optimus will be also able to play the piano and be able to thread a needle. I mean this is the level of precision no one has been able to achieve…

…Optimus is not design-locked. So when I say like we’re designing the train as it’s going — we’re redesigning the train as it’s going down the tracks while redesigning the tracks and the train stations…

…I think probably with version 2, it is a very rough guess because there’s so much uncertainty here, very rough guess that we start delivering Optimus robots to companies that are outside of Tesla in maybe the second half of next year, something like that…

…I’m confident at 1 million units a year, that the production cost of Optimus will be less than $20,000. If you compare the complexity of Optimus to the complexity of a car, so just the total mass and complexity of Optimus is much less than a car.

The buildout of Cortex accelerated the rollout of FSD Version 13; Tesla has invested $5 billion so far in total AI-related capex

The build-out of Cortex was accelerated because of the role — actually accelerate the rollout of FSD Version 13. Our cumulative AI-related CapEx, including infrastructure, so far has been approximately $5 billion. 

Tesla’s management is seeing significant interest from some car manufacturers in licensing Tesla’s FSD technology; management thinks that car manufacturers without FSD technology will go bust; management will only entertain situations where the volume would be very high

What we’re seeing is at this point, significant interest from a number of major car companies about licensing Tesla full self-driving technology…

…We’re only going to entertain situations where the volume would be very high. Otherwise, it’s just not worth the complexity. And we will not burden our engineering team with laborious discussions with other engineering teams until we obviously have unsupervised full self-driving working throughout the United States. I think the interest level from other manufacturers to license FSD will be extremely high once it is obvious that unless you have FSD, you’re dead.

Compared to Version 13, Version 14 of FSD will have a larger model size, longer context length, more memory, more driving-context, and more data on tricky corner cases

[Question] What technical breakthroughs will define V14 of FSD, given that V13 already covered photon to control? 

[Answer] So continuing to scale the model size a lot. We scale a bunch in V13 but then there’s still room to grow. So we’re going to continue to scale the model size. We’re going to increase the context length even more. The memory sort of like limited right now. We’re going to increase the amount of memory [indiscernible] minutes of context for driving. They’re going to add audio and emergency vehicles better. Add like data of the tricky corner cases that we get from the entire fleet, any interventions or any kind of like user intervention. We just add to the data set. So scaling in basically every access, training compute, [ asset ] size, model size, model context and also all the reinforcement learning objectives.

Tesla has difficulties training AI models for autonomous vehicles in China because the country previously did not allow Tesla to transfer training videos outside of China while the US government did not allow Tesla to do training in China; a workaround Tesla did was to train on publicly available videos of streets in China; Tesla also had to build a simulator for its AI models to train on bus lanes in China because they are complicated 

In China, which is a gigantic market, we do have some challenges because they weren’t — they currently allow us to transfer training video outside of China. And then the U.S. government won’t let us do training in China. So we’re in a bit of a buying there. It’s like a bit of a quandary. So what we’re really solving then is by literally looking at videos of streets in China that are available on the Internet to understand and then feeding that into our video training so that publicly available video of street signs and traffic rules in China can be used for training and then also putting it in a very accurate simulator. And so it will train using SIM for bus lanes in China. Like bus lanes in China, by the way, one of our biggest challenges in making FSD work in China is the bus lanes are very complicated. And there’s like literally like hours of the day that you’re allowed to be there and not be there. And then if you accidently go in at a bus lane at the wrong time, you get an automatic ticket instantly. And so it was kind of a big deal, bus lanes in China. So we put that into our simulator train on that, the car has to know what time of the day it is, read the sign. We’ll get this solved.

Elon Musk knows LiDAR technology really well because he built a LiDAR system for SpaceX that is in-use at the moment, but he thinks LiDAR is simply the wrong technology to use for autonomous vehicles because it has issues, and because humans are driving vehicles simply with our eyes and our biological neural nets

[Question] You’ve said in the past about LiDAR, for EVs, that LiDAR is a crutch, a fool’s errand. I think you even told me once, even if it was free, you’d say you wouldn’t use it. Do you still feel that way?

[Answer] Obviously humans drive without shooting lasers out of their eyes. I mean unless you’re superman. But like humans drive just with passive visual — humans drive with eyes and a neural net — and a brain neural net, sort of biological — so the digital equivalent of eyes and a brain are cameras and digital neural nets or AI. So that’s — the entire road system was designed for passive optical neural nets. That’s how the whole real system was designed and what everyone is expecting, that’s how we expect other cars to behave. So therefore, that is very obviously a solution for full self-driving as a generalized — but the generalized solution for full self-driving as opposed to the very specific neighborhood-by-neighborhood solution, which is very difficult to maintain, which is what our competitors are doing…

…LiDAR has a lot of issues. Like SpaceX Dragon docks with the space station using LiDAR, that’s a program that I personally spearheaded. I don’t have some fundamental bizarre dislike of LiDAR. It’s simply the wrong solution for driving cars on roads…

…I literally designed and built our own red LiDAR. I oversaw the project, the engineering thing. It was my decision to use LiDAR on Dragon. And I oversaw that engineering project directly. So I’m like we literally designed and made a LiDAR to dock with the space station. If I thought it was the right solution for cars, I would do that, but it isn’t.

The Trade Desk (NASDAQ: TTD)

Trade Desk’s management continues to invest in AI and thinks that AI is game-changing for forecasting and insights on identity and measurement; Trade Desk’s AI efforts started in 2017 with Koa, but management sees much bigger opportunities today; management is asking every development team in Trade Desk to look for opportunities to introduce AI into Trade Desk’s platform; there are already hundreds of AI enhancements to Trade Desk’s platform that have been shipped or are going to be shipped in 2025

AI is providing next-level performance in targeting and optimization, but it is also particularly game-changing in forecasting and identity and measurement. We continue to look at our technology stack and ask, where can we inject AI and enhance our product and client outcomes? Over and over again, we are finding new opportunities to make AI investments…

…We started our ML and AI efforts in 2017 with the launch of Koa, but today, the opportunities are much bigger. We’re asking every scrum inside of our company to look for opportunities to inject AI into our platform. Hundreds of enhancements recently shipped and coming in 2025 would not be possible without AI. We must keep the pedal to the metal, not to chest them on stages, which everyone else seems to be doing, but instead to produce results and win share.

Wix (NASDAQ: WIX)

Wix’s AI Website Builder was launched in 2024 and has driven stronger conversion and purchase behaviour from users; more than 1 million sites have been created and published with AI Website Builder; most new Wix users today are creating their websites through Wix’s AI tools and AI Website Builder and these users have higher rates of converting to paid subscribers

2024 was also the year of AI innovation. In addition to the significant number of AI tools introduced, we notably launched our AI Website Builder, the new generation of our previous AI site builder introduced in 2016. The new AI Website Builder continues to drive demonstrably stronger conversion and purchase behavior…

…Over 1 million sites have been created and published with the Website Builder…

…Most new users today are creating their websites through our AI-powered onboarding process and Website Builder which is leading to a meaningful increase in conversion of free users to paid subscriptions, particularly among Self Creators.

Wix’s management launched Wix’s first directly monetised AI product – AI Site Chat – in December 2024; AI Site Chat will help businesses converse with customers round the clock; users of AI Site Chat have free limited access, with an option to pay for additional usage; AI Site Chat’s preliminary results look very promising

In December, we also rolled out our first directly monetized AI product – the AI Site Chat…

…The AI Site-Chat was launched mid-December to Wix users in English, providing businesses with the ability to connect with visitors 24/7, answer their questions, and provide relevant information in real time, even when business owners are unavailable. By enhancing availability and engagement on their websites, the feature empowers businesses to meet the needs of their customers around the clock, ultimately improving the customer experience and driving potential sales. Users have free limited access with the option to upgrade to premium plans for additional usage…

…So if you’re a Wix customer, you can now install a chat, AI-powered chat on your website, and this will handle customer requests, product inquiries and support request. And from — and again, it’s very early in days and the preliminary results, but it looks very promising. 

AI agents and assistants are an important part of management’s product roadmap for Wix in 2025; Wix is testing (1) an AI assistant for its Wix Business Manager dashboard, and (2) Marketing Agent, a directly monetizable AI agent that helps users accomplish marketing tasks; Marketing Agent is the first of a number of specialised AI agents management will roll out in 2025; management intends to test monetisation opportunities with the new AI agents

AI remains a major part of our 2025 product roadmap with particular focus on AI-powered agents and assistants…

… Currently, we are testing our AI Assistant within the Wix Business Manager as well as our AI Marketing Agent.

The AI Assistant in the Wix Business Manager is a seamlessly integrated chat interface within the dashboard. Acting as a trusted aide, this assistant guides users through their management journey by providing answers to questions and valuable insights about their site. With its comprehensive knowledge, the AI Assistant empowers users to better understand and leverage available resources, assisting with site operations and business tasks. For instance, it can suggest content options, address support inquiries, and analyze analytics—all from a single entry point.

The AI Marketing Agent helps businesses to market themselves online by proactively generating tailored marketing plans that align with users’ goals and target audiences. By analyzing data from their website, the AI delivers personalized strategies to enhance SEO, create engaging content, manage social media, run email campaigns and optimize paid advertising—all with minimal effort from the user. This solution not only simplifies marketing but also drives Wix’s monetization strategy, seamlessly guiding users toward high-impact paid advertising and premium marketing solutions. As businesses invest in growing their online presence, Wix benefits through a share of ad spend and premium feature adoption—fueling both user success and revenue growth.

We will continue to release and optimize specialized AI agents that assist our users in building the online presence they envision. We are exploring various monetization strategies as we fully roll out these agents and adoption increases.

Wix’s management is seeing Wix’s gross margin improve because of AI integration in customer care

Creative Subscriptions non-GAAP gross margin improved to 85% in Q4’24 and to 84% for the full year 2024, up from 82% in 2023. Business Solutions non-GAAP gross margin increased to 32% in Q4’24 and to slightly above 30% for the full year 2024. Continued gross margin expansion is the product of multiple years of cost structure optimization and efficiencies from AI integration across our Customer Care operations.

Wix’s management believes the opportunity for Wix in the AI era is bigger than what came before

There’s a lot of discussions about a lot of theories about it. But I really believe that the opportunity there is bigger than anything else because what we have today are going to continue to dramatically evolve into something that is probably more powerful and more enabling for small businesses to be successful. Overall, the Internet has a tendency to do it every 10 years or so, right, in the ’90s, the Internet started and became HTML, then it became images and then later on videos and then it became mobile, right? And I think they became interactive, everything become an application, kind of an application. And I think how website will look at the AI universe is the next step, and I think there’s a lot of exciting things we can offer our users there.

Visa (NYSE: V)

Visa is an early adopter of AI and management continues to drive adoption; Visa has seen material gains in engineering productivity; Visa has deployed AI in many functions, such as analytics, sales, finance, and marketing

We were very early adopters of artificial intelligence, and we continue to drive hard at the adoption of generative AI as we have for the last couple of years. So we’ve been working to embed AI and AI tooling into our company’s operations, I guess, broadly. We’ve seen material gains in productivity, particularly in our engineering teams. We’ve deployed AI tooling in client services, sales, finance, marketing, really everywhere across the company. And we were a very early adopter of applied AI in the analytics and modeling space, very early by like decades, we’ve been using AI in that space. So our data science and risk management teams have, at this point, decades of applied experience with AI, and they’re aggressively adopting the current generations of AI technology to enhance both our internal and our market-facing predictive and detective modeling capabilities. Our product teams are also aggressively adopting gen AI to build and ship new products.

Zoom Communications (NASDAQ: ZM)

Zoom AI Companion’s monthly active users (MAUs) grew 68% quarter-on-quarter; management has added new agentic AI capabilities to Zoom AI Companion; management will launch the Custom AI Companion add-on in April 2025; management will launch AI Companion for clinicians in March 2025; Zoom AI Companion is added into a low-end Zoom subscription plan at no added cost, and customers do not want to leave their subscriptions because of the added benefit of Zoom AI Companion; Zoom will be monetising Zoom AI Companion from April 2025 onwards through the Custom AI Companion add-on; the Custom AI Companion add-on would be $12 a seat when it’s launched in April 2025 and management thinks this price would provide a really compelling TCO (total cost of ownership) for customers; management thinks Custom AI Companion would have a bigger impact on Zoom’s revenue in 2026 (FY2027) than in 2025 (FY2026); see Point 28 for use cases for Custom AI Companion

Growth in monthly active users of Zoom AI Companion has accelerated to 68% quarter-over-quarter, demonstrating the real value AI is providing customers…

…As part of AI Companion 2.0, we added advanced agentic capabilities, including memory, reasoning, orchestration and a seamless integration with Microsoft and Google services. In April, we’re launching Custom AI Companion add-on to automate workplace tasks through custom agents. This will personalize AI to fit customer needs, connect with their existing data, and work seamlessly with their third-party tools. We’re also enhancing Zoom Workplace for Clinicians with an upgraded AI Companion that will enable clinical note-taking capabilities and specialized medical features for healthcare providers starting in March…

…If you look at our low SMB customer online buyers, AI Companion is part of that at no additional cost, made our service very sticky and also the customers give a very basic example, like meeting summary, right? It works so well, more and more customers follow the value…

…For high end, for sure, and we understand that today’s AI Companion and additional cost we cannot monetize. However, in April, we are going to announce the customized Companion for interested customers. We can monetize…

…[Question] So in April, when the AI customization, the AI Companion becomes available, I think it’s $11 or $12 a seat. Can you maybe help us understand how you’re thinking about like what’s the real use case?

[Answer] In regards to your question about what are sort of the assumptions or what’s the targeting in our [ head ] with the $12 Custom AI Companion SKU. I would say, starting with enterprise customers, obviously, the easiest place to sort of pounce on them is our own customer base and talk about that, but certainly not just limited to that. But we’ll be probably giving a lot more, I would say, at Enterprise Connect, which you can see on the thing there. But I would say we’ve assumed some degree of monetization in FY ’26, I think you’ll see more of it in ’27. And we think that the $12 price point is going to be a really compelling TCO story for our customers, it’s differentiated from what others in the market are pricing now. 

The Zoom Virtual Agent feature will soon be able to handle complex tasks

Zoom Virtual Agent will soon expand reasoning abilities to handle complex tasks while maintaining conversational context for more natural and helpful outcomes.

Zoom’s management believes Zoom is uniquely positioned to win in agentic AI for a few reasons, including Zoom having exception context of users’ ongoing conversations, and Zoom’s federated AI approach where the company can use the best models for each task

We’re uniquely positioned to succeed in agentic AI for several reasons:

● Zoom is a system of engagement for our users with recent information in ongoing conversations. This exceptional context along with user engagement allows us to drive greater value for customers.

● Our federated AI approach lets us combine the best models for each task. We can use specialized small language models where appropriate, while leveraging larger models for more complex reasoning – driving both quality and cost efficiency

Zoom’s management is seeing large businesses want to use Zoom because of the AI features of its products

You take a Contact Center, for example, why we are winning? Because a lot of AI features like AI Expert Assist. AI, a lot of features built into our quality management and so on and so forth. 

Zoom’s management sees Zoom’s AI business services as a great way to monetise AI

You take a Contact Center, for example, why we are winning? Because a lot of AI features like AI Expert Assist. AI, a lot of features built into our quality management and so on and so forth. But all those business services, that’s another great way for us to monetize AI.

Zoom’s management thinks Zoom’s cost of ownership with AI is lower than what competitors are offering

And I look at our AI Companion, all those AI Companion core features today at no additional cost, right? And customer really like it because of the quality, they’re getting better and better every quarter and very useful, right? Not like some other competitors, right? They talk about their AI strategy and when customers realize that, wow, it’s very expensive. And the total cost of ownership is not getting better because cost of the value is not [ great ], but also it’s not [ free ] and they always try to increase price.

A good example of a use case for Custom AI Companion

[Question] So in April, when the AI customization, the AI Companion becomes available, I think it’s $11 or $12 a seat. Can you maybe help us understand how you’re thinking about like what’s the real use case?

[Answer] So regarding the Custom AI Combined on use cases, high levels, we give a customer ability to customize their needs. I’ll give a few examples. One feature like we have a Zoom Service Call video clip, and we are going to support the standard template, right? How to support every customer? They have a customized template for each of the users, and this is a part of combining AI Studio, right? And also all kinds of third-party integration, right? And they like they prefer, right, some of those kind of sort of third-party application integration. With their data, with the knowledge, whether the [ big scenery ], a lot of things, right? Each company is different, they would not customized, so we can leverage our combining studio to work together with the customer to support their needs and also at same time commodities.

Zoom’s management expects the cost from AI usage to increase and so that will impact Zoom’s margins in the future, but management is also building efficiencies to offset the higher cost of AI

[Question] As we think about a shift more towards AI contribution, aren’t we shifting more towards a consumption model rather than a seat model over time, why wouldn’t we see margin compression longer term?

[Answer] Around how to think about margins and business models and why we don’t see compression. And what I would say is that — what we expect to see is similar to what you saw in FY ’25, which is we’re seeing obvious increase in cost from AI.  And that we have an ongoing methodical kind of efficiency list to offset, and we certainly expect that broadly to continue into FY ’26. So I think we feel good about our ability to kind of moderate that. There’s other things we do more holistically where we can offset stuff that’s maybe not in AI in our margins, things like [ colos ], et cetera, that we’ve talked about previously. 

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Microsoft, Paycom Software, PayPal, Salesforce, Shopify, TSMC, Tesla, The Trade Desk, Wix, Visa, and Zoom. Holdings are subject to change at any time.

The Latest Thoughts From American Technology Companies On AI (2024 Q4) – Part 1

A collection of quotes on artificial intelligence, or AI, from the management teams of US-listed technology companies in the 2024 Q4 earnings season.

The way I see it, artificial intelligence (or AI), really leapt into the zeitgeist in late-2022 or early-2023 with the public introduction of DALL-E2 and ChatGPT. Both are provided by OpenAI and are software products that use AI to generate art and writing, respectively (and often at astounding quality). Since then, developments in AI have progressed at a breathtaking pace.

With the latest earnings season for the US stock market – for the fourth quarter of 2024 – coming to its tail-end, I thought it would be useful to collate some of the interesting commentary I’ve come across in earnings conference calls, from the leaders of technology companies that I follow or have a vested interest in, on the topic of AI and how the technology could impact their industry and the business world writ large. This is an ongoing series. For the older commentary:

I’ve split the latest commentary into two parts for the sake of brevity. This is Part 1, and you can Part 2 here. With that, I’ll let the management teams take the stand… 

Airbnb (NASDAQ: ABNB)

Airbnb’s management thinks AI is early and has yet to fundamentally change the travel market for any of the large travel platforms; most travel companies are starting with AI on trip planning but management thinks AI is still too early for trip planning

Here’s what I think about AI. I think it’s still really early. It’s probably similar to like the mid- to late ’90s for the Internet. So I think it’s going to have a profound impact on travel, but I don’t think it’s yet fundamentally changed for any of the large travel platforms…

…So most companies, what they’re actually doing is they’re doing integrations of these other platforms on trip planning. But the trip planning, it’s still early. I don’t think it’s quite bit ready for prime time.

Airbnb’s management is starting with AI in customer service; Airbnb will roll out AI-powered customer support later in 2025; management thinks AI can provide great customer support partly because it can speak all languages 24/7 and read thousands of pages of documents; management will eventually graduate the customer-support AI into a travel and living concierge; management thinks AI can help improve efficiency at Airbnb in customer service

We’re actually starting with customer service. So later this year, we’re going to be rolling out, as part of our Summer Release, AI-powered customer support. As you imagine, we get millions of contacts every year. AI can do an incredible job of customer service. It can speak every language 24/7. It can read a corpus of thousands of pages of documents. And so we’re starting with customer support. And over the coming years, what we’re going to do is we’re going to take the AI-powered customer service agent, and we’re going to bring it into essentially Airbnb search to eventually graduate to be a travel and living concierge…

…[Question] With respect to the AI, I appreciate your answer with respect to outward-looking and how it might change the landscape. What do you think the potential is internally to apply AI for efficiencies inside the company and create an additional layer of potential margin efficiency and/or free cash flow conversion in the years ahead?

[Answer] There’s like a couple like efficiencies that you could imagine at Airbnb. One is obviously customer service. I think that’s like one of the biggest ones. I’ve kind of already covered that, but I think that’s like a massive change for Airbnb.

Airbnb’s management thinks that AI models are getting cheaper and are starting to be commoditised

I think it’s a really exciting time in the space because you’ve seen like with DeepSeek and more competition with models is models are getting cheaper or nearly free. They’re getting faster and they’re getting more intelligent. And they are, for all intent and purpose, starting to get commoditized.

Airbnb’s management thinks that a lot of the value from AI is going to accrue to platforms, and they want Airbnb to be the platform for travel and living that will reap most of the value from AI

What I think that means is a lot of value is going to accrue to the platform. And ultimately, I think the best platform, the best applications are going to be the ones that like most accrue the value from AI. And I think we’re going to be the one to do that with traveling and living.

Airbnb’s management thinks AI can help improve efficiency at Airbnb in engineering productivity; in the short-term, the improvement in engineering productivity has not been material; over the next few years, management thinks AI can drive a 30% increase in engineering productivity at Airbnb; over the long-term, management thinks there can an order of magnitude more productivity; management think younger, more innovative companies, could benefit from AI more than incumbent enterprises

[Question] With respect to the AI, I appreciate your answer with respect to outward-looking and how it might change the landscape. What do you think the potential is internally to apply AI for efficiencies inside the company and create an additional layer of potential margin efficiency and/or free cash flow conversion in the years ahead?

[Answer] The other, I assume, you refer to is essentially engineering productivity. We are seeing some productivity gains. I’ve talked to a lot of other tech CEOs, and here’s what I’ve heard talking to other like tech CEOs. Most of them haven’t seen a material like change in engineering productivity. Most of the engineers are using AI tools. They’re seeing some productivity. I don’t think it’s flowing to like a fundamental step-change in productivity yet. I think a lot of us believe in some kind of medium term of a few years, you could easily see like a 30% increase in technology and engineering productivity. And then, of course, beyond that, I mean, I think it could be like an order of magnitude more productivity because — but that’s going to be like down the road. And I think that’s going to be something that almost all companies benefit from. I think the kind of younger, more innovative, startup-like companies might benefit a little bit more because they’ll have engineers who are more likely to adopt the tools.

Alphabet (NASDAQ: GOOG)

AI Overviews in Search is now available in more than 100 countries; AI Overviews drive higher user satisfaction and search usage; Google’s Gemini model is being used in AI Overviews; with AI Overviews, usage growth of Search is growing over time, especially with younger users; management recently launched ads in AI Overviews and AI Overviews is currently monetising at nearly the same rate as Google Search; Google Search has continued to perform well in this AI age, as overall usage has continued to grow, with stronger growth seen in AI Overviews across all segments

In Search, AI overviews are now available in more than 100 countries. They continue to drive higher satisfaction and search usage…

…That includes Search where Gemini is pairing our AI overviews. People use search more with AI overviews and usage growth increases over-time as people learn that they can ask new types of questions. This behavior is even more pronounced with younger users who really appreciate the speed and efficiency of this new format…

…We’ve already started testing Gemini 2.0 in AI overviews and plan to roll it out more broadly later in the year…

…We recently launched the ads within AI Overviews on mobile in the U.S., which builds on our previous rollout of ads above and below. And as I talked about before, for the AI Overviews, overall, we actually see monetization at approximately the same rate, which I think really gives us a strong base on which we can innovate even more…

…On Search usage, overall, our metrics are healthy. We are continuing to see growth in Search on a year-on-year basis in terms of overall usage. Of course, within that, AI Overviews has seen stronger growth, particularly across all segments of users, including younger users, so it’s being well received. But overall, I think through this AI moment, I think Search is continuing to perform well.

Circle to Search is now available on more than 200 million Android devices; Circle to Search is opening new Search use cases; Circle to Search is popular with younger users; Circle to Search is used to start more than 10% of searches among users who have tried it before

Circle to Search is now available on over 200 million Android devices…

…Circle to Search is driving additional Search use and opening up even more types of questions. This feature is also popular among younger users. Those who have tried Circle to Search before now use it to start more than 10% of their searches…

…In Search, we’re seeing people increasingly ask entirely new questions using their voice, camera or in ways that were not possible before, like with Circle to search.

Alphabet’s management believes Google has a unique infrastructure advantage in AI because the company has developed each component of its technology stack; Alphabet broke ground on 11 new cloud regions and data center campuses in 2024, and announced plans for 7 new subsea cable projects; Google data centers now deliver 4x more computing power per unit of electricity compared to 5 years ago

We have a unique advantage because we develop every component of our technology stack, including hardware, compilers, models and products. This approach allows us to drive efficiencies at every level from training and serving to develop our productivity. In 2024, we broke ground on 11 new cloud regions and data center campuses in places like South Carolina, Indiana, Missouri and around the world.

We also announced plans for seven new subsea cable projects, strengthening global connectivity. Our leading infrastructure is also among the world’s most efficient. Google data centers deliver nearly four times more computing power per unit of electricity compared to just five years ago.

Google Cloud customers consume 8x more compute capacity for training and inference compared to 18 months ago; first-time commitments to Google Cloud more than doubled in 2024; Google Cloud closed a few deals in 2024 worth more than $1 billion each; Google Cloud’s AI hypercomputer utilises both GPUs (graphics processing units) and TPUs (tensor processing units), and has helped Wayfair improve performance and scalability by 25%; Google saw strong uptake of Trillium, its 6th generation TPU, in 2024 Q4; Trillium is 4x better in training and has 3x higher inference throughput than the 5th generation TPU; Google Cloud is offering NVIDIA’s H200 GPUs to customers; Google Cloud is the first cloud provider to provide NVIDIA’s Blackwell GPUs; the capex for Google Cloud is mostly for Google’s own self-designed data centers and TPUs (tensor processing units)

Today, cloud customers consume more than eight times the compute capacity for training and inferencing compared to 18 months ago…

…In 2024, the number of first-time commitments more than double compared to 2023…

…Last year, we closed several strategic deals over $1 billion and the number of deals over $250 million doubled from the prior year…

…We continue to see strong growth across our broad portfolio of AI-powered cloud solutions. It begins with our AI hypercomputer, which delivers leading performance and cost across both GPUs and TPUs. These advantages help Citadel with modeling markets and training and enabled Wayfair to modernize its platform, improving performance and scalability by nearly 25%. 

In Q4, we saw strong uptake of Trillium, our sixth-generation TPU, which delivers four times better training performance and three times greater inference throughput compared to the previous generation. We also continue our strong relationship with NVIDIA. We recently delivered their H200 based platforms to customers. And just last week, we were the first to announce a customer running on the highly-anticipated Blackwell platform…

……Our strategy is mostly to rely our own self-design and build data centers. So, they’re industry-leading in terms of both cost and power efficiency at scale. We have our own customized TPUs. They’re customized for our own workload, so they do deliver outstanding the superior performance and capex efficiency. So, we’re going to be looking at all that when we make decisions as to how we’re going to progress capital investments throughout the coming years.

Google launched an experimental version of its Gemini 2.0 Flash model in December 2024, but the model will be generally available for developers and customers; Google debuted its experimental Gemini 2.0 Flash Thinking model in late-2024 and it has gathered extremely positive reviews; Google is working on even better thinking models; Gemini 2.0’s advances in multimodality and native tool use helps Google build a universal AI assistant; an example of this universal assistant can be seen in Deep Research; Deep Research was launched in Gemini Advanced in December and is being rolled out to Android users globally; the consumer Gemini app debuted on iOS in November 2024 and has seen great product momentum; Project Mariner and Project Astra are AI agent products currently being tested and they will appear in the Gemini app sometime in 2025; Gemini and Google’s video and image generation models consistently excel in industry leaderboards and benchmarks; 4.4 million developers are using Gemini models today, double from just six months ago; Google has 7 products with over 2 billion users each, and all 7 products use Gemini; all Google Workspace business and enterprise customers were recently given access to all of Gemini’s AI capabilities; Gemini 2.0 Flash is one of the most capable models people can access for free; management’s current intention for monetisation of Gemini is through subscriptions and improving the user experience, but they have an eye on advertising revenues

In December, we unveiled Gemini 2.0, our most capable AI model yet, built for the agent era. We launched an experimental version of Gemini 2.0 Flash, our workhorse model with low-latency and enhanced performance. Flash has already rolled-out to the Gemini app, and tomorrow we are making 2.0 Flash generally available for developers and customers, along with other model updates…

…Late last year, we also debuted our experimental Gemini 2.0 Flash Thinking model. The progress to scale thinking has been super-fast and the review so-far have been extremely positive. We are working on even better thinking models and look-forward to sharing those with the developer community soon.

Gemini 2.0’s advances in multimodality and native tool use enable us to build new agents that bring us closer to our vision of a universal assistant. One early example is deep research. It uses agent capabilities to explore complex topics on your behalf and give key findings along with sources. It launched in Gemini Advanced in December and is rolling out to Android users all over the world.

We are seeing great product momentum with our consumer Gemini app, which debuted on iOS last November.

…We have opened up trusted tester access to a handful of research prototypes, including Project Mariner, which can understand and reason across information on a browser screen to complete tasks and Project Astra. We expect to bring features from both to the Gemini app later this year…

…Veo2, our state-of-the-art video generation model and Imagine3, our highest-quality text image model. These generative media models as well as Gemini consistently top industry leaderboards and score top marks across industry benchmarks. That’s why more than 4.4 million developers are using our Gemini models today, double the number from just six months ago…

…We have seven products and platforms with over 2 billion users and all are using Gemini…

…We recently gave all Google Workspace business and enterprise customers access to all of our powerful Gemini AI capabilities to help boost their productivity…

…2.0 Flash. I mean, I think that’s one of the most capable models you can access at the free tier…

…[Question] How should we think about the future monetization opportunity of Gemini? Today, it’s really a premium subscription offering or a free offering. Over time, do you see an ad component?

[Answer] On the monetization side, obviously, for now, we are focused on a free tier and subscriptions. But obviously, as you’ve seen in Google over time, we always want to lead with user experience. And we do have very good ideas for native ad concepts, but you’ll see us lead with the user experience. And — but I do think we’re always committed to making the products work and reach billions of users at scale. And advertising has been a great aspect of that strategy. And so, just like you’ve seen with YouTube, we’ll give people options over time. But for this year, I think you’ll see us be focused on the subscription direction.

Google Cloud’s AI developer platform, Vertex AI, had a 5x year-on-year increase in customers in 2024 Q4; Vertex AI offers more than 200 foundation models form Google; Vertex AI’s usage grow 20x in 2024

Our AI developer platform, Vertex AI, saw a 5x increase in customers year-over-year with brands like, International and WPP building new applications and benefiting from our 200 plus foundation models. Vertex usage increased 20x during 2024 with particularly strong developer adoption of Gemini Flash, Gemini 2.0, and most recently VEO.

Alphabet’s management will be introducing Veo2, Google’s video generation model, for creators in Youtube in 2025; advertisers around the world can now promote Youtube creator videos and ad campaigns across all AI-powered campaign types and Google ads

Expanding on our state-of-the-art video generation model, we announced Veo2, which creates incredibly high-quality video in a wide range of subjects and styles. It’s been inspiring to see how people are experimenting with it. We’ll make it available to creators on YouTube in the coming months…

…. All advertisers globally can now promote YouTube creator videos and ad campaigns across all AI-powered campaign types and Google Ads, and creators can tag partners in their brand videos…

Alphabet’s management announced the first beta of Android 16 in January 2025; there will be deeper Gemini integration for the new Samsung Galaxy S25 smartphone series; Alphabet has announced Android XR, the first Android platform for the Gemini era

Last month, we announced the first beta of Android 16 plus new Android updates, including a deeper Gemini integration coming to the new Samsung Galaxy S25 series. We also recently-announced Android XR, the first Android platform built for the Gemini era. Created with Samsung and Qualcomm, Android XR is designed to power an ecosystem of next-generation extended reality devices like headsets and glasses.

Waymo is now serving more than 150,000 trips per week (was 150,000 in 2024 Q3); Waymo is expanding in new markets in the USA this year and in 2026; Waymo will soon be launched in Tokyo; Waymo is developing its 6th-gen driver, which will significantly reduce hardware costs

It’s now averaging over 150,000 trips each week and growing. Looking ahead, Waymo will be expanding its network and operations partnerships to open up new markets, including Austin and Atlanta this year and Miami next year. And in the coming weeks, Waymo One vehicles will arrive in Tokyo for their first international road trip. We are also developing the sixth-generation Waymo driver, which will significantly lower hardware costs.

Alphabet’s management introduced a new Google shopping experience, infused with AI, in 2024 Q4, and there was 13% more daily active users in Google shopping in December 2024 compared to a year ago; the new Google Shopping experience helps users speed up their shopping research

Google is already present in over half of journeys where a new brand, product or retailer are discovered by offering new ways for people to search, we’re expanding commercial opportunities for our advertisers…

…Retail was particularly strong this holiday season, especially on Black Friday and Cyber Monday, which each generated over $1 billion in ad revenue. Interestingly, despite the U.S. holiday shopping season being the shortest since 2019, retail sales began much earlier in October, causing the season to extend longer than anticipated.

People shop more than 1 billion times a day across Google. Last quarter, we introduced a reinvented Google shopping experience, rebuilt from the ground up with AI. This December saw roughly 13% more daily active users in Google shopping in the U.S., compared to the same period in 2023…

…The new Google Shopping experiences specifically to your question, users to really intelligently show the most relevant products, helping to speed up and simplify your research. You get an AI-generated brief with top things to consider for your search plus maybe products that meet your needs. So, shoppers very often want low prices. So, the new page not only includes like deal-finding tools like price comparison, price insights, price tracking throughout. But it’s also a new and dedicated personalized deals page, which can browse deals for you, and all this is really built on the backbone of AI.

Shoppers can now take a photo and use Lens to quickly find information about the product; Lens is now used in over 20 billion visual searches per month (was over 20 billion in 2024 Q3); majority of Lens searches are incremental

Shoppers can now take a photo of a product and using Lens quickly find information about the product, reviews, similar products and where they can get it for a great price. Lens is used for over 20 billion visual search queries every month and the majority of these searches are incremental.

Alphabet’s management continues to infuse AI capabilities into Google’s advertising business; Petco used Demand Gen campaigns to achieve a 275% increase in return on ad spend and a 74% increase in click-through rates compared to social benchmarks; Youtube Select Creator Takeovers is now generally available in the US and will be rolled out across the world; PMax was recently strengthened with new controls and easier reporting functions; Event Ticket Center used PMax and saw a 5x increase in production of creative assets, driving a 300% increase in conversions compared to using manual assets; Meridian, Google’s marketing mix model, was recently made generally available and it delivers 17% higher return on advertising spend on Youtube compared to manual campaigns

We continue investing in AI capabilities across media buying, creative and measurement. As I’ve said before, we believe that AI will revolutionize every part of the marketing value chain.

And over the past quarter, we’ve seen how our customers are increasingly focusing on optimizing the use of AI. As an example, [ Petco ], used Demand Gen campaigns across targeting, creative generation and bidding to find new pet parent audiences across YouTube. They achieved a 275% higher return on ad spend and a 74% higher click-through rate than their social benchmarks.

On media buying, we made YouTube Select Creator Takeovers generally available in the U.S. and will be expanding to more markets this year. Creators know their audience the best and creator takeovers help businesses connect with consumers through authentic and relevant content.

Looking at Creative, we introduced new controls and made reporting easier in PMax, helping customers better understand and reinvest into their best-performing assets. Using asset generation in PMax, Event Ticket Center achieved a 5x increase in production of creative assets saving time and effort. They also increased conversions by 300% compared to the previous period when they used manual assets…

…Last week, we made Meridian, our marketing mix model, generally available for customers, helping more business reinvest into creative and media buying strategies that they know work. Based on the Nielsen meta analysis of marketing mix models, on average, Google AI-powered video campaigns on YouTube delivered 17% higher return on advertising spend than manual campaigns.

Sephora used Demand Gen Shorts-only channel for advertising that drove an 82% increase in searches for Sephora Holiday

Sephora used demand gen Shorts-only channel to boost traffic and brand searches for the holiday gift guide campaign and leverage greater collaborations to find the best gift. This drove an 82% relative uplift in searches for Sephora holiday.

Citi is using Google Cloud for its generative AI initiatives across customer service, document summarisation, and search

Another expanding partnership is with Citi, who is modernizing its technology infrastructure with Google Cloud to transform employee and customer experiences. Using Google Cloud, it will improve its digital products, streamline employee workflows and use advanced high-performance computing to enable millions of daily computations. This partnership also fuels Citi’s generate AI initiatives across customer service, document summarization and search to reduce manual processing.

Google Cloud had 30% revenue growth in 2024 Q4 (was 35% in 2024 Q3) driven by growth in core GCP products, AI infrastructure, and generative AI solutions; operating margin was 17.5% (was 17% in 2024 Q3 and was 9.4% in 2023 Q4); GCP grew at a much higher rate than Google Cloud overall; Google Cloud had more AI demand than capacity in 2024 Q4; management is thinking about Google’s capital intensity, but they want to invest because they are seeing strong AI demand both internally and externally; the capex Google is making can be repurposed across its different businesses

Turning to the Google Cloud segment, which continued to deliver very strong results this quarter. Revenue increased by 30% to $12 billion in the fourth quarter, reflecting growth in GCP, across core GCP products, AI infrastructure, and generative AI solutions. Once again, GCP grew at a rate that was much higher than cloud overall. Healthy Google Workspace growth was primarily driven by increase in average revenue per seat. Google Cloud operating income increased to $2.1 billion and operating margin increased from 9.4% to 17.5%…

…We do see and have been seeing very strong demand for our AI products in the fourth quarter in 2024. And we exited the year with more demand than we had available capacity. So, we are in a tight supply demand situation, working very hard to bring more capacity online…

…[Question] How do you think about long-term capital intensity for this business?

[Answer] On the first one, certainly, we’re looking ahead, but we’re managing very responsibly. It was a very rigorous, even internal governance process, looking at how do we allocate the capacity and what would we need to support the customer demand externally, but also across the Google — the Alphabet business. And as you’ve seen in the comment I’ve just made on Cloud, we do have demand that exceeds our available capacity. So, we’ll be working hard to address that and make sure we bring more capacity online. We do have the benefit of having a very broad business, and we can repurpose capacity, whether it’s through Google Services or Google Cloud to support, as I said, whether it’s search or GDM, or Google Cloud customers, we can do that in a more efficient manner.

Alphabet’s management thinks Google’s AI models are in the lead when compared to DeepSeek’s, and this is because of Google’s full-stack development

If you look at one of the areas in which the Gemini model shines is the Pareto frontier of cost, performance, and latency. And if you look at all three attributes, I think we are — we lead this period of frontier. And I would say both our 2.0 Flash models, our 2.0 Flash thinking models, they are some of the most efficient models out there, including comparing to DeepSeek’s V3 and R1. And I think a lot of it is our strength of the full stack development, end-to-end optimization, our obsession with cost per query.

Alphabet’s management has seen the proportion of AI spend on inference growing over the last 3 years when compared to training; management thinks reasoning AI models will accelerate this trend

A couple of things I would say are if you look at the trajectory over the past three years, the proportion of the spend toward inference compared to training has been increasing, which is good because, obviously, inferences to support businesses with good ROIC…

…I think the reasoning models, if anything, accelerates that trend because it’s obviously scaling upon inference dimension as well.

Alphabet’s management thinks that AI agents and Google Search are not competing in a zero-sum game

[Question] With your own project Mariner efforts and a competitor’s recent launch, it seems there’s suddenly really strong momentum on AI consumer agents and kind of catching up to that old Google Duplex Vision. I think when you look a few years ahead, where do you see consumer agents going? And really, what does it mean to Google Search outside of Lens? Is there room for both to flourish?

[Answer] Gemini 2.0 was definitely built with the view of enabling more agentic use cases. And so, I actually — we are definitely seeing progress inside. And I think we’ll be able to do more agentic experiences for our users. Look, I actually think all of this expands the opportunity space. I think it — historically, we’ve had information use cases, but now you can have — you can act on your information needs in a much deeper way. It’s always been our vision when we have talked about Google Assistant, etc. So, I think the opportunity space expands. I think there’s plenty of it, feels very far from a zero-sum game. There’s plenty of room, I think, for many new types of use cases to flourish. And I think for us, we have a clear sense of additional use cases we can start to tackle for our users in Google Search.

Alphabet’s management has been passing on cost differentiations arising from Google Cloud’s end-to-end stack approach to customers

Part of the reason we have taken the end-to-end stack approach is so that we can definitely drive a strong differentiation in end-to-end optimizing and not only on a cost but on a latency basis, on a performance basis. Be it the Pareto frontier we mentioned, and I think our full stack approach and our TPU efforts all play give a meaningful advantage. And we plan — you already see that. I know you asked about the cost, but it’s effectively captured when we price outside, we pass on the differentiation. 

Amazon (NASDAQ: AMZN)

AWS grew 19% year-on-year in 2024 Q4, and is now at a US$115 billion annualised revenue run rate; management expects lumpiness in AWS’s growth in the next few years, but is incredibly optimistic about AWS’s growth; management thinks the future will be one where (a) every app is infused with generative AI that has inference as a core building block, and (b) companies will have AI agents accomplishing tasks and interacting with each other; management believes this future will be built on the cloud, and mostly on AWS; the shift by enterprises from on-premises to the cloud, which is a non-AI activity, continues for AWS; AWS continues to innovate in non-AI areas; AWS’s growth in 2024 Q4 was driven by both generative AI and non-generative AI offerings; AWS had a massive 48% year-on-year jump in operating income in 2024 Q4, helped partly by an increase in estimated useful life of servers that started in 2024; management sees AWS being capable of faster growth today if not for supply constraints; the constraints relate to (1) chips from 3rd-party partners (most likely referring to NVIDIA), (2) AWS’s own Trainium chips, (3) power for data centers, and (4) other supply chain components; management sees the AWS constraints starting to relax in 2025 H2; AWS’s AI services come with lower margins right now, but management thinks the AI-related margin will over time be on par with the non-AI margin

In Q4, AWS grew 19% year-over-year and now has a $115 billion annualized revenue run rate. AWS is a reasonably large business by most folks’ standards. And though we expect growth will be lumpy over the next few years as enterprise adoption cycles, capacity considerations and technology advancements impact timing, it’s hard to overstate how optimistic we are about what lies ahead for AWS’ customers and business…

…While it may be hard for some to fathom a world where virtually every app has generative AI infused in it, with inference being a core building block just like compute, storage and database, and most companies having their own agents that accomplish various tasks and interact with one another, this is the world we’re thinking about all the time. And we continue to believe that this world will mostly be built on top of the cloud with the largest portion of it on AWS…

…While AI continues to be a compelling new driver in the business, we haven’t lost our focus on core modernization of companies’ technology infrastructure from on-premises to the cloud. We signed new AWS agreements with companies, including Intuit, PayPal, Norwegian Cruise Line Holdings, Northrop Grumman, The Guardian Life Insurance Company of America, Reddit, Japan Airlines, Baker Hughes, The Hertz Corporation, Redfin, Chime Financial, Asana, and many others. Consistent customer feedback from our recent AWS re:Invent gathering was appreciation that we’re still inventing rapidly in non-AI key infrastructure areas like storage, compute, database and analytics…

…During the fourth quarter, we continued to see growth in both generative AI and non-generative AI offerings as companies turn their attention to newer initiatives, bring more workloads to the cloud, restart or accelerate existing migrations from on-premise to the cloud, and tap into the power of generative AI…

…AWS reported operating income of $10.6 billion, an increase of $3.5 billion year-over-year. This is a result of strong growth, innovation in our software and infrastructure to drive efficiencies, and continued focus on cost control across the business. As we’ve said in the past, we expect AWS operating margins to fluctuate over time driven in part by the level of investments we’re making. Additionally, we increased the estimated useful life of our servers starting in 2024, which contributed approximately 200 basis points to the AWS margin increase year-over-year in Q4…

……It is true that we could be growing faster, if not for some of the constraints on capacity. And they come in the form of, I would say, chips from our third-party partners, come a little bit slower than before with a lot of midstream changes that take a little bit of time to get the hardware actually yielding the percentage-healthy and high-quality servers we expect. It comes with our own big new launch of our own hardware and our own chips and Trainium2, which we just went to general availability at re:Invent, but the majority of the volume is coming in really over the next couple of quarters, the next few months. It comes in the form of power constraints where I think the world is still constrained on power from where I think we all believe we could serve customers if we were unconstrained. There are some components in the supply chain, like motherboards, too, that are a little bit short in supply for various types of servers…

…I predict those constraints really start to relax in the second half of ’25…

…At the stage we’re in right now, AI is still early stage. It does come originally with lower margins and a heavy investment load as we’ve talked about. And in the short term, over time, that should be a headwind on margins. But over the long term, we feel the margins will be comparable in non-AI business as well.

Amazon’s management sees NVIDIA being an important partner of AWS for a long time; management does not see many large-scale generative AI apps existing right now; when generative AI apps reach scale, their costs to operate can rise very quickly, and management believes this will drive customers to demand better price performance from chips, which is why AWS built its custom AI chips; Trainium 2, AWS’s custom AI chip, was launched in December 2024; EC2 instances powered by Trainium 2 is 30%-40% more price performant than instances powered by other GPUs; important technology companies such as Adobe, Databricks, and Qualcomm have seen impressive results after testing Trainium 2; Anthropic is building its future frontier models on Trainium 2; AWS is collaborating with Anthropic on Project Rainier, which is a cluster of a few hundred thousand Trainium 2 chips that have 5x the exaflops Anthropic used to train its current set of models; management is already 

Most AI compute has been driven by NVIDIA chips, and we obviously have a deep partnership with NVIDIA and will for as long as we can see into the future. However, there aren’t that many generative AI applications of large scale yet. And when you get there, as we have with apps like Alexa and Rufus, cost can get steep quickly. Customers want better price performance and it’s why we built our own custom AI silicon. Trainium2 just launched at our AWS re:Invent Conference in December. And EC2 instances with these chips are typically 30% to 40% more price performant than other current GPU-powered instances available. That’s very compelling at scale. Several technically-capable companies like Adobe, Databricks, Poolside and Qualcomm have seen impressive results in early testing of Trainium2. It’s also why you’re seeing Anthropic build their future frontier models on Trainium2. We’re collaborating with Anthropic to build Project Rainier, a cluster of Trainium2 UltraServers containing hundreds of thousands of Trainium2 chips. This cluster is going to be 5x the number of exaflops as the cluster that Anthropic used to train their current leading set of cloud models. We’re already hard at work on Trainium3, which we expect to preview late in ’25 and defining Trainium4 thereafter.

Building outstanding performant chips that deliver leading price performance has become a core strength of AWS’, starting with our Nitro and Graviton chips in our core business and now extending to Trainium and AI and something unique to AWS relative to other competing cloud providers.

Amazon’s management has seen Amazon SageMaker AI, AWS’s fully-managed AI service, become the go-to service for AI model builders; SageMaker’s HyperPod automatically splits training workloads across many AI accelerators and prevents interruptions, saving training time up tp 40%; management recently released new features for SageMaker, such as the ability to prioritise which workloads to receive capacity when budgets are reached; the latest version of SageMaker is able to integrate all of AWS’s data analytics and AI services into one surface

I won’t spend a lot of time in these comments on Amazon SageMaker AI, which has become the go-to service for AI model builders to manage their AI data, build models, experiment and deploy these models, except to say that SageMaker’s HyperPod capability, which automatically splits training workloads across many AI accelerators, prevents interruptions by periodically saving checkpoints, and automatically repairing faulty instances from their last saved checkpoint and saving training time by up to 40%. It continues to be a differentiator, received several new compelling capabilities at re:Invent, including the ability to manage costs at a cluster level and prioritize which workloads should receive capacity when budgets are reached, and is increasingly being adopted by model builders…

…There were several key launches customers were abuzz about, including Amazon Aurora DSQL, our new serverless distributed SQL database that enables applications with the highest availability, strong consistency, PostgreS compatibility and 4x faster reads and writes compared to other popular distributed SQL databases; Amazon S3 tables, which make S3 the first cloud object store with fully managed support for Apache Iceberg for faster analytics; Amazon S3 Metadata, which automatically generates queryable metadata, simplifying data discovery, business analytics, and real-time inference to help customers unlock the value of their data in S3; and the next generation of Amazon SageMaker, which brings together all of the data analytics services and AI services into one interface to do analytics and AI more easily at scale.

Amazon Bedrock is AWS’s fully-managed service for developers to build generative AI applications by leverage on frontier models; management recently introduced more than 100 popular emerging models on Bedrock, including DeepSeek’s R1 models; management recently introduced new features to Bedrock to help customers lower cost and latency in inference workloads; management is seeing Bedrock resonate strongly with customers; management recently released Amazon’s own Nova family of frontier models on Bedrock; customers are starting to experiment with DeepSeek’s models

Amazon Bedrock is our fully managed service that offers the broadest choice of high-performing foundation models with the most compelling set of features that make it easy to build a high-quality generative AI application. We continue to iterate quickly on Bedrock announcing Luma AI poolside and over 100 other popular emerging models to Bedrock at re:Invent. In short order, we also just added DeepSeek’s R1 models to Bedrock and SageMaker…

…We delivered several compelling new Bedrock features at re:Invent, including prompt caching, intelligent prompt routing and model distillation, all of which help customers achieve lower cost and latency in their inference. Like SageMaker AI, Bedrock is growing quickly and resonating strongly with customers…

…We also just launched Amazon’s own family of frontier models in Bedrock called Nova…

…We moved so quickly to make sure that DeepSeek was available both in Bedrock and in SageMaker faster than you saw from others. And we already have customers starting to experiment with that.

The Nova family has comparable intelligence with other leading AI models, but also offers lower latency and price, and integration with important Bedrock features; many large enterprises, including Palantir, Deloitte, and SAP, are already using Nova

We also just launched Amazon’s own family of frontier models in Bedrock called Nova. These models compare favorably in intelligence against the leading models in the world but offer lower latency; lower price, about 75% lower than other models in Bedrock; and are integrated with key Bedrock features like fine-tuning, model distillation, knowledge bases of RAG and agentic capabilities. Thousands of AWS customers are already taking advantage of the capabilities and price performance of Amazon Nova models, including Palantir, Deloitte, SAP, Dentsu, Fortinet, Trellix, and Robinhood, and we’ve just gotten started.

Amazon’s management still thinks Amazon Q is the most capable AI-powered software development assistant; early testing shows that Amazon Q can now shorten a multi-year mainframe migration by 50%

Amazon Q is the most capable generative AI-powered assistant for software development and to leverage your own data…

…We obliged with our recent deliveries of Q Transformations that enable moves from Windows.NET applications to Linux, VMware to EC2, and accelerates mainframe migrations. Early customer testing indicates that Q can turn what was going to be a multiyear effort to do a mainframe migration into a multi-quarter effort, cutting by more than 50% the time to migrate mainframes. This is a big deal and these transformations are good examples of practical AI.

Amazon’s management expects capital expenditures of around US$105 billion for the whole of 2025 (was around $75 billion in 2024); the capex in 2025 will be for AWS as well as the retail business, but will primarily be for AWS’s AI infrastructure; reminder that the faster AWS grows, the faster Amazon needs to invest capital for hardware; management will only spend on capex if they see significant signals of demand; management thinks AI is a once-in-a-lifetime business opportunity, and that it’s a good sign on the long-term growth opportunities AWS has when capex is expanding

Capital investments were $26.3 billion in the fourth quarter, and we think that run rate will be reasonably representative of our 2025 capital investment rate. Similar to 2024, the majority of the spend will be to support the growing need for technology infrastructure. This primarily relates to AWS, including to support demand for our AI services, as well as tech infrastructure to support our North America and International segments. Additionally, we’re continuing to invest in capacity for our fulfillment and transportation network to support future growth. We’re also investing in same-day delivery facilities and our inbound network as well as robotics and automation to improve delivery speeds and to lower our cost to serve. These capital investments will support growth for many years to come…

…The vast majority of that CapEx spend is on AI for AWS. The way that AWS business works and the way the cash cycle works is that the faster we grow, the more CapEx we end up spending because we have to procure data center and hardware and chips and networking gear ahead of when we’re able to monetize it. We don’t procure it unless we see significant signals of demand. And so when AWS is expanding its CapEx, particularly in what we think is one of these once-in-a-lifetime type of business opportunities like AI represents, I think it’s actually quite a good sign, medium to long term, for the AWS business…

…We also have CapEx that we’re spending this year in our Stores business, really with an aim towards trying to continue to improve the delivery speed and our cost to serve. And so you’ll see us expanding the number of same-day facilities from where we are right now. You’ll also see us expand the number of delivery stations that we have in rural areas so we can get items to people who live in rural areas much more quickly, and then a pretty significant investment as well on robotics and automation so we can take our cost to serve down and continue to improve our productivity.

Amazon’s management completed a useful life study for its servers and network equipment in 2024 Q4 and has decreased the useful life estimate; management early retired some servers and network equipment in 2024 Q4; the decrease in useful life estimate and the early retirement will lower Amazon’s operating income, primarily in the AWS segment

In Q4, we completed a useful life study for our servers and network equipment, and observed an increased pace of technology development, particularly in the area of artificial intelligence and machine learning. As a result, we’re decreasing the useful life for a subset of our servers and network equipment from 6 years to 5 years, beginning in January 2025. We anticipate this will decrease full year 2025 operating income by approximately $700 million. In addition, we also early retired a subset of our servers and network equipment. We recorded a Q4 2024 expense of approximately $920 million from accelerated depreciation and related charges and expect this will also decrease full year 2025 operating income by approximately $600 million. Both of these server and network equipment useful life changes primarily impact our AWS segment.

Amazon’s management sees AI as the biggest opportunity since cloud and the internet

From our perspective, we think virtually every application that we know of today is going to be reinvented with AI inside of it and with inference being a core building block, just like compute and storage and database. If you believe that, plus altogether new experiences that we’ve only dreamed about are going to actually be available to us with AI, AI represents, for sure, the biggest opportunity since cloud and probably the biggest technology shift and opportunity in business since the Internet.

Amazon’s management has been impressed with DeepSeek’s innovations

I think like many others, we were impressed with what DeepSeek has done, I think in part impressed with some of the training techniques, primarily in flipping the sequencing of reinforcement learning being earlier and without the human-in-the-loop. We thought that was interesting ahead of the supervised fine-tuning. We also thought some of the inference optimizations they did were also quite interesting

Amazon’s management’s core belief remains that generative AI apps will use multiple models and different customers will use different AI models for different workloads 

You have a core belief like we do that virtually all the big generative AI apps are going to use multiple model types, and different customers are going to use different models for different types of workloads.

Amazon’s management thinks that the cheaper AI inference becomes, the more inference spending there will be; management believes that the cost of AI inference will fall substantially over time

Sometimes people make the assumptions that if you’re able to decrease the cost of any type of technology component, in this case, we’re really talking about inference, that somehow it’s going to lead to less total spend in technology. And we have never seen that to be the case. We did the same thing in the cloud where we launched AWS in 2006, where we offered S3 object storage for $0.15 a gigabyte and compute for $0.10 an hour, which, of course, is much lower now many years later, people thought that people would spend a lot less money on infrastructure technology. And what happens is companies will spend a lot less per unit of infrastructure, and that is very, very useful for their businesses, but then they get excited about what else they could build that they always thought was cost prohibitive before, and they usually end up spending a lot more in total on technology once you make the per unit cost less. And I think that is very much what’s going to happen here in AI, which is the cost of inference will substantially come down. What you heard in the last couple of weeks, DeepSeek is a piece of it, but everybody is working on this. I believe the cost of inference will meaningfully come down. I think it will make it much easier for companies to be able to infuse all their applications with inference and with generative AI.

Amazon’s management currently sees 2 main ways that companies are getting value out of AI; the 1st way is through productivity and cost savings, and it is the lowest-hanging fruit; the 2nd way is by building new experiences

There’s kind of two macro buckets of how we see people, both ourselves inside Amazon as well as other companies using AWS, how we see them getting value out of AI today. The first macro bucket, I would say, is really around productivity and cost savings. And in many ways, this is the lowest-hanging fruit in AI…

….I’d say the other big macro bucket are really altogether new experiences.

Amazon has built a chatbot with generative AI and it has lifted customer satisfaction by 500 basis points; Amazon has built a generative AI application for 3rd-party sellers to easily fill up their product detail pages; Amazon has built generative AI applications for inventory management that improve inventory forecasting by 10% and regional predictions by 20%; the brains of Amazon’s robotics are infused with generative AI

If you look at customer service and you look at the chatbot that we’ve built, we completely rearchitected it with generative AI. It’s delivering. It already had pretty high satisfaction. It’s delivering 500 basis points better satisfaction from customers with the new generative AI-infused chatbot.

If you look at our millions of third-party selling partners, one of their biggest pain points is, because we put a high premium on really organizing our marketplace so that it’s easy to find things, there’s a bunch of different fields you have to fill out when you’re creating a new product detail page, but we’ve built a generative AI application for them where they can either fill in just a couple of lines of text or take a picture of an image or point to a URL, and the generative AI app will fill in most of the rest of the information they have to fill out, which speeds up getting selection on the website and easier for sellers.

If you look at how we do inventory management and trying to understand what inventory we need, at what facility, at what time, the generative AI applications we’ve built there have led to 10% better forecasting on our part and 20% better regional predictions.

In our robotics, we were just talking about the brains in a lot of those robotics are generative AI-infused that do things like tell the robotic claw what’s in a bin, what it should pick up, how it should move it, where it should place it in the other bin that it’s sitting next to. So it’s really in the brains of most of our robotics.

Amazon’s Rufus is an AI-infused shopping assistant that is growing significantly; users can take a picture of a product with Amazon Lens and have the service surface the exact item through the use of AI; Amazon is using AI to know the relative sizing of clothes and shoes from different brands so that it can recommend the right sizes to shoppers; Amazon is using AI to improve the viewing experience of sporting events; Rufus provides a significant improvement to the shopping experience for shoppers and management expects the usage of Rufus to increase throughout 2025

You see lots of those in our retail business, ranging from Rufus, which is our AI-infused shopping assistant, which continues to grow very significantly; to things like Amazon Lens, where you can take a picture of a product that’s in front of you, you check it out in the app, you can find it in the little box at the top, you take a picture of an item in front of you, and it uses computer vision and generative AI to pull up the exact item in search result; to things like sizing, where we basically have taken the catalogs of all these different clothing manufacturers and then compare them against one another so we know which brands tend to run big or small relative to each other. So when you come to buy a pair of shoes, for instance, it can recommend what size you need; to even what we’re doing in Thursday Night Football, where we’re using generative AI for really inventive features like it sends alerts where we predict which players are going to put quarterback or defensive vulnerabilities, where we were able to show viewers what area of the field is vulnerable…

…I do think that Rufus, if you look at how it impacts the customer experience and if you actually use it month-to-month, it continues to get better and better. If you’re buying something and you’re on our product detail page, our product detail pages provide so much information that sometimes it’s hard, if you’re trying to find something quickly, to scroll through and find that little piece of information. And so we have so many customers now who just use Rufus to help them find a quick fact about a product. They also use Rufus to figure out how to summarize customer reviews so they don’t have to read 100 customer reviews to get a sense of what people think about that product. If you look at the personalization, really, most prominently today, your ability to go into Rufus and ask what’s happened to an order or what did I just order or can you pull up for me this item that I ordered 2 months ago, the personalization keeps getting much better. And so we expect throughout 2025, that the number of occasions where you’re not sure what you want to buy and you want help from Rufus are going to continue to increase and be more and more helpful to customers.

Amazon has around 1,000 generative AI applications that it has built or is building

We’ve got about 1,000 different generative AI applications we’ve either built or in the process of building right now.

Apple (NASDAQ: AAPL)

Apple Intelligence was first released in the USA in October 2024, with more features and countries introduced in December 2024; Apple Intelligence will be rolled out to even more countries in April 2025; management sees Apple Intelligence as a breakthrough for privacy in AI; SAP is using Apple Intelligence in the USA to improve the employee as well as customer experience; the Apple Intelligence features that people are using include Writing Tools, Image Playground, Genmoji, Visual Intelligence, Clean Up, and more; management has found Apple Intelligence’s email summarisation feature to be very useful; management thinks that different users will find their own “killer feature” within Apple Intelligence

In October, we released the first set of Apple Intelligence features in U.S. English for iPhone, iPad and Mac, and we rolled out more features and expanded to more countries in December.

Now users can discover the benefits of these new features in the things they do every day. They can use Writing Tools to help find just the right words, create fun and unique images with Image Playground and Genmoji, handle daily tasks and seek out information with a more natural and conversational Siri, create movies of their memories with a simple prompt and touch up their photos with Clean Up. We introduced visual intelligence with Camera Control to help users instantly learn about their surroundings. Users can also seamlessly access ChatGPT across iOS, iPadOS and macOS.

And we were excited to recently begin our international expansion with Apple Intelligence now available in Australia, Canada, New Zealand, South Africa and the U.K. We’re working hard to take Apple Intelligence even further. In April, we’re bringing Apple Intelligence to more languages, including French, German, Italian, Portuguese, Spanish, Japanese, Korean and simplified Chinese as well as localized English to Singapore and India. And we’ll continue to roll out more features in the future, including an even more capable Siri.

Apple Intelligence builds on years of innovations we’ve made across hardware and software to transform how users experience our products. Apple Intelligence also empowers users by delivering personal context that’s relevant to them. And importantly, Apple Intelligence is a breakthrough for privacy in AI with innovations like Private Cloud Compute, which extends the industry-leading security and privacy of Apple devices into the cloud…

…We’re excited to see leading enterprises such as SAP leverage Apple Intelligence in the U.S. with features like Writing Tools, summarize and priority notifications to enhance both their employee and customer experiences…

…In terms of the features that people are using, they’re using all of the ones that I had referenced in my opening comments, from Writing Tools to Image Playground and Genmoji, to visual intelligence and more. And so we see all of those being used. Clean Up is another one that is popular, and people love seeing that one demoed in the stores as well…

…I know from my own personal experience, once you start using the features, you can’t imagine not using them anymore. I know I get hundreds of e-mails a day, and the summarization function is so important…

…[Question] Do you guys see the upgraded Siri expected in April as something that will, let’s say, be the killer application among the suite of features that you have announced in Apple Intelligence?

[Answer] I think the killer feature is different for different people. But I think for most, they’re going to find that they’re going to use many of the features every day. And certainly, one of those is the — is Siri, and that will be coming over the next several months.

Many customers are excited about the iPhone 16 because of Apple Intelligence; the iPhone 16’s year-on-year performance was stronger in countries where Apple Intelligence was available compared to countries where Apple Intelligence was not available

Our iPhone 16 lineup takes the smartphone experience to the next level in so many ways, and Apple Intelligence is one of many reasons why customers are excited…

…We did see that the markets where we had rolled out Apple Intelligence, that the year-over-year performance on the iPhone 16 family was stronger than those where Apple Intelligence was not available…

Apple’s management thinks the developments in the AI industry brought on by DeepSeek’s emergence is a positive for Apple

[Question] There’s a perception that you’re a big beneficiary of lower cost of compute. And I was wondering if you could give your worldly perspective here on the DeepSeek situation.

[Answer] In general, I think innovation that drives efficiency is a good thing. And that’s what you see in that model. Our tight integration of silicon and software, I think, will continue to serve us very well.

Arista Networks (NYSE: ANET)

Cloud and AI titans were a significant contributor to Arista Networks’ revenue in 2024; management considers Oracle an AI titan too

Now shifting to annual sector revenue for 2024. Our cloud and AI titans contributed significantly at approximately 48%, keeping in mind that Oracle is a new member of this category.

Arista Networks’ core cloud AI and data center products are built off its extensible OS (operating system) and goes up to 800 gigabit Ethernet speeds

Our core cloud AI and data center products are built off a highly differentiated, extensible OS stack and is successfully deployed across 10, 25, 100, 200, 400 and 800 gigabit Ethernet speeds. It delivers power efficiency, high availability, automation and agility as the data centers demand, insatiable bandwidth capacity and network speeds for both front-end and back-end storage, compute and AI zones.

Arista Networks’ management expects the company’s 800 gigabit Ethernet switch to emerge as an AI back-end cluster in 2025

We expect 800 gigabit Ethernet to emerge as an AI back-end cluster in 2025.

Arista Networks’ management is still optimistic that AI revenues will reach $1.5 billion in 2025, including $750 million in AI back-end clusters; the $750 million in revenue from AI back-end clusters will have a major helping hand from 3 of the 5 major AI trials Arista Networks is working on that are rolling out a cumulative 100,000 GPUs in 2025 (see more below)

We remain optimistic about achieving our AI revenue goal of $1.5 billion in AI centers, which includes the $750 million in AI back-end clusters in 2025…

…[Question] You are reiterating $750 million AI back-end sales this year despite the stalled or the fifth customer. Can you talk about where is the upside coming from this year? Is it broad-based or 1 or 2 customers?

[Answer] We’re well on our way and 3 customers deploying a cumulative of 100,000 GPUs is going to help us with that number this year. And as we increased our guidance to $8.2 billion, I think we’re going to see momentum both in AI, cloud and enterprises. I’m not ready to break it down and tell you which where. I think we’ll see — we’ll know that much better in the second half. But Chantelle and I feel confident that we can definitely do the $8.2 billion that we historically don’t call out so early in the year. So having visibility if that helps.

Arista Networks is building some of the world’s greatest Arista AI centers at production scale and it’s involved with both the back-end clusters and front-end networks; Arista Networks’ management sees the data traffic flow of AI workloads as having significant differences from traditional cloud workloads and Arista AI centers can seamlessly connect to the front end compute storage with its backend Ethernet portfolio; Arista’s AI networking portfolio consists of 3 families and over 20 Etherlink switches

Networking for AI is also gaining traction as we move into 2025, building some of the world’s greatest Arista AI centers at production scale. These are constructed with both back-end clusters and front-end networks…

…The fidelity of the AI traffic differs greatly from cloud workloads in terms of diversity, duration and size of flow. Just one slow flow can flow the entire job completion time for a training workload. Therefore, Arista AI centers seamlessly connect to the front end of compute storage WAN and classic cloud networks with our back-end Arista Etherlink portfolio. This AI accelerated networking portfolio consists of 3 families and over 20 Etherlink switches, not just 1 point switch.

Arista Networks’ management’s AI for Networking strategy is doing well and it includes software that have superior AI ops

Our AI for networking strategy is also doing well, and it’s about curating the data for higher-level network functions. We instrument our customer’s networks with our published subscribed state Foundation with our software called Network Data Lake to deliver proactive, predictive and prescriptive platforms that have superior AI ops with A care support and product functions.

Arista Networks’ management is still committed to 4 of the 5 major AI trials that they have been discussing in recent earnings calls; the remaining AI trial is still stalled and the customer is not a Cloud Titan and is waiting for funding; 3 of the 4 trials that are active are expected to roll out a cumulative 100,000 GPUs in 2025 and they are all waiting for the next-generation NVIDIA GPU; Arista Networks’ management expects to do very well on the back-end with those 3 trials; the remaining trial of the 4 active trials is migrating from Infiniband to Ethernet to test the viability of Ethernet, and Arista Networks’ management expects to enter production in 2026

I want to say Arista is still committed to 4 out of our 5 AI clusters that I mentioned in prior calls, but just one is a little bit stalled. It is not a Cloud Titan. They are awaiting GPUs and some funding too, I think. So I hope they’ll come back next year, but for this year, we won’t talk about them. But the remaining 4, let me spend some — jgive you some color, 3 out of the 4 customers are expected to this year rolled out a cumulative of 100,000 GPUs. So we’re going to do very well with 3 of them on the back end. And you can imagine, they’re all pretty much one major NVIDIA class of GPU — it’s — they will be waiting for the next generation of GPUs. But independent of that, we’ll be rolling out fairly large numbers. On the fourth one, we are migrating right now from InfiniBand to proving that Ethernet is a viable solution, so we’re still — they’ve historically been InfiniBand. And so we’re still in pilot and we expect to go into production next year. We’re doing very well in 4 out of 4, the Fifth one installed and 3 out of the 4 expected to be 100,000 GPUs this year.

Arista Networks thinks the market for AI networking is large enough that there will be enough room for both the company and other whitebox networking manufacturers; management also thinks Arista Networks’ products have significant differentiation from whitebox products, especially in the AI spine in a typical leaf-spine network, because Arista Networks’ products can automatically provide an alternate connection when a GPU in the network is in trouble

[Question] Can you maybe share your perspective that when it comes to AI network especially the back-end networks, how do you see the mix evolving white box versus OEM solution?

[Answer] This TAM is so huge and so large. We will always coexist with white boxes and operating systems that are non-EOS, much like Apple coexists on the iPhone with other phones of different types. When you look at the back end of an AI cluster, there are typically 2 components, the AI lead and the AI spine. The AI lead connects to the GPUs and therefore, is the first, if you will, point of connection. And the AI spine aggregates all of these AI leads. Almost in all the back-end examples we’ve seen, the AI spine is generally 100% Arista-branded EOS. You’ve got to do an awful lot of routing, scale, features, capabilities that are very rich that would be difficult to do in any other environment. The AI leads can vary. So for example, the — let’s take the example of the 5 customers I mentioned a lot, 3 out of the 5 are all EOS in the [indiscernible] spine. 2 out of the 5 are kind of hybrids. Some of them have some form of SONic or FBOSS. And as you know, we co-develop with them and coexist in a number of use cases where it’s a real hybrid combination of EOS and an open OS. So for most part, I’d just like to say that white box and Arista will coexist and will provide different strokes for different folks…

…A lot of our deployments right now is 400 and 800 gig, and you see a tremendous amount of differentiation, not only like I explained to you in scale and routing features, but cost and load balancing, AI visibility and analytics at real time, personal queuing, congestion control, visibility and most importantly, smart system upgrade because you sort of want your GPUs to come down because you don’t have the right software to accelerate so that the network provides the ideal foundation that if the GPU is in trouble, we can automatically give a different connection and an alternate connection. So tremendous amount of differentiation there and even more valid in a GPU which costs typically 5x as much as a CPU…

…When you’re buying these expensive GPUs that cost $25,000, they’re like diamonds, right? You’re not going to string a diamond on a piece of thread. So first thing I want to say is you need a mission-critical network, whether you want to call it white box, blue box, EOS or some other software, you’ve got to have mission-critical functions, analytics, visibility, high availability, et cetera. As I mentioned, and I want to reiterate, they’re also typically a leaf spine network. And I have yet to see an AI spine deployment that is not EOS-based. I’m not saying it can’t happen or won’t happen. But in all 5 major installations, the benefit of our EOS features for high availability for routing, for VXLAN, for telemetry, our customers really see that. And the 7800 is the flagship AI spine product that we have been deploying last year, this year and in the future. Coming soon, of course, is also the product we jointly engineered with Meta, which is the distributed [Ecolink] switch. And that is also an example of a product that provides that kind of leaf spine combination, both with FBOSS and EOS options in it. So in my view, it’s difficult to imagine a highly resilient system without Arista EOS in AI or non-AI use cases.

On the leaf, you can cut corners. You can go with smaller buffers, you may have a smaller installation. So I can imagine that some people will want to experiment and do experiment in smaller configurations with non-EOS. But again, to do that, you have to have a fairly large staff to build the operations for it. So that’s also a critical element. So unless you’re a large Cloud Titan customer, you’re less likely to take that chance because you don’t have the staff.

Arista Networks’ management is seeing strong demand from its Cloud Titan customers

Speaking specifically to Meta, we are obviously in a number of use cases in Meta. Keep in mind that our 2024 Meta numbers is influenced by more of their 2023 CapEx, and that was Meta’s year of efficiency where their CapEx was down 15% to 20%. So you’re probably seeing some correlation between their CapEx being down and our revenue numbers being slightly lower in ’24. In general, I would just say all our cloud titans are performing well in demand, and we shouldn’t confuse that with timing of our shipments. And I fully expect Microsoft and Meta to be greater than 10% customers in a strong manner in 2025 as well. Specific to the others we added in, they’re not 10% customers, but they’re doing very well, and we’re happy with their cloud and AI use cases.

Arista Networks’ management thinks the emergence of DeepSeek will lead to AI development evolving from back-end training that’s concentrated in a handful of users, to being distributed more widely across CPUs and GPUs; management also thinks DeepSeek’s emergence is a positive for Arista Networks because DeepSeek’s innovations can drive the AI industry towards a new class of CPUs, GPUs, AI accelerators and Arista Networks is able to scale up network for all kinds of XPUs

DeepSeek certainly deep fixed many stocks, but I actually see this as a positive because I think you’re now going to see a new class of CPUs, GPUs, AI accelerators and where you can have substantial efficiency gains that go beyond training. So that could be some sort of inference or mixture of experts or reasoning and which lowers the token count and therefore, the cost. So what I like about all these different options is Arista can scale up network for all kinds of XPUs and accelerators. And I think the eye-opening thing here for all of our experts who are building all these engineering models is there are many different types and training isn’t the only one. So I think this is a nice evolution of how AI will not just be a back-end training only limited to 5 customers type phenomenon, but will become more and more distributed across a range of CPUs and GPUs.

Arista Networks’ management thinks hyper-scale GPU clusters, such as Project Stargate, will drive the development of vertical rack integration in the next few years and Andy Bechtolsheim, an Arista Networks co-founder, is personally involved in these projects

If you look at how we have classically approached GPUs and connected libraries, we’ve largely looked at it as 2 separate building blocks. There’s the vendor who provides the GPUs and then there’s us who provides the scale-out networking. But when you look at Stargate and projects like this, I think you’ll start to see more of a vertical rack integration where the processor, the scale up, the scale out and all of the software to provide a single point of control and visibility starts to come more and more together. This is not a 2025 phenomenon, but definitely in ’26 and ’27, you’re going to see a new class of AI accelerators for — and a new class of training and inference, which is extremely different than the current more pluggable label type of version. So we’re very optimistic about it.

Andy Bechtolsheim is personally involved in the design of a number of these next-generation projects and the need for this type of shall we say, pushing Moore’s Law of improvements in density of performance that we saw in the 2000s is coming back, and you can boost more and more performance per XPU, which means you have to boost the network scale from 800 gig to 1.16.

Arista Networks’ management sees a $70 billion total addressable market in 2028, of which roughly a third is related to AI

[Question] If you can talk to the $70 billion TAM number for 2028, how much is AI?

[Answer] On the $70 billion TAM in 2028, I would roughly say 1/3 is AI, 1/3 is data center and cloud and 1/3 is campus and enterprise. And obviously, absorbed into that is routing and security and observability. I’m not calling them out separately for the purpose of this discussion.

Arista Networks’ management sees co-packaged optics (CPO) as having weak adoption compared to co-packaged copper (CPC) because CPO has been experiencing field failures

Co-packaged optics is not a new idea. It’s been around 10 to 20 years. So the fundamental reason, let’s go through why co-packaged optics has had a relatively weak adoption so far is because of field failures and most of it is still in proof of concept today. So going back to networking, the most important attribute of a network switch is reliability and troubleshooting. And once you solder a co-packaged optics on a PCB, you lose some of that flexibility and you don’t get the serviceability and manufacturing. That’s been the problem. Now a number of alternatives are emerging, and we’re a big fan of co-packaged copper as well as pluggable optics that can complement this like linear drive or LTO as we call it.

Now we also see that if co-packaged optics improves some of the metrics it has right now. For example, it has a higher channel count than the industry standard of 8-channel pluggable optics, but we can do higher channel pluggable optics as well. So some of these things improve, we can see that both CPC and CPO will be important technologies at 224 gig or even 448 gig. But so far, our customers have preferred a LEGO approach that they can mix and match pluggable switches and pluggable optics and haven’t committed to soldering them on the PCB. And we feel that will change only if CPO gets better and more reliable. And I think CPC can be a nice alternative to that.

Arista Networks’ management is seeing customers start moving towards actual use-cases for AI, but the customers are saying that these AI projects take time to implement

For the AI perspective, speaking with the customers, it’s great to move from kind of a theory to more specific conversation, and you’re seeing that in the banks and some of the higher tier Global 2000, Fortune 500 companies. And so they’re moving from theory to actual use cases they’re speaking to. And the way they describe it is it takes a bit of time. They’re working mostly with cloud service providers at the beginning, kind of doing some training and then they’re deciding whether they bring that on-prem and inference. So they’re making those decisions.

Arista Networks’ management is seeing a new class of Tier 2 specialty AI cloud providers emerge

We are seeing a new class of Tier 2 specialty cloud providers emerge that want to provide AI as a service and want to be differentiated there. And there’s a whole lot of funding, grant money, real money going in there. So service providers, too early to call. But Neo clouds and specialty providers, yes, we’re seeing lots of examples of that.

The advent of AI has accelerated the speed-transitions in networking data switches, but there’s still going to be a long runway for Arista Networks’ 400 gig and 800 gig products, with 1.6 tera products being deployed in a measured way

The speed transitions because of AI are certainly getting faster. It used to take when we went from 200 gig, for example, at Meta or 100 gig in some of our Cloud Titans to 400, that speed transition typically took 3 to 4, maybe even 5 years, right? In AI, we see that cycle being almost every 2 years…

…2024 was the year of real 400 gig. ’25 and ’26, I would say, is more 800 gig. And I really see 1.6T coming into the picture because we don’t have chips yet, maybe in what do you say, John, late ’26 and real production maybe in ’27. So there’s a lot of talk and hype on it, just like I remember talk and hype on 400 gig 5 years ago. But I think realistically, you’re going to see a long runway for 400 and 800 gig. Now as we get into 1.6T, part of the reason I think it’s going to be measured and thoughtful is many of our customers are still awaiting their own AI accelerators or NVIDIA GPUs, which with liquid cooling that would actually push that kind of bandwidth. So new GPUs will require new bandwidth, and that’s going to push it out a year or 2.

Arista Networks’ management sees a future where the market share between NVIDIA GPUs and custom AI accelerators (ASICs) is roughly evenly-split, but Arista Networks’ products will be GPU-agnostic

[Question] There’s been a lot of discussion over the last few months between the general purpose GPU clusters from NVIDIA and then the custom ASIC solutions from some of your popular customers. I guess just in your view, over the longer term, does Arista’s opportunity differ across these 2 chip types?

[Answer] I think I’ve always said this, you guys often spoke about NVIDIA as a competitor. And I don’t see it that way. I see that — thank you, NVIDIA. Thank you, Jensen, for the GPUs because that gives us an opportunity to connect to them, and that’s been a predominant market for us. As we move forward, we see not only that we connect to them, but we can connect to AMD GPUs and built in in-house AI accelerators. So a lot of them are in active development or in early stages. NVIDIA is the dominant market share holder with probably 80%, 90%. But if you ask me to guess what it would look like 2, 3 years from now, I think it could be 50-50. So Arista could be the scale-out network for all types of accelerators. We’ll be GPU agnostic. And I think there’ll be less opportunity to bundle by specific vendors and more opportunity for customers to choose best-of-breed. 

ASML (NASDAQ: ASML)

AI will be the biggest driver of ASML’s growth and management sees customers benefiting very strongly from it; management thinks ASML will hit the upper end of the revenue guidance range for 2025 if its customers can bring on additional AI-related capacity during the year, but there are also risks that could result in only the lower end of the guidance coming true

We see total revenue for 2025 between €30 billion and €35 billion and the gross margin between 51% and 53%. AI is the clear driver. I think we started to see that last year. In fact, at this point, we really believe that AI is creating a shift in the market and we have seen customers benefiting from it very strongly…

…If AI demand continues to be strong and customers are successful in bringing on additional capacity online to support that demand, there is potential opportunity towards the upper end of our range. On the other hand, there are also risks related to customers and geopolitics that could drive results towards the lower end of the range.

ASML’s management is still very positive on the long-term outlook for ASML, with AI being a driver for growth; management expects AI to create a shift in ASML’s end-market products towards more HPC (high performance computing) and HBM (high bandwidth memory), which requires more advanced logic and DRAM, which in turn needs more critical lithography exposures

I think our view on the long term is still, I would say, very positive…

…Looking longer term, overall the semiconductor market remains strong with artificial intelligence creating growth but also a shift in market dynamics as I highlighted earlier. These dynamics will lead to a shift in the mix of end market products towards more HPC and HBM which requires more advanced logic and DRAM. For ASML, we anticipate that an increased number of critical lithography exposures for these advanced Logic and Memory processes will drive increasing demand for ASML products and services. As a result, we see a 2030 revenue opportunity between 44 billion euros and 60 billion euros with gross margins expected between 56 percent and 60 percent, as we presented in Investor Day 2024.

ASML’s management is seeing aggressive capacity addition among some DRAM memory customers because of demand for high bandwidth memory (HBM), but apart from HBM, other DRAM memory customers have a slower recovery

 I think that high-bandwidth memory is driving today, I would say, also an aggressive capacity addition, at least for some of the customer. I think on the normal DRAM, I would say, my comment is similar to the one on mobile [ photology ] before. I think there are also nothing spectacular, but there is some recovery, which also called for more capacity. So that’s why we still see DRAM pretty strong in 2025.

Datadog (NASDAQ: DDOG)

Datadog launched LLM Observability in 2024; management continues to see increased interest in next-gen AI; 3,500 Datadog customers at the end of 2024 Q4 used 1 or more Datadog AI integrations; when it comes to AI inference, management is seeing most customers using a 3rd-party AI model through an API or a 3rd-party inference platform, and these customers want to observe whether the model is doing the right thing, and this need is what LLM Observability is serving; management is seeing very few customers running the full AI inference stack currently, but they think this could happen soon and it would be an exciting development

We launched LLM observability, in general availability to help customers evaluate, safely deploy and manage their models in production, and we continue to see increased interest in next-gen AI. At the end of Q4, about 3,500 customers use 1 or more Datadog AI integrations to send this data about their machine learning, AI, and LLM usage…

…On the inference side, the — mostly still what customers do is they use a third-party model either through an API or through a third-party inference platform. And what they’re interested in is measuring whether that model is doing the right thing. And that’s what we serve right now with LLM observability, for example, as well, we see quite a bit of adoption that does not come largely from the AI-native companies. So that’s what we see today.

In terms of operating the inference stack fully and how we see relatively few customers with that yet, we think that’s something that’s going to come next. And by the way, we’re very excited by the developments we see in the space. So it looks like there is many, many different options that are going to be viable for running your AI inference. There’s a very healthy set of commercial API-gated services. There’s models that you can install in the open source. There are models in the open source today that are rivalling in quality with the best closed API models. So we think the ecosystem is developing into a rich diversification that will allow customers to have a diversity of modalities for using AI, which is exciting. 

AI-native customers accounted for 6% of Datadog’s ARR in 2024 Q4 (was 6% 2024 Q3); AI-native customers contributed 5 percentage points to Datadog’s year-on-year growth in 2024 Q4, compared to 3 percentage points in 2023 Q4; among customers in the AI-native cohort, management has seen optimisation of usage and volume discounts related to contracts in 2024 Q4, and management thinks these customers will continue to optimise cloud and observability usage in the future; the dynamics with the AI-native cohort that happened in 2024 Q4 was inline with management’s expectations

We continue to see robust contribution from AI native customers who represented about 6% of Q4 ARR roughly the same as the quarter — as last quarter and up from about 3% of ARR in the year-ago quarter. AI native customers contributed about 5 percentage points of year-over-year revenue growth in Q4 versus 4 points in the last quarter and about 3 points in the year-ago quarter. So we saw strong growth from AI native customers in Q4. We believe that adoption of AI will continue to benefit Datadog in the long term. Meanwhile, we did see some optimization and volume discounts related to contract renewals in Q4. We remain mindful that we may see volatility in our revenue growth on the backdrop of long-term volume growth from this cohort as customers renew with us on different terms, and as they may choose to optimize cloud and observability usage. ..

… [Question] I’m trying to understand if the AI usage and commits are kind of on the same trajectory that they were on or whether you feel that there are some oscillations there.

[Answer] What happened during the quarter is pretty much what we thought would happen when we discussed it in the last earnings call. When you look at the AI cohort, we definitely saw some renewals with higher commit, better terms and optimization usage all at the same time, which is fairly typical, which typically happens with larger end customers in particular is at the time of renewal, customers are going to trying and optimize what they can. They’re going to get better prices from us, up their commitments and we might see a flat or down a month or quarter after that, with a still sharp growth from the year before and growth to come in the year to come. So that’s what we typically see. When you look at the cohort as a whole, even with that significant renewal optimization and better unit economics this quarter is wholly stable, this quarter as a whole is stable quarter-to-quarter in its revenue and it’s growing a lot from the quarter before, even with all that.

Datadog’s management sees some emerging opportunities in Infrastructure Monitoring that are related to the usage of GPUs (Graphics Processing Units), but the opportunities will only make sense if there is broad usage of GPUs by a large number of customers, which is not happening today

There’s a number of new use cases that are emerging that are related to infrastructure that we might want to cover. Again, we — when I say they’re emerging, they’re actually emerging, like we still have to see what the actual need is from a large number of customers. I’m talking in particular about infrastructure concerns around GPU management, GPU optimization, like there’s quite a lot going on there that we can potentially do. But we — for that, we need to see broad usage of the raw GPUs by a large number of customers as opposed to usage by a smaller number of native customers, which is mostly what we still see today.

Datadog’s management thinks it’s hard to tell where AI agents can be the most beneficial for Datadog’s observability platform because it’s still a very nascent field and management has observed that things change really quickly; when management built LLM Observability, the initial use cases were for AI chatbots and RAG (retrieval augmented generation), but now the use cases are starting to shift towards AI agents

[Question] Just curious, when we think about agents, which parts of the core observability platform that you think are most relevant or going to be most beneficial to your business as you start to monitor those?

[Answer] It’s a bit hard to tell because it’s a very nascent field. So my guess is in a year if we probably look different from what it looks like today. Just like this year, it looks very different from what it looks like last year. What we do see, though, is that — so when we built — we started building our LLM Observability product, most of the use cases we saw there from customers were chatbot in nature or RAG in nature, trying to access information and return the information. Now we see more and more customers building agents on top of that and sending the data from their agents. So we definitely see a growing trend there of adoption and the LLM Observability product is a good level of abstraction, at least for the current iteration of these agents to get them. So that’s what we can see today.

Datadog’s management sees AI touching many different areas of Datadog, such as how software is being written and deployed, how customer-support is improved, and more

What’s fascinating about the current evolution of AI, in particular, is that it touches a lot of the different areas of the business. The first area for company like ours the first area to be transformed is really the way software is being built. What engineers use, how they write software, how they debug software, how do they also operate systems. And part of that is outside tooling we’re using for writing software. Part of that is dogfooding, or new products for incident resolution and that sort of thing. So that’s the first area. There’s a number of other areas that are going to see large improvements in productivity. Typically, everything that has to do with supporting customers, helping with onboarding and helping troubleshoot issues like all of that is in acceleration. In the end, we expect to see improvements everywhere, from front office to back office.

Fiverr (NYSE: FVRR)

Fiverr’s management launched Fiverr Go in February 2025, an open platform for personalised AI tools designed to give creators full control over their creative processes and rights; Fiverr Go enables freelancers to build personalised AI models (there was a presentation on this recently) without having to know AI engineering; Fiverr Go is designed to be personalised for the creator, so the creator becomes more important compared to the AI technology; Fiverr Go is generative AI technology with human accountability (will be interesting to see if Fiverr Go is popular; people can create designs/images with other AI models, so customers who use Fiverr Go are those who need the special features that Fiverr Go offers); Fiverr Go generates content that is good enough for mission critical business tasks, unlike what’s commonly happening with other AI-generated content; Fiverr Go is no different from a direct order from the freelancer themself, except it is faster and easier for buyers; Fiverr Go has personalised AI assistants for freelancers; Fiverr Go has an open developer platform for 3rd-party developers to build generative AI apps

Fiverr Go is an open platform for personalized AI tools that include the personalized AI assistant and the AI creation model. Different from other AI platforms that often exploit human creativity without proper attribution or compensation, Fiverr Go is uniquely designed to reshape this power dynamic by giving creators full control over their creative process and rights. It also enables freelancers to build personalized AI models without the need to collect training data sets or understand AI engineering, thanks to Fiverr’s unparalleled foundation of over 6.5 billion interactions and nearly 150 million transactions on the marketplace and most importantly, it allows freelancers to become a one-person production house, making more money while focusing on the things that matter. By giving freelancers control over configuration, pricing and creative rights and leveling the playing field of implementing AI technology, Fiverr Go ensures that creators remain at the center of the creative economy. It decisively turned the power dynamic between humans and AI towards the human side…

…For customers, Fiverr Go is also fundamentally different from other AI platforms. It is GenAI with human accountability. AI results often feel unreliable, generic and very hard to edit. What is good enough for a simple question and answer on ChatGPT does not cut it for business mission-critical tasks. In fact, many customers come to Fiverr today with AI-generated content because they miss the confidence that comes from another human eye and talent, helping them perfect the results for their needs. Fiverr Go eliminates all of this friction and frustration. Every delivery on Fiverr Go is backed by the full faith of the creator behind it with an included revision as the freelancer defines. This means that the quality and service you get from Fiverr Go is no different from a direct order from the freelancers themselves, except for a faster, easier and more delightful experience. The personalized AI assistant on Fiverr Go can communicate with potential clients when the freelancer is away or busy, handle routine tasks and provide actionable business insights, effectively becoming the creator’s business partner. It often feels smarter than an average human assistant because it’s equipped with all the history of how the freelancer works as well as knowledge of trends and best practices on the Fiverr marketplace…

…We’ve also announced an open developer platform on Fiverr Go to allow AI specialists and model developers to build generative AI applications across any discipline. We provide them with an ecosystem to collaborate with domain experts on Fiverr and the ability to leverage Fiverr’s massive data assets so that these applications can solve real-world problems and most important of all, Fiverr provides them an avenue to generate revenue from those applications through our marketplace…

…So from our experience with AI, what we come to learn is that a lot of the creation process using AI is very random and take you through figuring out what are the best tools because there’s thousands of different options around AI. And each one operates slightly differently. And you need to master each one of them. And you need to become a prompt engineer. And then editing is extremely, extremely hard. Plus you don’t get the feedback that comes from working with a human being that can actually look at the creation from a human eye and give you a sense if this is actually capturing what you’re trying to do. It allows us or allows freelancers to design their own model in a way that rewards them but remains extremely accurate to their style, allowing customers to get the results they expect to get because they see the portfolio of their freelancer, like the style of writing or design or singing or narration, and they can get exactly this. So we think that, that combination and that confidence that comes from the fact that the creator itself is always there.

The AI personal assistant in Fiverr Go can help to respond to customer questions based on individual freelancers; the first 3 minutes after a buyer writes to a freelancer is the most crucial time for conversion and this is where the AI assistant can help; there are already thousands of AI assistants running on Fiverr Go, converting customers

Fiverr Go is actually a tool for conversion. That’s the entire idea because we know that customers these days expect instant responses and instant results. And as a result of that, we designed those 2 tools, the AI personal assistant, which is able to answer customer questions immediately even if the freelancer is away or busy. We know that the first 3 minutes after a customer writes to a freelancer are the most crucial time for conversion and this is why we designed this tool. And this tool is essentially encapsulating the entire knowledge of the freelancer and basing itself on it, being able to address any possible question and bring it to conversion…

…It’s fresh from yesterday, but we have many thousands of assistants running on the system, converting customers already, which is an amazing sign.

Fiverr Go is a creator tool that can create works based off freelancers’ style and allows customers to get highly-accurate samples of a freelancers’ work to lower friction in selecting freelancers

When we think about the creation model, the creation model allows customers to get the confidence that this is the freelancer, this is the style that they’re looking for, because now instead of asking a freelancer for samples, waiting for it, causing the freelancer to essentially work for free, they can get those samples right away. Now the quality of these samples is just mind-blowing. The level of accuracy that these samples produce are exact match with the style of the freelancer, which gives the customer the confidence that if they played and liked it, this is the type of freelancer that they should engage with.

The Fiverr Go open developer platform is essentially an app store for AI apps; the open developer platform allows developers to train AI models on Fiverr’s transactional data set, which is probably the largest dataset of its kind in existence

Now what we’re doing with this is actually we’re opening up the Go platform to outside developers. Think about it as an app store in essence. So what we’re doing is we’re allowing them to develop models, APIs, workflows, but then train those models on probably the biggest transactional data set in existence today that we hold so that they can actually help us build models that freelancers can join — can enjoy from. And we believe that by doing so and giving those developers incentives to do so because every time their app is going to be used for a transaction, they’re going to make money out of it.

Fiverr Go’s take rate will be the same for now and management will learn as they go

[Question] Would your take rate be different in Fiverr Go?

[Answer] For now, the take rate remains the same for Go. And as we roll it out and as we see usage, we will figure out what to do or what’s the right thing to do. For now, we treat it as a normal transaction with the same take rate.

Mastercard (NYSE: MA)

Mastercard closed the Recorded Future acquisition in 2024 Q4 (Recorded Future provides AI-powered solutions for real-time visibility into potential threats related to fraud); Recorded Future has been deploying AI for over a decade, just like Mastercard has; Recorded Future uses AI to analyse threat data across the entire Internet; the acquisition of Recorded Future improves Mastercard’s cybersecurity solutions

Our diverse capabilities in payments and services and solutions, including the acquisition of Recorded Future this quarter set us apart…

…Recorded Future is the world’s largest threat intelligence company with more than 1,900 customers across 75 countries. Customers include over 50% of the Fortune 100 and government agencies in 45 countries, including more than half of the G20. We’ve been deploying AI at scale for well over a decade, so has Recorded Future. They leverage AI-powered insights to analyze threat data from every corner of the Internet and customers gain real-time visibility and actionable insights to proactively reduce risks. We now have an even more robust set of powerful intelligence, identity, dispute, fraud and scan prevention solutions. Together, these uniquely differentiated technologies will enable us to create smarter models, distribute these capabilities more broadly and help our customers anticipate threats before cyber-attacks can take place. That means better protection for governments, businesses, banks, consumers the entire ecosystem and well beyond the payment transactions. We’re also leveraging our distribution at scale to deepen market penetration of our services and solutions.

Meta Platforms (NASDAQ: META)

Meta’s management expects Meta AI to be the leading AI assistant in 2025, reaching more than 1 billion people; Meta AI is already the most-used AI assistant in the world with more than 700 million monthly actives; management believes Meta AI is at a scale that allows it to develop a durable long-term advantage; management has an exciting road map for Meta AI in 2025 that focuses on personalisation; management does not believe that there’s going to be only one big AI that is the same for everyone; there are some fun surprises for Meta AI in 2025 that management has up their sleeves; Meta AI can now remember certain details of people’s prior queries; management sees a few possible paths for Meta AI’s monetisation, but their focus right now is just on building a great user experience; WhatsApp has the strongest Meta AI usage, followed by Facebook; people are using Meta AI on WhatsApp for informational, educational, and emotional purposes 

 In AI, I expect that this is going to be the year when a highly intelligent and personalized AI assistant reaches more than 1 billion people, and I expect Meta AI to be that leading AI assistant. Meta AI is already used by more people than any other assistant. And once a service reaches that kind of scale, it usually develops a durable long-term advantage.

We have a really exciting road map for this year with a unique vision focused on personalization. We believe that people don’t all want to use the same AI. People want their AI to be personalized to their context, their interests, their personality, their culture, and how they think about the world. I don’t think that there’s just going to be one big AI that everyone uses that does the same thing. People are going to get to choose how their AI works and what it looks like for them. I continue to think that this is going to be one of the most transformative products that we’ve made, and we have some fun surprises that I think people are going to like this year…

… Meta AI usage continues to scale with more than 700 million monthly actives. We’re now introducing updates that will enable Meta AI to deliver more personalized and relevant responses by remembering certain details from people’s prior queries and considering what they engage with on Facebook and Instagram to develop better intuition for their interest and preferences…

…Our initial focus for Meta AI is really about building a great consumer experience, and that’s frankly, where all of our energies are kind of directed to right now. There will, I think, be pretty clear monetization opportunities over time, including paid recommendations and including a premium offering, but really not where we are focused in terms of the development of Meta AI today…

…WhatsApp continues to see the strongest Meta AI usage across our family of apps. People there are using it most frequently for information seeking and educational queries along with emotional support use cases. Most of the WhatsApp engagement is in one-on-one threads, though we see some usage in group messaging. And on Facebook, which is the second largest driver of Meta AI engagement, we’re seeing strong engagement from our feed deep dives integration that lets people ask Meta AI questions about the content that is recommended to that. So across, I would say, all query types, we continue to see signs that Meta AI is helping people leverage our apps for new use cases. We talked about information gathering, social interaction and communication Lots of people use it for humor and casual conversation. They use it for writing and editing research recommendations. 

Meta’s management thinks Llama will become the most advanced and widely-used AI model in 2025; Llama 4 is making great progress; Meta has a reasoning Llama model in the works; management’s goal for Llama 4 is for it be the leading AI model; Llama 4 is built to be multi-modal and agentic; management expects Llama 4 to unlock a lot of new use cases

I think this will very well be the year when Llama and open-source become the most advanced and widely used AI models as well. Llama 4 is making great progress in training, Llama 4 Mini is doing with pretraining and our reasoning models and larger model are looking good too. 

Our goal with Llama 3 was to make open source competitive with closed models. And our goal for Llama 4 is to lead. Llama 4 will be natively multimodal. It’s an omni model, and it will have agenetic capabilities. So it’s going to be novel, and it’s going to unlock a lot of new use cases.

Meta’s management thinks it will be possible in 2025 to build an AI engineering agent that is as capable as a human mid-level software engineer; management believes that the company that builds this AI engineering agent first will have a meaningful advantage in advancing AI research; Meta already has internal AI coding assistants, powered by Llama; management has no plan to release the AI engineer as an external product anytime soon, but sees the potential for it in the longer-term; management does not expect the AI engineer to be extremely widely deployed in 2025, with the dramatic changes happening in 2026 and beyond

I also expect that 2025 will be the year when it becomes possible to build an AI engineering agent that has coding and problem-solving abilities of around a good mid-level engineer. And this is going to be a profound milestone and potentially one of the most important innovations in history, like as well as over time, potentially a very large market. Whichever company builds this first, I think it’s going to have a meaningful advantage in deploying it to advance their AI research and shape the field…

…As part of our efficiency focus over the past 2 years, we’ve made significant improvements in our internal processes and developer tools and introduce new tools like our AI-powered coding assistant, which is helping our engineers write code more quickly. Looking forward, we expect that the continuous advancements in Llama’s coding capabilities will provide even greater leverage to our engineers, and we are focused on expanding its capabilities to not only assist our engineers in writing and reviewing our code but to also begin generating code changes to automate tool updates and improve the quality of our code base…

…And then the AI engineer piece, I’m really excited about it. I mean, I don’t know that that’s going to be an external product anytime soon. But I think for what we’re working on, our goal is to advance AI research and advance our own development internally. And I think it’s just going to be a very profound thing. So I mean that’s something that I think will show up through making our products better over time. But — and then as that works, there will potentially be a market opportunity down the road. But I mean, for now and this year, we’re really — I think this is — I don’t think you’re going to see this year like an AI engineer that is extremely widely deployed, changing all of development. I think this is going to be the year where that really starts to become possible and lays the groundwork for a much more dramatic change in ’26 and beyond.

The Ray-Ban Meta AI glasses are a big hit so far but management thinks 2025 will be the pivotal year to determine if the AI glasses can be on a path towards being the next computing platform and selling hundreds of millions, or more, units; management continues to think that glasses are the perfect form factor for AI; management is optimistic about AI glasses, but there’s still uncertainty about the long-term trajectory

Our Ray-Ban Meta AI glasses are a real hit. And this will be the year when we understand the trajectory for AI glasses as a category. Many breakout products in the history of consumer electronics have sold 5 million to 10 million units and they’re third generation. This will be a defining year that determines if we’re on a path towards many hundreds of millions and eventually billions of AI glasses and glasses being the next computing platform like we’ve been talking about for some time or if this is just going to be a longer grind. But it’s great overall to see people recognizing that these glasses are the perfect form factor for AI as well as just great stylish glasses…

…There are a lot of people in the world who have glasses. It’s kind of hard for me to imagine that a decade or more from now, all the glasses aren’t going to basically be AI glasses as well as a lot of people who don’t wear glasses today, finding that to be a useful thing. So I’m incredibly optimistic about this…

…But look, the Ray-Ban Metas were hit. We still don’t know what the long-term trajectory for this is going to be. And I think we’re going to learn a lot this year. 

Meta will bring ~1 gigawatt of AI data center capacity online in 2025 and is building an AI data center that is at least 2 gigawatts in capacity; management intends to fund the investments through revenue growth that is driven by its AI advances; most of Meta’s new headcount growth will go towards its AI infrastructure and AI advances; management expects compute will be very important for the opportunities they want Meta to pursue; management is simultaneously growing Meta’s capacity and increasing the efficiency of its workloads; Meta is increasing the useful lives of its non-AI and AI servers to 5.5 years (from 4-5 years previously), which will lead to lower depreciation expenses per year; Meta started deploying its own MTIA (Meta Training and Inference Accelerator) AI chips in 2024 for inference workloads; management expects to ramp up MTIA usage for inference in 2025 and for training workloads in 2026; management will continue to buy third-party AI chips (likely referring to NVIDIA), but wants to use in-house chips for unique workloads; management thinks MTIA helps Meta achieve greater compute efficiency and performance per cost and power; management has been thinking about the balance of compute used in pre-training versus inference, but this does not mean that Meta will need less compute; management thinks that inference-time compute (or test-time compute) scaling will help Meta deliver a higher quality of service and that Meta has a strong business model to support the delivery of inference-time compute scaling; management believes that investing heavily in AI infrastructure is still going to be a strategic advantage over time, but it’s possible the reverse may be true in the future; management thinks it’s too early to tell what the long-run capacity intensity will look like

I announced last week that we expect to bring online almost a gigawatt of capacity this year. And we’re building a 2 gigawatt and potentially bigger AI data center that is so big that it will cover a significant part of Manhattan if we were placed there. We’re planning to fund all of this by, at the same time, investing aggressively in initiatives that use these AI advances to increase revenue growth…

…That’s what a lot of our new headcount growth is going towards and how well we execute on this will also determine our financial trajectory over the next few years…

…We expect compute will be central to many of the opportunities we’re pursuing as we advance the capabilities of Llama, drive increased usage of generative AI products and features across our platform and fuel core ads and organic engagement initiatives. We’re working to meet the growing capacity needs for these services by both scaling our infrastructure footprint and increasing the efficiency of our workloads…

…Our expectation going forward is that we’ll be able to use both our non-AI and AI [indiscernible] servers for a longer period of time before replacing them, which we estimate will be approximately 5.5 years. This will deliver savings in annual CapEx and resulting depreciation expense, which is already included in our guidance.

Finally, we’re pursuing cost efficiencies by deploying our custom MTIA silicon in areas where we can achieve a lower cost of compute by optimizing the chip to our unique workloads. In 2024, we started deploying MTIA to our ranking and recommendation inference workloads for ads and organic content. We expect to further ramp adoption of MTIA for these use cases throughout 2025, before extending our custom silicon efforts to training workloads for ranking and recommendations next year…

…We expect that we are continuing to purchase third-party silicon from leading providers in the industry. And we are certainly committed to those long-standing partnerships, but we’re also very invested in developing our own custom silicon for unique workloads, where off-the-shelf silicon isn’t necessarily optimal and specifically because we’re able to optimize the full stack to achieve greater compute efficiency and performance per cost and power because our workloads might require a different mix of memory versus network, bandwidth versus compute and so we can optimize that really to the specific needs of our different types of workloads.

Right now, the in-house MTIA program is focused on supporting our core ranking and recommendation inference workloads. We started adopting MTIA in the first half of 2024 for core ranking and recommendations in [indiscernible]. We’ll continue ramping adoption for those workloads over the course of 2025 as we use it for both incremental capacity and to replace some GPU-based servers when they reach the end of their useful lives. Next year, we’re hoping to expand MTIA to support some of our core AI training workloads and over time, some of our Gen AI use cases…

…There’s already sort of a debate around how much of the compute infrastructure that we’re using is going to go towards pretraining versus as you get more of these reasoning time models or reasoning models where you get more of the intelligence by putting more of the compute into inference, whether just will mix shift how we use our compute infrastructure towards that. That was already something that I think a lot of the — the other labs and ourselves were starting to think more about and already seemed pretty likely even before this, that — like of all the compute that we’re using, that the largest pieces aren’t necessarily going to go towards pre-training. But that doesn’t mean that you need less compute because one of the new properties that’s emerged is the ability to apply more compute at inference time in order to generate a higher level of intelligence and a higher quality of service, which means that as a company that has a strong business model to support this, I think that’s generally an advantage that we’re now going to be able to provide a higher quality of service than others who don’t necessarily have the business model to support it on a sustainable basis…

…I continue to think that investing very heavily in CapEx and infra is going to be a strategic advantage over time. It’s possible that we’ll learn otherwise at some point, but I just think it’s way too early to call that…

…I think it is really too early to determine what long-run capital intensity is going to look like. There are so many different factors. The pace of advancement in underlying models, how efficient can they be? What is the adoption and use case of our Gen AI products, what performance gains come from next-generation hardware innovations, both our own and third party and then ultimately, what monetization or other efficiency gains our AI investments unlock. 

In 2024 H2, Meta introduced a new machine learning system for ads ranking, in partnership with Nvidia, named Andromeda; Andromeda has enabled a 10,000x increase in the complexity of AI models Meta uses for ads retrieval, driving an 8% increase in quality of ads that people see; Andromeda can process large volumes of ads and positions Meta well for a future where advertisers use the company’s generative AI tools to create and test more ads

In the second half of 2024, we introduced an innovative new machine learning system in partnership with NVIDIA called Andromeda. This more efficient system enabled a 10,000x increase in the complexity of models we use for ads retrieval, which is the part of the ranking process where we narrow down a pool of tens of millions of ads to the few thousand we consider showing someone. The increase in model complexity is enabling us to run far more sophisticated prediction models to better personalize which ads we show someone. This has driven an 8% increase in the quality of ads that people see on objectives we’ve tested. Andromeda’s ability to efficiently process larger volumes of ads also positions us well for the future as advertisers use our generative AI tools to create and test more ads.

Advantage+ has surpassed a $20 billion annual revenue run rate and grew 70% year-on-year in 2024 Q4; Advantage+ will now be turned on by default for all campaigns that optimise for sales, app, or lead objectives; more than 4 million advertisers are now using at least one of Advantage+’s generative AI ad creative tools, up from 1 million six months ago; Meta’s first video generation tool, released in October, already has hundreds of thousands of advertisers using it monthly

 Adoption of Advantage+ shopping campaigns continues to scale with revenues surpassing a $20 billion annual run rate and growing 70% year-over-year in Q4. Given the strong performance and interest we’re seeing in Advantage+ shopping and our other end-to-end solutions, we’re testing a new streamlined campaign creation flow. So advertisers no longer need to choose between running a manual or Advantage+ sales or app campaign. In this new setup, all campaigns optimizing for sales, app or lead objectives will have Advantage+ turned on from the beginning. This will allow more advertisers to take advantage of the performance Advantage+ offers while still having the ability to further customize aspects of their campaigns when they need to. We plan to expand to more advertisers in the coming months before fully rolling it out later in the year.

Advantage+ Creative is another area where we’re seeing momentum. More than 4 million advertisers are now using at least one of our generative AI ad creative tools, up from 1 million six months ago. There has been significant early adoption of our first video generation tool that we rolled out in October, Image Animation, with hundreds of thousands of advertisers already using it monthly.

Meta’s management thinks the emergence of DeepSeek makes it even more likely for a global open source standard for AI models to develop; the presence of DeepSeek also makes management think it’s important that the open source standard be made in America and that it ’s even more important for Meta to focus on building open source AI models; Meta is learning from DeepSeek’s innovations in building AI models; management currently does not have a strong opinion on how Meta’s capex plans for AI infrastructure will change because of the recent news with DeepSeek 

I also just think in light of some of the recent news, the new competitor DeepSeek from China, I think it also just puts — it’s one of the things that we’re talking about is there’s going to be an open source standard globally. And I think for our kind of national advantage, it’s important that it’s an American standard. So we take that seriously, and we want to build the AI system that people around the world are using and I think that if anything, some of the recent news has only strengthened our conviction that this is the right thing for us to be focused on…

…I can start on the DeepSeek question. I think there’s a number of novel things that they did that I think we’re still digesting. And there are a number of things that they have advances that we will hope to implement in our systems. And that’s part of the nature of how this works, whether it’s a Chinese competitor or not…

…It’s probably too early to really have a strong opinion on what this means for the trajectory around infrastructure and CapEx and things like that. There are a bunch of trends that are happening here all at once.

Meta’s capex in 2025 is going to grow across servers, data centers, and networking; within each of servers, data centers, and networking, management expects growth in both AI and non-AI capex; management expects most of the AI-related capex in 2025 to be directed specifically towards Meta’s core AI infrastructure, but the infrastructure Meta is building can support both AI and non-AI workloads, and the GPU servers purchased can be used for both generative AI and core AI purposes

[Question] As we think about the $60 billion to $65 billion CapEx this year, does the composition change much from last year when you talked about servers as the largest part followed by data centers and networking equipment. And how should we think about that mix between like training and inference

[Answer] We certainly expect that 2025 CapEx is going to grow across all 3 of those components you described.

Servers will be the biggest growth driver that remains the largest portion of our overall CapEx budget. We expect both growth in AI capacity as we support our gen AI efforts and continue to invest meaningfully in core AI, but we are also expecting growth in non-AI capacity as we invest in the core business, including to support a higher base of engagement and to refresh our existing servers.

On the data center side, we’re anticipating higher data center spend in 2025 to be driven by build-outs of our large training clusters and our higher power density data centers that are entering the core construction phase. We’re expecting to use that capacity primarily for core AI and non-AI use cases.

On the networking side, we expect networking spend to grow in ’25 as we build higher-capacity networks to accommodate the growth in non-AI and core AI-related traffic along with our large Gen AI training clusters. We’re also investing in fiber to handle future cross-region training traffic.

And then in terms of the breakdown for core versus Gen AI use cases, we’re expecting total infrastructure spend within each of Gen AI, non-AI and core AI to increase in ’25 with the majority of our CapEx directed to our core business with some caveat that, that is — that’s not easy to measure perfectly as the data centers we’re building can support AI or non-AI workloads and the GPU-based servers, we procure for gen AI can be repurposed for core AI use cases and so on and so forth.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Alphabet, Amazon, Apple, ASML, Coupang, Datadog, Fiverr, Mastercard, and Meta Platforms. Holdings are subject to change at any time.

What We’re Reading (Week Ending 02 March 2025)

The best articles we’ve read in recent times on a wide range of topics, including investing, business, and the world in general.

We’ve constantly been sharing a list of our recent reads in our weekly emails for The Good Investors.

Do subscribe for our weekly updates through the orange box in the blog (it’s on the side if you’re using a computer, and all the way at the bottom if you’re using mobile) – it’s free!

But since our readership-audience for The Good Investors is wider than our subscriber base, we think sharing the reading list regularly on the blog itself can benefit even more people. The articles we share touch on a wide range of topics, including investing, business, and the world in general. 

Here are the articles for the week ending 02 March 2025:

1. Satya Nadella – Microsoft’s AGI Plan & Quantum Breakthrough – Dwarkesh Patel and Satya Nadella

Dwarkesh Patel

Where is the value going to be created in AI?

Satya Nadella

That’s a great one. So I think there are two places where I can say with some confidence. One is the hyperscalers that do well, because the fundamental thing is if you sort of go back to even how Sam and others describe it, if intelligence is log of compute, whoever can do lots of compute is a big winner.

The other interesting thing is, if you look at underneath even any AI workload, like take ChatGPT, it’s not like everybody’s excited about what’s happening on the GPU side, it’s great. In fact, I think of my fleet even as a ratio of the AI accelerator to storage, to compute. And at scale, you’ve got to grow it…

…Satya Nadella

So in fact it’s manna from heaven to have these AI workloads because guess what? They’re more hungry for more compute, not just for training, but we now know, for test time. When you think of an AI agent, it turns out the AI agent is going to exponentially increase compute usage because you’re not even bound by just one human invoking a program. It’s one human invoking programs that invoke lots more programs. That’s going to create massive, massive demand and scale for compute infrastructure. So our hyperscale business, Azure business, and other hyperscalers, I think that’s a big thing.

Then after that, it becomes a little fuzzy. You could say, hey, there is a winner-take-all model- I just don’t see it. This, by the way, is the other thing I’ve learned: being very good at understanding what are winner-take-all markets and what are not winner-take-all markets is, in some sense, everything. I remember even in the early days when I was getting into Azure, Amazon had a very significant lead and people would come to me, and investors would come to me, and say, “Oh, it’s game over. You’ll never make it. Amazon, it’s winner-take-all.”

Having competed against Oracle and IBM in client-server, I knew that the buyers will not tolerate winner-take-all. Structurally, hyperscale will never be a winner-take-all because buyers are smart.

Consumer markets sometimes can be winner-take-all, but anything where the buyer is a corporation, an enterprise, an IT department, they will want multiple suppliers. And so you got to be one of the multiple suppliers.

That, I think, is what will happen even on the model side. There will be open-source. There will be a governor. Just like on Windows, one of the big lessons learned for me was, if you have a closed-source operating system, there will be a complement to it, which will be open source.

And so to some degree that’s a real check on what happens. I think in models there is one dimension of, maybe there will be a few closed source, but there will definitely be an open source alternative, and the open-source alternative will actually make sure that the closed-source, winner-take-all is mitigated.

That’s my feeling on the model side. And by the way, let’s not discount if this thing is really as powerful as people make it out to be, the state is not going to sit around and wait for private companies to go around and… all over the world. So, I don’t see it as a winner-take-all.

Then above that, I think it’s going to be the same old stuff, which is in consumer, in some categories, there may be some winner-take-all network effect. After all, ChatGPT is a great example.

It’s an at-scale consumer property that has already got real escape velocity. I go to the App Store, and I see it’s always there in the top five, and I say “wow, that’s pretty unbelievable”.

So they were able to use that early advantage and parlay that into an app advantage. In consumer, that could happen. In the enterprise again, I think there will be, by category, different winners. That’s sort of at least how I analyze it…

…Satya Nadella

The way I come at it, Dwarkesh, it’s a great question because at some level, if you’re going to have this explosion, abundance, whatever, commodity of intelligence available, the first thing we have to observe is GDP growth.

Before I get to what Microsoft’s revenue will look like, there’s only one governor in all of this. This is where we get a little bit ahead of ourselves with all this AGI hype. Remember the developed world, which is what? 2% growth and if you adjust for inflation it’s zero?

So in 2025, as we sit here, I’m not an economist, at least I look at it and say we have a real growth challenge. So, the first thing that we all have to do is, when we say this is like the Industrial Revolution, let’s have that Industrial Revolution type of growth.

That means to me, 10%, 7%, developed world, inflation-adjusted, growing at 5%. That’s the real marker. It can’t just be supply-side.

In fact that’s the thing, a lot of people are writing about it, and I’m glad they are, which is the big winners here are not going to be tech companies. The winners are going to be the broader industry that uses this commodity that, by the way, is abundant. Suddenly productivity goes up and the economy is growing at a faster rate. When that happens, we’ll be fine as an industry.

But that’s to me the moment. Us self-claiming some AGI milestone, that’s just nonsensical benchmark hacking to me. The real benchmark is: the world growing at 10%.

Dwarkesh Patel

Okay, so if the world grew at 10%, the world economy is $100 trillion or something, if the world grew at 10%, that’s like an extra $10 trillion in value produced every single year. If that is the case, you as a hyperscaler… It seems like $80 billion is a lot of money. Shouldn’t you be doing like $800 billion?

If you really think in a couple of years, we could be really growing the world economy at this rate, and the key bottleneck would be: do you have the compute necessary to deploy these AIs to do all this work?

Satya Nadella

That is correct. But by the way, the classic supply side is, “Hey, let me build it and they’ll come.” That’s an argument, and after all we’ve done that, we’ve taken enough risk to go do it.

But at some point, the supply and demand have to map. That’s why I’m tracking both sides of it. You can go off the rails completely when you are hyping yourself with the supply-side, versus really understanding how to translate that into real value to customers.

That’s why I look at my inference revenue. That’s one of the reasons why even the disclosure on the inference revenue… It’s interesting that not many people are talking about their real revenue, but to me, that is important as a governor for how you think about it.

You’re not going to say they have to symmetrically meet at any given point in time, but you need to have existence proof that you are able to parlay yesterday’s, let’s call it capital, into today’s demand, so that then you can again invest, maybe exponentially even, knowing that you’re not going to be completely rate mismatched.

Dwarkesh Patel

I wonder if there’s a contradiction in these two different viewpoints, because one of the things you’ve done wonderfully is make these early bets. You invested in OpenAI in 2019, even before there was Copilot and any applications.

If you look at the Industrial Revolution, these 6%, 10% build-outs of railways and whatever things, many of those were not like, “We’ve got revenue from the tickets, and now we’re going to…”

Satya Nadella

There was a lot of money lost.

Dwarkesh Patel

That’s true. So, if you really think there’s some potential here to 10x or 5x the growth rate of the world, and then you’re like, “Well, what is the revenue from GPT-4?”

If you really think that’s the possibility from the next level up, shouldn’t you just, “Let’s go crazy, let’s do the hundreds of billions of dollars of compute?” I mean, there’s some chance, right?

Satya Nadella

Here’s the interesting thing, right? That’s why even that balanced approach to the fleet, at least, is very important to me. It’s not about building compute. It’s about building compute that can actually help me not only train the next big model but also serve the next big model. Until you do those two things, you’re not going to be able to really be in a position to take advantage of even your investment.

So, that’s kind of where it’s not a race to just building a model, it’s a race to creating a commodity that is getting used in the world to drive… You have to have a complete thought, not just one thing that you’re thinking about.

And by the way, one of the things is that there will be overbuild. To your point about what happened in the dotcom era, the memo has gone out that, hey, you know, you need more energy, and you need more compute. Thank God for it. So, everybody’s going to race.

In fact, it’s not just companies deploying, countries are going to deploy capital, and there will be clearly… I’m so excited to be a leaser, because, by the way; I build a lot, I lease a lot. I am thrilled that I’m going to be leasing a lot of capacity in ’27, ’28 because I look at the builds, and I’m saying, “This is fantastic.” The only thing that’s going to happen with all the compute builds is the prices are going to come down…

…Satya Nadella

This has been another 30-year journey for us. It’s unbelievable. I’m the third CEO of Microsoft who’s been excited about quantum.

The fundamental breakthrough here, or the vision that we’ve always had is, you need a physics breakthrough in order to build a utility-scale quantum computer that works. We took the path of saying, the one way for having a less noisy or more reliable qubit is to bet on a physical property that by definition is more reliable and that’s what led us to the Majorana zero modes, which was theorized in the 1930s. The question was, can we actually physically fabricate these things? Can we actually build them?

So the big breakthrough effectively, and I know you talked to Chetan, was that we now finally have existence proof and a physics breakthrough of Majorana zero modes in a new phase of matter effectively. This is why we like the analogy of thinking of this as the transistor moment of quantum computing, where we effectively have a new phase, which is the topological phase, which means we can even now reliably hide the quantum information, measure it, and we can fabricate it. And so now that we have it, we feel like with that core foundational fabrication technique out of the way, we can start building a Majorana chip.

That Majorana One which I think is going to basically be the first chip that will be capable of a million qubits, physical. And then on that, thousands of logical qubits, error-corrected. And then it’s game on. You suddenly have the ability to build a real utility-scale quantum computer, and that to me is now so much more feasible. Without something like this, you will still be able to achieve milestones, but you’ll never be able to build a utility-scale computer. That’s why we’re excited about it…

…Satya Nadella

It’s a great question. One thing that I’ve been excited about is, even in today’s world… we had this quantum program, and we added some APIs to it. The breakthrough we had maybe two years ago was to think of this HPC stack, and AI stack, and quantum together.

In fact, if you think about it, AI is like an emulator of the simulator. Quantum is like a simulator of nature. What is quantum going to do? By the way, quantum is not going to replace classical. Quantum is great at what quantum can do, and classical will also…

Quantum is going to be fantastic for anything that is not data-heavy but is exploration-heavy in terms of the state space. It should be data-light but exponential states that you want to explore. Simulation is a great one: chemical physics, what have you, biology.

One of the things that we’ve started doing is really using AI as the emulation engine. But you can then train. So the way I think of it is, if you have AI plus quantum, maybe you’ll use quantum to generate synthetic data that then gets used by AI to train better models that know how to model something like chemistry or physics or what have you. These two things will get used together.

So even today, that’s effectively what we’re doing with the combination of HPC and AI. I hope to replace some of the HPC pieces with quantum computers.

2. Microsoft’s Majorana 1 chip carves new path for quantum computing – Catherine Bolgar

Microsoft today introduced Majorana 1, the world’s first quantum chip powered by a new Topological Core architecture that it expects will realize quantum computers capable of solving meaningful, industrial-scale problems in years, not decades.

It leverages the world’s first topoconductor, a breakthrough type of material which can observe and control Majorana particles to produce more reliable and scalable qubits, which are the building blocks for quantum computers.

In the same way that the invention of semiconductors made today’s smartphones, computers and electronics possible, topoconductors and the new type of chip they enable offer a path to developing quantum systems that can scale to a million qubits and are capable of tackling the most complex industrial and societal problems, Microsoft said…

…This new architecture used to develop the Majorana 1 processor offers a clear path to fit a million qubits on a single chip that can fit in the palm of one’s hand, Microsoft said. This is a needed threshold for quantum computers to deliver transformative, real-world solutions – such as breaking down microplastics into harmless byproducts or inventing self-healing materials for construction, manufacturing or healthcare. All the world’s current computers operating together can’t do what a one-million-qubit quantum computer will be able to do…

…The topoconductor, or topological superconductor, is a special category of material that can create an entirely new state of matter – not a solid, liquid or gas but a topological state. This is harnessed to produce a more stable qubit that is fast, small and can be digitally controlled, without the tradeoffs required by current alternatives…

…This breakthrough required developing an entirely new materials stack made of indium arsenide and aluminum, much of which Microsoft designed and fabricated atom by atom…

…Commercially important applications will also require trillions of operations on a million qubits, which would be prohibitive with current approaches that rely on fine-tuned analog control of each qubit. The Microsoft team’s new measurement approach enables qubits to be controlled digitally, redefining and vastly simplifying how quantum computing works.

This progress validates Microsoft’s choice years ago to pursue a topological qubit design – a high risk, high reward scientific and engineering challenge that is now paying off. Today, the company has placed eight topological qubits on a chip designed to scale to one million…

…But reaching the next horizon of quantum computing will require a quantum architecture that can provide a million qubits or more and reach trillions of fast and reliable operations. Today’s announcement puts that horizon within years, not decades, Microsoft said.

Because they can use quantum mechanics to mathematically map how nature behaves with incredible precision – from chemical reactions to molecular interactions and enzyme energies – million-qubit machines should be able to solve certain types of problems in chemistry, materials science and other industries that are impossible for today’s classical computers to accurately calculate…

…Most of all, quantum computing could allow engineers, scientists, companies and others to simply design things right the first time – which would be transformative for everything from healthcare to product development. The power of quantum computing, combined with AI tools, would allow someone to describe what kind of new material or molecule they want to create in plain language and get an answer that works straightaway – no guesswork or years of trial and error.

“Any company that makes anything could just design it perfectly the first time out. It would just give you the answer,” Troyer said. “The quantum computer teaches the AI the language of nature so the AI can just tell you the recipe for what you want to make.”…

…Qubits can be created in different ways, each with advantages and disadvantages. Nearly 20 years ago, Microsoft decided to pursue a unique approach: developing topological qubits, which it believed would offer more stable qubits requiring less error correction, thereby unlocking speed, size and controllability advantages. The approach posed a steep learning curve, requiring uncharted scientific and engineering breakthroughs, but also the most promising path to creating scalable and controllable qubits capable of doing commercially valuable work.

The disadvantage is – or was – that until recently the exotic particles Microsoft sought to use, called Majoranas, had never been seen or made. They don’t exist in nature and can only be coaxed into existence with magnetic fields and superconductors. The difficulty of developing the right materials to create the exotic particles and their associated topological state of matter is why most quantum efforts have focused on other kinds of qubits…

…Majoranas hide quantum information, making it more robust, but also harder to measure. The Microsoft team’s new measurement approach is so precise it can detect the difference between one billion and one billion and one electrons in a superconducting wire – which tells the computer what state the qubit is in and forms the basis for quantum computation.

The measurements can be turned on and off with voltage pulses, like flicking a light switch, rather than finetuning dials for each individual qubit. This simpler measurement approach that enables digital control simplifies the quantum computing process and the physical requirements to build a scalable machine…

…Majorana 1, Microsoft’s quantum chip that contains both qubits as well as surrounding control electronics, can be held in the palm of one’s hand and fits neatly into a quantum computer that can be easily deployed inside Azure datacenters.

3. The most underreported and important story in AI right now is that pure scaling has failed to produce AGI – Gary Marcus

On the order of half a trillion dollars has been invested on a premise that I have long argued was unlikely to succeed: the idea (sometimes informally referred to as the scaling hypothesis) that we could get to “artificial general intelligence” simply by adding more and more data and GPUs…

…Virtually all of the generative AI industry has been built on this presumption; projects like the OpenAI/Oracle/Softbank joint venture Stargate, allegedly another half trillion dollars, are also largely based on this premise…

…But I always knew it couldn’t last forever. When I said as much, the field was absolutely furious at me…

…The first signs that the pure scaling of data and compute might in fact be hitting a wall came from industry leaks from people like famed investor Marc Andreessen, who said in early November 2024 that current models are “sort of hitting the same ceiling on capabilities.” Then, in December, Microsoft CEO Satya Nadella echoed many of my 2022 themes, saying at a Microsoft Ignite event, “in the last multiple weeks there is a lot of debate on have we hit the wall with scaling laws. Is it gonna continue? Again, the thing to remember at the end of the day these are not physical laws. There are just empirical observations that hold true just like Moore’s Law did for a long period of time and so therefore it’s actually good to have some skepticism some debate.”…

…Finally, and perhaps most significantly: Elon Musk said over that weekend that Grok 3, with 15x the compute of Grok 2, and immense energy (and construction and chop) bills, would be “the smartest AI on the earth.” Yet the world quickly saw that Grok 3 is still afflicted by the kind of unreliability that has hobbled earlier models. The famous ML expert Andrej Karpathy reported that Grok 3 occasionally stumbles on basics like math and spelling. In my own experiments, I quickly found a wide array of errors, such as hallucinations (e.g, it told me with certainty that there was a significant 5.6-sized earthquake on Feb. 10 in Billings, Montana, when no such thing had happened) and extremely poor visual comprehension (e.g. it could not properly label the basic parts of a bicycle).

Nadella, in his December speech, pointed to test-time compute, in which systems are allowed extra time for “reasoning” as the next big thing, and to some degree he is right; it is the next big thing, a new thing to try to scale, since merely scaling compute and data is no longer bringing the massive returns it once did. At least for a while, adding more and more computing time will help, at least on some kinds of problems…

…although DeepSeek lowered the costs of training these new systems, they are still expensive to operate, which is why companies like OpenAI are limiting their usage. When customers begin to realize that even with the greater expenses, errors still seep in, they are likely to be disappointed. One irate customer cc:d me yesterday on a several page demand for a refund, writing in part that “GPT-4o Pro [which includes access to test time compute] has consistently underperformed,” and enumerated problems such as “Severely degraded memory” and “Hallucinations and Unreliable Answers.”…

…the illustrious Stanford Natural Language Processing group reached a similar conclusion, reading between the lines of OpenAI’s recent announcement in the same way I did. In their words, Altman’s recent OpenAI roadmap was “the final admission that the 2023 strategy of OpenAI, Anthropic, etc. ‘“simply scaling up model size, data, compute, and dollars spent will get us to AGI/ASI’) is no longer working!”

In short, half a trillion dollars have been invested in a misguided premise; a great deal more funding seems to be headed in the same direction for now.

4. Is Microsoft’s Copilot push the biggest bait-and-switch in AI – Tien Tzuo

Over a year ago, Microsoft launched Copilot Pro, an AI assistant embedded in its Office suite, with a $20/month price. The uptake apparently was abysmal. By October, they admitted that the way they were selling Copilot was not working out.

So what did they do? They forced it on Microsoft users by including Copilot in Office, and hiking up subscription fees. Microsoft first made this change in Asia, then fully pulled the plug across the globe last month, impacting 84 million subscribers. To add insult to injury, Microsoft renamed the product to Microsoft 365 Copilot. You didn’t want to pay for Copilot? Well, now you are…

…Not only is Microsoft’s Copilot rollout deceptive, it’s also embarrassingly disastrous.

This goes to show that even tech giants, including a major backer of AI pioneer OpenAI, can suffer the hype and competitive pressure surrounding AI. And it’s a stark reminder that what businesses should really be focused on instead is value — communicating it clearly and delivering it tangibly to customers…

…Well, there’s a good contrast to Microsoft’s approach — from Adobe.

Adobe took a different approach with its AI rollout last fall, resisting the temptation to immediately monetize its new video generator, instead using it to boost adoption and engagement. By positioning AI as a value-add rather than a paid extra, Adobe was playing the long game, building a loyal user base that would be ripe for future upselling once they experienced AI’s benefits for themselves.

5. Broken Markets!? – The Brooklyn Investor

So, I keep hearing that markets are broken, or that the market is as expensive as ever. I know I keep saying this and I am sounding like a broken record, but I am not so sure…

…But if you look at individual stocks, markets are clearly differentiating between individual stocks. Look at Nvidia vs. Intel. If nobody was really evaluating them and the market was ignoring fundamentals, you would think both stocks would be performing similarly. But they are clearly not. People are clearly differentiating between winners and losers. It’s a separate question whether they are over-discounting their views. That’s a different discussion, and contrary to the view that passive investing is killing fundamental analysis.

Another example: JPM, the better bank, is trading at 2.4x book, versus C, which is a crappier one, selling at 0.8x book. You can’t complain that the market is broken just because you don’t agree with it. On the other hand, Buffett in the 50s loved the fact that institutional investors of the time completely ignored company analysis / valuation…

…Look at all the rich folks at the Berkshire Hathaway annual meeting. Look at BRK stock since 1970. How often did it look ‘overpriced’? What about the market? What if people sold out when they thought BRK was overpriced? Or the market? Would they be as rich as they are now? Probably not. Maybe there are some smart folks that got in and out of BRK over the years, but I would bet that the richest of them are the ones that just held it and did nothing.

I keep telling people this, but if you look at all the richest people in the world, a lot of them are people who held a single asset, and held it for decades. Look at Bill Gates. What if he was hip to value investing and knew more about stocks and values. He may have sold out of MSFT when it looked really overvalued. What about Buffett? What about Bezos?

A lot of the rich used to be real estate moguls, and I thought a lot of them were wealthy because real estate was not very liquid, so they had no choice but to hold on even in bad times. Stocks have what you may call the “curse of liquidity”. It’s so easy to say, holy sh*t, something bad is going to happen, and *click*, you can get out of the market. Back in the 90s, we used to fear a 1-800 crash; people will call their brokers’ automated trade execution lines, 1-800-Sell-Everything, and go “Get me out of the market!!! See everything NOW!!!”, and the market would not be able to open the next morning. Some of us truly feared that, and hedge funds talked about that sort of thing all the time. But you can’t do that with your house.

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. We currently have a vested interest in Amazon and Microsoft. Holdings are subject to change at any time.