All articles

Divergence of Returns In The Index

Some companies have fallen hard.

The S&P 500 index is up about 8% in US dollar terms so far in 2026. That’s a decent return for less than half a year.  The index is also on track to surpass its historical 10% annual return for the year.

Similarly, the tech-heavy NASDAQ index is up 13% year-to-date.

Both indexes are also sitting near an all-time high.

Despite this, there is an interesting phenomenon happening.

Usually when the indexes are at an all-time high, you’d expect to see most companies to be up year-to-date. But that’s not the case this year

Although the indexes have performed well, nearly half of the indexes’ components are currently underwater for the year.

Of the 503 stocks in the S&P 500 index, 218 are negative for the year, while 45 of the NASDAQ’s 101 component stocks are down. The depth of some of the drawdowns are also quite steep.

Of the 218 stocks that are down in the S&P 500, 122 are currently more than 10% below where they started the year. And 56 are 20% or more underwater for the year. Meanwhile, more than 20% of the NASDAQ components are down more than 20% in 2026 thus far.

This has created a really interesting scenario where despite the indexes being near all-time highs, there are pockets of the stock market, even in the large cap arena, that are spotting materially cheaper valuations than they did just five months ago.

Hunting for value

Lower share prices naturally mean lower valuations and could be a good hunting ground for value investors. 

With the index at all time highs, searching amidst beaten-down individual names could be a great way to gain exposure to the market.

The list of stocks that are down year-to-date include some well known companies such as FICO, Lululemon, Tractor Supply, and Accenture to name a few. The four listed companies are down 35%, 43%, 39% and 37% respectively this year. 

This being said, just because a company is trading at a lower stock price does not automatically make it good value. Even seemingly stable companies can run ahead of fundamentals and corrections may just be stock prices coming down to more sane valuations.

Previously “deep-moat” companies can also run into trouble or face disruption, as is the fear surrounding software-as-a-service companies as they are potentially facing disruption from artificial intelligence.

Nevertheless, as an investor, seeing that there is a substantial list of big cap stocks that are trading down for the year does excite me.

Why is the stock down?

When hunting for value, it is important to understand why the stock is down. There could be a legitimate reason for a stock to fall.

For instance, FICO, the company behind the FICO credit score that banks in the US use to assess whether to provide loans to someone, is facing a potential new competitor in the form of Vantage Score which could lead to market share losses in the future. (Vantage Score has existed for many years, but there are recent regulatory changes that have eaten away at FICO’s previous monopolistic status.)

A stock could also be down simply because its price had run ahead of its fundamentals.

Take Palantir for instance. The company just reported stellar revenue growth of 85% in the first quarter of 2026 and is guiding for revenue growth of more than 100% for 2026.

Yet the stock price is down 25% year-to-date. This could simply be because Palantir was trading at an overly expensive valuation of 100 times its 2026 free cash flow guidance. With the aforementioned year-to-date decline, Palantir now trades at a more reasonable but still expensive 74 times its 2026 free cash flow guidance.

Happy hunting

Although there are companies that are facing challenges and so have stock prices that are down for a reason, there are potentially also companies that may be mispriced after a steep drawdown.

This could provide a nice entry point for patient investors who are willing to ride out the negative sentiment and potential downward momentum.

It is also a great way to enter the market if you are not keen to buy directly into the index which is trading at an all-time high price.

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 don’t have a vested interest in any company mentioned. Holdings are subject to change at any time.

What We’re Reading (Week Ending 17 May 2026)

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 17 May 2026:

1. A Government Debt Crisis? – Ben Carlson

One of my favorites is the 1972 Time Magazine cover story:

This sounds like it could have been written today:

Debt service is now the third highest public expense, exceeded only by spending for defense and education; most of the money goes to banks, which are the major buyers of bonds that governments at all levels sell to cover their deficits. Moreover, debt functions as a wrong-way income redistribution device, channeling tax money that is paid in large part by the poor and the middle class into the pockets of wealthy holders of trust accounts or stock in banks.

When this cover was published, government debt was roughly $430 billion.

Today it’s fast approaching $40 trillion in total…

…The Wall Street Journal shows that publicly held debt to GDP is now 100% for the first time since WWII..

…Here’s the trillion dollar question — why have none of the government debt crisis predictions come to fruition?…

There are two big mistakes people make when they predict a catastrophe from U.S. government debt levles:

1. Conflating U.S. government debt with household debt. Government debt is not like a mortgage that needs to be paid back. As long as the economy keep growing, debt levels will likely keep rising.1 Plus, the U.S. government has the ability to print the global reserve currency. You can’t print more dollar bills in your basement.

2. The government’s liabilities are someone else’s assets. Treasuries are bonds owned by pensions, insurance companies, fund managers, and households. It’s the largest, most liquid bond in the world and there isn’t an alternative…

…So what would make me worry about government debt levels?

The biggest risk of large deficits and government spending is inflation…

…Continuously rising interest rates would also be cause for concern…

…Another concern is the fact that interest expenses are becoming a larger share of the government’s budget…

…Interest expenses now exceed the defense budget.

The good news is that interest expense as a percentage of GDP is at 1980s levels.

The bad news is that it has risen like a rocket and rates were a lot higher back then…

…Is there a line in the sand where a government debt crisis automatically kicks in?

No one knows.

2. China’s $3 Trillion of Hidden Bad Debt Prolongs Economic Pain – Bloomberg News

By any measure, Tom Hu should be in default on a $730,000 bank loan for his plastics business in China. He barely brings in enough revenue to pay expenses and can’t cover the debt costs.

Yet rather than calling in the loan, his bank lets him defer payments — keeping him afloat, while avoiding another past-due loan on its books…

…Stories like Hu’s are playing out across China as banks grapple with a growing pile of bad debt. It’s impossible to quantify the true extent of the problem, though most economists say the ratio of bad loans is significantly higher than the 1.5% official rate. One analyst at Absolute Strategy Research in London pegs it at about 10%, which would mean a staggering $3 trillion in loans that should be classified as past due are not. Others say it could be double that amount…

…The apparent stability of the official bad loan rate is all the more surprising given that the economy has experienced a major property collapse and posted the slowest nominal growth outside Covid since the 1970s. In March, China lowered its 2026 growth target to between 4.5% and 5% — its least ambitious goal since 1991.

Regulators have taken note. Despite seemingly strong capital buffers and stable NPL ratios, officials have moved to bolster the nation’s six biggest banks with more than $100 billion in fresh capital…

…The primary culprit for the surge in bad loans is a mountain of credit extended to companies whose earnings are insufficient to cover interest payments. About 10% of listed non-financial firms have failed to cover interest payments from their earnings before interest and tax for three consecutive years, according to Absolute Strategy Research. As a result, the non-performing loan ratio is probably closer to 10% than 1.5%, according to Adam Wolfe, an emerging markets economist at the firm…

…China’s official NPL ratio has always been a bit of a mystery. In good times and bad, it’s rarely wavered much from 1.5%, and most economists say it greatly understates the true stress in the system. The figure captures only loans officially classified as “substandard,” “doubtful,” or “loss.”

In reality, the classification is often a subjective assessment and banks have different internal criteria. A much larger pool of troubled credit remains in the “special mention” — those that may have already become overdue but yet to be categorized as nonperforming — or “normal” categories, thanks to an aggressive use of leniency known as forbearance.

Existing rules stipulate that when repayment on a loan is overdue by more than 90 days and the borrower can’t fully repay the amount, it should be marked as nonperforming.

Economists including Wolfe estimate that about 40% of loans are either eligible or already in some sort of forbearance program, where banks are strongly discouraged from seeking repayment or recognizing losses…

…In other words, rather than cracking down on deadbeat borrowers, China’s banks are encouraged to cut them some slack. Regulators have for years urged the big banks to keep their reported bad loan ratio under 2%, according to people familiar with the guidance.

With the forbearance policy — a legacy of Covid support programs that’s been extended to property developers and other firms — Beijing is signaling its desire to maintain financial stability. It wants to avoid a rash of bank failures that would follow a surge in reported bad credits and company defaults.

A leniency policy for small businesses that was introduced during the pandemic was extended in 2024 to encourage banks to roll over loans for companies enduring temporary difficulties. This policy is effective until late next year, and applies to 9.4 trillion yuan ($1.38 trillion) worth of loans, according to officials.

As a result, banks routinely roll over maturing loans, extend repayment periods, or allow interest to be capitalized to avoid triggering NPL recognition. Local governments also exert pressure on lenders to maintain stability by avoiding cuts to risk classifications on loans tied to sensitive sectors. Those include property developers, local government debt and small businesses in weaker regions, according to a dozen bankers interviewed by Bloomberg News…

…All this leniency comes at a cost. Financial resources are trapped in unprofitable and even inactive firms, hindering banks’ ability to promote growth in healthy businesses. Overall loan growth is slowing significantly after fixed-asset investment experienced an unprecedented contraction last year…

…Chinese banks are also accelerating write-offs and transfers of bad assets. Lenders have disposed of more than 3 trillion yuan of non-performing assets a year since 2020, with the total rising to roughly 3.8 trillion yuan in 2024, the highest on record.

Banks have stepped up transfers of NPL portfolios to asset management companies, which typically hoover up bad assets in China. Still, these firms entrust collection back to the originating banks in many cases, according to people familiar with the matter. The funds used to purchase bad loans largely come from the banks, meaning the risks aren’t fully removed from the financial system.

3. The Inference Shift – Ben Thompson

Specifically, coding with LLMs requires a human in the loop. It’s the human that defines what is to be coded, checks the work, commits the pull request, etc.; it’s not hard to envision a future, however, where all of this is completely handled by machines. This will apply to agentic work broadly: the true power of agents will not be that they do work for humans, but rather that they do work without human involvement at all.

This, by extension, will mean that the likely best approach to solving agentic inference will look a lot different than answer inference. The most important aspect for answer inference is token speed; the most important aspect for agentic inference, however, is memory. Agents need context, state, and history. Some of that will live as active KV cache; some will live in host memory or SSDs; much of it will live in databases, logs, embeddings, and object stores. The important point is that agentic inference will be less about GPUs answering a question and more about the memory hierarchy wrapped around a model.

Critically, this articulation of an agentic-specific memory hierarchy implies a necessary trade-off of speed for capacity. Here’s the thing, though: lower speed isn’t nearly as important a consideration if there isn’t a human in the loop. If an agent is waiting around for a job that is being run overnight, the agent doesn’t know or care about the user experience impact; what is most important is being able to accomplish a task, and if entirely new approaches to memory make that possible, then delays are fine.

Meanwhile, if delays are fine, then all of the focus on pure compute power and high-bandwidth memory seems out of place: if latency isn’t the top priority, then slower and cheaper memory — like traditional DRAM, for example — makes a lot more sense. And if the entire system is mostly waiting on memory, then chips don’t need to be as fast as the cutting edge either. This represents a profound shift in future architectures, but it also doesn’t mean that current architectures are going away:

  • Training will continue to matter, and Nvidia’s current architecture, including high-speed compute, large amounts of high-bandwidth memory, and high-speed networking, will likely continue to dominate.
  • Answer inference will be a meaningful market, albeit a relatively small one, and speed from chips like Cerebras or Groq (I explained how Nvidia is deploying Groq’s LPUs here) will be very useful.
  • Agentic inference will gradually unbundle the GPU, which alternates between stranding high-bandwidth memory (during the prefill process) and stranding compute (during the decode process), in favor of increasingly sophisticated memory hierarchies dominated by high capacity and relatively lower cost memory types, with “good enough” compute; indeed, if anything it will be the speed of CPUs for things like tool use that will matter more than the speed of GPUs…

…To date the invocation of “scaling with compute” has implicitly meant Nvidia bullishness. However, much of Nvidia’s relative advantage to date has been a function of latency: Nvidia chips have fast compute, but keeping that compute busy has required big investments in ever-expanding HBM memory and networking. If latency isn’t the key constraint, however, then Nvidia’s approach seems less worth paying a premium for…

…China, meanwhile, for all of its lack of leading edge compute, has everything it needs for agentic inference: fast-enough (but not leading-edge) GPUs, fast-enough (but not leading-edge) CPUs, DRAM, hard drives, etc. The challenge, of course, is compute for training; it’s also possible that answer inference is more important for national security, at least when it comes to military applications.

4. 50 Learnings from the War in Iran – Tomas Pueyo

Missile and drone launching can be dramatically curtailed, because you can track where they’re launched from and destroy that.

But they’re very hard to fully eliminate. This is the beginning of aerial drone warfare. It suggests it will be super important in the future as an asymmetric weapon: Countries can produce drones in a decentralized way and launch them from many different, constantly changing places.

The other way in which drones and missiles can be intercepted is at the destination. Israel has proven that this can work quite well: Iran has been unable to cause critical damage in the country despite trying over and over again…

…Iran’s entire fleet was destroyed in a matter of days (Ukraine did something similar over the last few years, virtually wiping out Russia’s fleet in the Black Sea).

This marks the end of naval warfare as we know it. Few countries will invest in a full traditional naval force anymore…

…Israel and the US blew up a lot of the command chain, but they couldn’t have done that just with airplanes. They needed intelligence, satellites, cyber penetration, AI, amazing communications, and fast command decisions. Doing all of these steps well and integrating them seamlessly is beyond the capability of most countries today…

…For the first time in history, Israel deployed an Iron Dome system in a foreign country—the UAE—manned by Israeli soldiers. This is unprecedented: Israel defending Arabs against other Muslims!…

…Iran finally executed their biggest threat, which gave them lots of leverage in negotiations: They closed the Strait of Hormuz.

It wasn’t clear that this was a threat they could actually follow through with. But it is. They closed it.

They did so even without air supremacy or a naval force. This is very counterintuitive! It turns out you can use small boats and drones to close a big international highway…

…Although US opponents have more incentives to de-dollarize, one thing is to want it and the other to succeed. The dollar has actually risen during the war, and its position as a reserve currency hasn’t changed.

5. An Ode to Restraint: Lessons from the Tim Cook Legacy! – Aswath Damodaran

If you were to create a profile of Tim Cook, the manager, based upon the choices that he has made at Apple during his tenure as CEO, two very divergent views emerge. To his admirers, his actions on some fronts (initiating dividends, massive stock buybacks, borrowing money) and inaction on other fronts (no big acquisitions, diffidence on AI investments), represent an exercise in discipline and restraint, preserving the company’s crown jewel (the iPhone) and fending off the bankers and consultants, with their false promises. To his critics, and there are quite a few, Cook’s caution has cost Apple its disruptor status, when it could have used its ample cash reserves to buy its way or invest in into almost every new business that has bloomed in the last fifteen years. In fact, they point to chances that Apple has had to buy some of the biggest stars in the market, from Tesla and Netflix more than a decade ago to Anthropic, Mistral and Perplexity in more recent years.

It is impossible to argue that one side is right and the other side wrong, but it is undeniable that both pathways (the restrained pathway that Apple adopted and the more aggressive pathway that it could have taken) include trade offs. It is true that Apple’s restraint has led it to miss out on some of the biggest trends in technology over the last decade, but it has also avoided the overpayment that is so common with high profile acquisitions of big companies. The argument that Apple would be worth a lot more today if it had bought Netflix or Tesla a decade ago falls flat for two reasons. The first is the selection bias in picking two companies that, in hindsight, have emerged as winners, when in fact there were at least a dozen other worse-performing companies that were also on Apple’s radar. The second is the presumption that companies like Tesla or Netflix would have been just as successful, owned by Apple, as they were as stand alone enterprises. The clash of corporate cultures that would have ensued if Apple had bought either Tesla, a company that reinvents its business narrative every few hours, or Netflix, an entity that makes content in quantity with the hope that some it sticks, would have been epic, with the risk that both Apple and its acquired target would have gone down in flames.

More generally, though, the question of whether you want a visionary or a disciplined business builder at the top of a firm is not one that has an easy answer, since it depends on the firm in question. In my work on corporate life cycles, I focus on the management skills that are needed most in a company, based upon where it is the life cycle, and that may help address the choice between vision and restraint:…

…With young companies, vision dominates, as managers work to sway investors, employees and nascent customers that their product or service will find a market. As the vision takes hold, converting it into commercial products and services requires trading off some portions of vision for pragmatism, in the interest of getting the business going. As products and services find demand among customers, business building becomes a key difference-maker, with the grunt work of marketing, production facilities and supply chains coming into play. Assuming that you have made it through these three stages, the trade offs of scaling up come into focus, and as you hit market limits, success depends on being opportunistic in finding new products and markets, but only if they exist. In corporate middle age, pathways to easy growth, especially at scale, become difficult to find, and to the extent that value comes from moats and core products, playing defense against competitors takes priority. Finally, in decline, a phase that no company ever wants to enter, but is inevitable at some point, you need to be willing to shrink a firm, shutting down businesses that no longer deliver value and selling other assets to high bidders.

Given these very divergent management functions, it should come as no surprise that there is no prototype for the perfect CEO, McKinsey and Harvard Business School blueprints notwithstanding.

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 Apple and Netflix. Holdings are subject to change at any time. 

More Of The Latest Thoughts From American Technology Companies On AI (2026 Q1)

A collection of quotes on artificial intelligence, or AI, from the management teams of US-listed technology companies in the 2026 Q1 earnings season.

Last week, I published The Latest Thoughts From American Technology Companies On AI (2026 Q1). In it, I shared commentary in earnings conference calls for the first quarter of 2026, from the leaders of US-listed 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 2026’s first 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:

With that, here are the latest commentary, in no particular order:

Airbnb (NASDAQ: ABNB)

AI is now writing nearly 60% of the code Airbnb’s engineers produce, 2x higher than the industry average; AI code-writing is helping Airbnb ship more features faster and deliver better experiences for guests and hosts; management thinks AI makes companies move faster; management thinks AI requires a company’s employees to be hands-on; management is seeing many of the company’s design managers and engineering managers return to coding with the help of Claude Code

Nearly 60% of the code our engineers produce is now written by AI, which we estimate is about twice the industry average. That means our teams are shipping more features and iterating more quickly. But it’s not just about speed, it’s about delivering a better experience for our guests and hosts…

…AI, I think we should think of as an accelerant to everything. And we can think of it as a disruptive technology. I actually think of it more as an accelerating technology. I think the #1 characteristic of AI is speed. It just speeds every single thing up.

I also think it makes — it requires everyone to be more hands-on and requires everyone to be more nimble and more adaptive to change. I think one of the benefits of the way Airbnb is run is that — and I think there was a term that was coined. Paul Graham, Founder Mode based on a talk I gave, but it’s really this notion that leaders should be hands on. I do not think there’s going to be as much of a role for pure people managers. Said differently, 30,000 feet hands-off managers. I think everyone is going to have to be much more hands-on, much more in the details of the company and all the data. I think now data inside a company is completely democratized. You don’t need to inquire with the data scientists to get data, we all have self-serve dashboards.

I’m seeing like many of our design managers and engineering managers going back to coding or using Claude Code.

Airbnb’s AI assistant now solves more than 40% of issues that guests face, up from 33% in 2025 Q4 and at a significantly faster pace; Airbnb’s AI assistant has helped to reduce cost per booking by 10% year-on-year in 2026 Q1; it’s really difficult to use AI for customer service; management believes that Airbnb’s 40% rate of using AI to solve customer issues is industry-leading

When guests contact us through our AI assistant, over 40% of issues are now resolved without a human agent. And this is up from about 1/3 in Q4 with significantly faster resolution time. We’ve seen the cost per booking decrease about 10% year-over-year in Q1, and we expect to see more of this as we improve AI customer support this year…

…We want to focus on the hardest problem in AI, which we thought was customer service. The reason why is the stakes are high, you have — you cannot hallucinate, you have to answer things very, very quickly because they are calling and they have problems. You have to be multilingual, often in the same conversation because sometimes guests and hosts don’t speak the same language. You have to adjudicate very difficult things. You have to escalate to human accurately, especially if it’s timely or there’s a trust and safety incident. And you have to deal with personally identifiable information that means that you have to be able to protect people’s data, you have to be able to read and train based on nearly 100 policies, tens of thousands of evolving conversations and look at like millions of data points of how a prior case was adjudicated to be able to answer correctly…

…Over 40% of people connect with our AI assistant self-solve. And I believe it’s, by far, the best AI self-solve in all of travel. I’m pretty confident of that.

Airbnb’s management thinks the ultimate search experience in Airbnb in the AI paradigm will be deep personalisation; Airbnb knows details about its users, which makes deep personalisation possible; management thinks this is similar to what all e-commerce sites will look like eventually; management’s AI strategy for search starts at the bottom of the funnel, unlike competitors; Airbnb now has AI summaries in its listings page; Airbnb is using AI for matching; management is currently testing AI search in Airbnb 

I think the ultimate, like, paradigm is not this tab versus co-mingle inventory. I believe that’s a pre-AI paradigm. I think post an AI paradigm that we’re moving towards and this relates in a second to AI search is deep personalization, understanding every user, every member. And I just want to remind everyone listening that 100% of people who booked have an account, and they have to have a verified ID. You cannot book as a guest. You have to have account, you have to be a member of the community. Therefore, we know something about you. We can infer a lot, not only about what you’re clicking on the site, but all of your past booking activities…

…We have hundreds of millions of reviews on Airbnb. And one of the things our guests told us is when they get to an Airbnb, it’s great when they see like 100 reviews, it’s awesome, but they don’t have time to read all 100 reviews. So we now have AI summaries. And AI summaries are really great. We have filters, we have AI summaries. We’re now using AI for matching. AI is really helping our search ranking and our relevance…

…Finally, it’s top of funnel, which you would call AI search. This is top of funnel. And this is what we’re currently testing.

Airbnb’s management thinks that a company needs to be really good at technology, data, and infrastructure in order to be good at AI; management has been cleaning up Airbnb’s data for the last few years to prepare for AI

When you break AI under the hood, you realize that you need — in order to be good at AI, you need to be really good at technology, foundational. You need to be really good data and infrastructure. So what we have been doing over the last few years is really getting our data warehouse really, really clean because your AI is only as good as your data.

Airbnb has an AI-native executive running its technology stack, which management believes is the only example of its kind within the travel industry

I mentioned in our last earnings call, we hired Ahmad, our CTO, who was the leader of the Meta LLaMa model. So we are probably one of the only technology companies in the world certainly only in travel that has an AI-native person running as the entire technology stack.

Airbnb’s management is currently experimenting with the best ways to implement AI in the business; management thinks nobody has figured out AI for travel e-commerce yet, even though ChatGPT traffic converts on Airbnb at a higher rate than Google traffic, for 5 reasons, namely (1) travel e-commerce is photo-forward, whereas AI chatbots are text-based, (2) chatbots do not allow users to directly manipulate search results, (3) chatbots do not allow easy comparison between a wide variety of options, (4) chatbots are not multiplayer, and (5) chatbots are not map-native; management thinks AI is a risk to Airbnb, but it’s also an opportunity; management foresees a lot of AI-focused innovation from Airbnb in 2027

We are essentially piloting a variety of different ways to use AI, whether it’s in the search box, whether it’s once you search, interrupting on the search, it’s the filter panel, once you book a trip. So we’re trying a lot of different things. We’re really in the exploration, research development mode…

…I don’t think anyone figured out AI for travel or e-commerce yet. Let me use an example, ChatGPT. Last year, ChatGPT announced the creation or of third-party apps. And then this past March, they shut that project down. And one of the things we noticed is that while ChatGPT is — traffic converts higher than Google traffic when it’s sent to Airbnb, we think the design of a chatbot fundamentally as its currently constructed today does not work for travel e-commerce. There’s essentially four problems.

The first problem with the chatbot is there’s too much text. Chatbot are LLMs, large language models. They’re language. And most of e-commerce is not language forward, it’s photo forward. That’s the first problem. The second is there’s no direct manipulation. You can’t touch anything. You have to type everything. And that’s great for a conversation. But if you want to like move the price slider, that’s much easier than type, well show me X, Y and Z. The third problem is comparison. You go to Airbnb in Paris, there’s tens of thousands of homes, I think over 100,000 homes. Imagine trying to compare 100,000 homes in a chat bot, you get lost. And so it wants to show you just three options. You want to see more than three and pretty soon you get confused in a thread. And the fourth problem is that almost all bookings of Airbnb have multiple guests, what we call multiplayer. Chat bots are primarily single player. This doesn’t account for the fact that 85% of people booking Airbnb send a message, 100% have an account. And also chat bots are not map-native…

…AI is a risk to us and everyone. If it’s a risk to us, it’s a risk to everyone. So risk to everyone is an opportunity for us…

…I believe that over the next year, you can see a lot of innovation around AI search, AI-native interfaces.

Airbnb’s management buckets its alternative accommodations supply into 2 buckets, namely, the API (application programming interface) bucket, and the primary homes and vacation homes bucket; for the API bucket, management thinks AI enables Airbnb to build more tools to serve hosts; management thinks Airbnb has been lagging behind 3rd parties in building great tools for the API bucket; hosts within the API bucket have sounded out to Airbnb that they need better tools to manage their businesses, and Airbnb has struggled in the past for resources to build these tools, but now the company has a productivity-boost from AI in software development and so are able start building the tools; for the primary homes and vacation homes bucket, management thinks AI can make it much easier for primary homes hosts to list their properties

You can think about our core accommodations business of homes as a few different categories. So you have essentially hosts that connect via an API. You might call that host API partners. These are primarily property managers. That’s one category. Then we have primary homes, homes that people live in primarily, so typically more than 180 days a year. Then you have vacation homes, then you have things like private rooms. So you have to think about each. And I would break them into two, the API and the primary homes or vacation homes. These are two buckets.

I think with the host API partners, I think it’s more about AI enabling us to build more tools. I think we’ve been a little bit lagging behind third parties and building great tools for host API partners. And as a segment, the host API hosts are growing really, really fast, and we see a really big opportunity to better serve them. One of the things we found is that the more properties you manage at Airbnb, the lower your rating is. And so said differently, our customers have higher satisfaction with individual hosts over property managers. Now on the one hand, that’s encouraging because that inventory is more unique and exclusive to Airbnb. Other hand, we see that as opportunity. And one of the things those API partners say is, well, we want to be better host, but we need better tools. So AI is a like — maybe here’s an analogy. In the old world, you might need a team of 20 engineers. In a new world, an engineer can spin up 10 agents. And those agents can work 24/7. I mean I’m kind of exaggerating a little bit. You have to be there to prompt them and the amount of work they can do without supervision isn’t overnight, typically for most tasks, but you can see a huge amount of leverage. So the fact that we’re adopting AI tools is a way for us to get a lot more leverage around the software for most API partners…

…Originally, we didn’t have the resources to do all of the host API work we want to do. And now with AI, we’re reevaluating how much productivity we have, and we’re able to accelerate the development of this work…

…AI, especially though, can help the sourcing discovery in the listing of primary homes. So without, again, giving away some of the things we’ll show in 20 — May 20, we do find that AI can make it much easier to list your property. So right now, you have to type everything in, you type in your address, you type in your title, you have to type in your listed description. Eventually, I imagine a world where you can just say like, list my place, you put in your address, it can scrape information on the Internet. You can take photos. It can even write your description based on computer visioning of the photo. So it’s very, very difficult for a regular person to list a property.

Airbnb’s management thinks that AI agents still cannot work for long hours in an unsupervised manner

So AI is a like — maybe here’s an analogy. In the old world, you might need a team of 20 engineers. In a new world, an engineer can spin up 10 agents. And those agents can work 24/7. I mean I’m kind of exaggerating a little bit. You have to be there to prompt them and the amount of work they can do without supervision isn’t overnight, typically for most tasks.

Arista Networks (NYSE: ANET)

Arista Networks’ management sees AI workflow patterns as being different from typical cloud computing workflows; AI workflows have 2 main categories, namely, long-lived massive flows, and short-lived, unpredictable flows; the difference between AI workflows and typical cloud workflows mean the performance of a flow is important

Unlike typical workloads, AI workflow patterns can be long-lived elephant flows or short-lived and simply not predictable. This implies careful attention to performance where a flow can cause burstiness for a long duration of milliseconds. The intensity of a flow can determine the line weight throughput, the shifting traffic patterns to massive flows synchronized to all-in-all or all-reduce or burst with collective communication are all important for AI training and inference applications.

In the scale-up AI networking use case, Arista Networks’ management sees ESUN (Ethernet for Scale-Up Networking) paving the way for Ethernet technologies to increase and decrease computing power flexibly to match workload demands; Arista Networks will be entering the scale-up networking business in 2027; Arista Networks will be working with its customers to build AI racks with rapid interconnects for CPC (co-packaged copper) and CPO (co-packaged optics); management has no doubt that Arista Networks will have a number of scale-up use cases in 2027 and most of them will start with 1.6 terabit switches; the scale-up use cases in 2027 include 5-7 rack opportunities that Arista Networks is actively designing with customers; today’s scale-up AI networking products are mostly from NVIDIA’s NVLink and PCIe; CPOs are very much still science experiments in the eyes of Arista Networks’ management; management thinks scale-up racks would not be possible with XPO 

In scale-up mode, we have familiar technologies such as NVLink and PCIe that have enabled vertical scaling of single compute nodes or racks. The advent of ESUN, Ethernet for Scale-Up Networking, specifications allows for increasing or decreasing computing power in a flexible manner with Ethernet to automatically adapt to workload demands. Scale-Up will be a new entry for Arista in 2027 and beyond, where we will be working closely with our customers to build AI racks with very fast interconnects for co-packaged copper, CPC, or open co-packaged optics, CPO, as well as supporting collectives and memory acceleration…

…there is no doubt in our minds that we will have a number of racks and number of scale-up use cases in 2027. Maybe some of them will be in early trials, but majority of them are looking at really starting with 1.6T, and 1.6T chips will really happen in 2027. There may be a few, a handful of them that tried some experimental stuff at 800 gig. But we continue to see at least 5 to 7 rack opportunities. Some of them are multiple racks with the same customer. We’re actively designing with them. There’s a huge amount of liquid cooling designs with very dense cabling options, acceleration of collectives and memory, features we have to work on for low latency. So I definitely feel we’re in active engineering phase with Ken and Hugh’s teams this year. But unlike the ODMs, I think we’re held to a higher bar, and we have to just make sure that this thing is production worthy and specification adhering to ESUN. So I would say today’s scale-up is mostly limited to NVLink from NVIDIA and maybe some PCIe switching. But majority of the Ethernet scale-up will only really happen in ’27 and ’28…

…While the industry has been talking a lot about co-packaged optics, these are still science experiments, and they’re very proprietary with individual vendors doing their own thing…

…We embrace open CPO a few years from now, but we think XPO has a 10-year run, especially at 1.6T and 3.2T where you need liquid cooling and you need that kind of capacity. So all the scale-up racks we’re talking about wouldn’t be possible without XPO or CPC or any one of those technologies.

In the scale-out AI networking use case, Arista Networks already has more than 100 cumulative customers to-date in 800 gigabit Ethernet deployments; management expects to see 1.6 terabit Ethernet solutions in 2027 at production scale

Scale-out or horizontal scaling involves adding more machines to a leaf-spine fabric, moving workloads across multiple servers or nodes or even connecting other elements like storage or CPUs. As you scale up or out with massive data sets, bottlenecks can be resolved with collective and protocol acceleration at L2, L3, cluster load balancing, all at wire rate. The system must deliver consistent performance without degradation as more nodes participate. Arista is a shining example here with greater than 100 cumulative customers to date in 800 gigabit Ethernet deployments, and we expect the addition of 1.6 terabit in 2027 at production scale.

In the scale-across AI networking use case, Arista Networks’ management thinks the company’s 7800 R3 and R4 series of products, which provides sophisticated traffic engineering, deep routing, encryption properties, and integrated optics atop its EOS (Extensible Operating System) stack, are a great solution; management sees the 7800 series as the premier scale-across product; scale-across AI networking was only a small part of Arista Networks’ business in 2025, but will contribute at least 1/3 of the company’s $3.5 billion in AI networking revenue in 2026; the presence of Alphabet’s TPUs and AMD’s GPUs has created a huge opportunity for Arista Networks in scale-across AI networking; management thinks scale-across is the most significant and differentiated opportunity in AI networking for Arista Networks

Scale across — drives across the cloud and AI as the AI accelerators in a location may need to be distributed to achieve the appropriate bandwidth capacity with the optimal power. As workloads become more complex and more distributed, the bi-sectional bandwidth must scale smoothly to avoid bottlenecks and preserve performance. This demands sophisticated traffic engineering, deep routing, encryption properties, and integrated optics based on Arista EOS stack, and using Arista’s flagship 7800R3 or R4 series. The 7800 has established itself in this category as the premier scale across choice…

…I think last year, on scale-across, we were just beginning. So I think they were small numbers. And majority of the numbers were really scale-out. That’s sort of our heritage and that’s where we excel. If I were to anticipate how it would be this year, again, scale-up is virtually 0 and nonexistent because it really only comes to play after the ESUN spec. So consider that more a’27, ’28 kind of number. So I think the number will be really shared between scale-across and scale-out. I don’t know if I can say it’s 50-50 or 70-30 or 60-40, but scale-across will definitely contribute at least 1/3 of our AI number…

…In general, we are seeing diverse accelerators. Last time I spoke about the AMD accelerators. This time, I will definitely give a nod to the TPUs because in particularly scale across use cases, we’re seeing multitenants connecting to different AI accelerators, including TPUs as well. So I think the diversity of accelerators is creating tremendous multiaccelerator opportunity and multiprotocol features that we can provide for them in our network…

…Scale-across is by far the most significant and differentiated opportunity that really highlights Arista’s prowess in both platforms and software.

Arista Networks’ management thinks the company’s Etherlink portfolio handles both massive synchronous flows for AI training, and low latency flows for real-time inference

Arista’s Etherlink portfolio addresses both the synchronous flows for massive training and the low latency for concurrent swarms of real-time inference in this era of trillions of tokens, terabits of performance, and terawatts of power.

Of Arista Networks’ 4 major AI customers that are deploying AI with Ethernet, 3 had deployed 100,000 GPUs each with Ethernet as of 2025 Q4; the last remaining customer has migrated from Infiniband to Ethernet at production scale; since 2024, Arista Networks has expanded to many more customers beyond the 4 major ones

In 2024, you may recall, we discussed 4 Ethernet-based AI training deployments. And of course, since then, we’ve expanded and exploded to countless others. This fourth customer from the group has officially moved from InfiniBand to Ethernet at production scale over the last 2 years.

Arista Networks’ management thinks the high-speed Ethernet AI leaf-spine architecture, with flexible air or liquid cooling, can overcome the constraints of power and space for AI workloads; management thinks the architecture can help build a low latency distributed AI supercomputer fabric globally

The high-speed Ethernet AI leaf-spine with flexible air or liquid cooled infrastructure overcomes the physical constraints of power and space for AI workloads. It results in a low latency distributed AI supercomputer fabric across global regions.

Arista Networks’ management recently introduced its extended pluggable optics, the XPO form factor; management thinks the company’s networking progress has been important for high-speed optics transmission; the XPO form factor is now endorsed by more than 100 vendors and delivers a record-breaking 12.8 terabits of throughput per pluggable module, and unprecedented rack density, among other traits; management thinks XPO will have a 10-year run; management thinks scale-up racks would not be possible with XPO; management thinks XPO is a very important innovation for the industry; management sees XO unlocking a standard multivendor way to obtain 4x the network density in liquid cooling, which is critical for AI use cases; management thinks XPO and OSFP (Octal Small Form-factor Pluggable) are partnering technologies, where XPO is more suitable for higher data speeds; management thinks XPO will be more suitable for scale-out and scale-across workloads compared to scale-up

What is clear to me and us is our networking progress with data, control and management, and multiplanar orchestration is not only central to our AI switching performance, but also important for high-speed optics transmission. At the recent Optical Fiber Conference, Arista unveiled its extended pluggable optics, XPO form factor, designed specifically for optics innovations at high speed. Now endorsed by greater than 100 vendors, salient features include record-breaking throughput, delivering 12.8 terabits per pluggable module, unprecedented rack density achieving 204.8 terabits per OCP rack unit, integrated cold plate capable of cooling up to 400 watts power per module, and the universality and flexibility across a range of pluggable optics, copper as well as linear halftime or retimed interfaces…

…We embrace open CPO a few years from now, but we think XPO has a 10-year run, especially at 1.6T and 3.2T where you need liquid cooling and you need that kind of capacity. So all the scale-up racks we’re talking about wouldn’t be possible without XPO or CPC or any one of those technologies…

…99% of the optical market today that we connect to is all pluggable optics. So this is a very crucial invention and innovation, not just for Arista, but the industry at large…

…What XPO unlocks is a standard, interoperable multivendor way to get to 4x the network density in liquid cooling, which is absolutely critical for these AI use cases. Without that, you’ve got this huge bottleneck at the front panel, the amount of extra rack space is required to get through OSFPs. It’s — so we’re really enabling the future growth of our industry this way, which we benefit and others benefit as well…

…You should look at XPO as a partner to OSFP. So at 400 gig and 800 gig you’ll be fine with OSFP. And as we go to higher speeds in ’27, ’28 or even beyond, OSFP will run out of steam, and this will be the new connector of choice. So the migration to higher speeds equals the migration to XPO, particularly for scale-out and scale-across. Within a rack and scale-up, there’s still a number of choices. I think within short distances of 2 to 3 meters, you’re still going to see a lot of co-packaged copper and I think XPO in terms of density will be another alternative. But I don’t rule out open CPO as well over there. They’re really looking to maximize the density in a minimum amount of space. So I think XPO will be particularly prevalent in scale-out and scale-across and will be one of the choices in scale-up.

Arista Networks recently won a neocloud as a customer for AI networking; the neocloud’s initial white box architecture could not handle massive scale-out requirements; Arista Networks was selected by the neocloud for its scale-out architecture, which could connect with AMD XPUs; the neocloud is also using AVD (Arista’s Validated Design framework) to automate networking provisioning and thus lower the total cost of ownership; Arista Networks’ management is seeing tremendous opportunity with neocloud and sovereign cloud customers; management thinks the neoclouds are a very important sector for AI networking because they do not have the resources to tackle networking, and so will rely on vendors such as Arista Networks

Our first highlighted win is a neocloud AI network. The customer was constrained by an incumbent white box architecture that simply could not keep pace with the massive scale-out requirements of AI. Arista was selected as a commercially proven and reliable scale-out architecture with unmatched stability of EOS and the ability to connect AMD MI Series XPUs. Arista’s AI leaf and spine Etherlink products were deployed at 800 gigabits to provide the incredible performance modern AI networks require. The AI fabric was tuned using Arista’s cluster load balancing to scale out to thousands of XPUs minimizing hotspots and congestion. On the software side, the customer leveraged AVD, Arista’s Validated Design framework, to automate network provisioning, which both reduces the total cost of ownership, but also provides an easy path to reliable network deployment at scale, where without AVD automation, a small mistake can cause precious days of debugging time. This was a strategic neocloud win with large potential for upside growth in an area where we are seeing enormous opportunity and velocity in both neocloud and sovereign cloud customers…

…It’s easy to talk about the titans because the numbers are so ginormous, right? But the neoclouds are a very important sector because they don’t always have the staff to do everything they want to do, and they really lean on Arista’s design expertise, EOS expertise, network design configurations we can provide them, a family of 22 products we have in AI. 

Arista Networks’ management is seeing industry-wide supply shortages across the silicon board, which has led to higher supply costs and thus gross margin pressure; demand for Arista Networks’ networking products is outstripping supply; management hopes the supply shortages will ease in 1-2 years; despite the supply chain challenges, management has raised guidance for Arista Networks’ AI networking revenue for 2026 to $3.5 billion (previous guidance was for $3.26 billion); Arista Networks’ purchase commitments at the end of 2026 Q1 was $8.9 billion, up 31% sequentially; the sequential increase in purchase commitments was for chips related to new products and AI deployments; management is willing to hurt Arista Networks’ gross margin in order to meet demand for AI networking; management is seeing shortage of power in data center sites; management has chosen not to raise prices, which explains the gross margin pressure; Arista Networks’ purchase commitments extend to multiple years because the lead times for chips are that long

Our demand is actually the best I’ve ever seen in my Arista tenure. The supply, however, is a slightly different and opposite tale. We are experiencing industry-wide shortages across the board, be it wafers, silicon chips, CPUs, optics, and of course, memory that I referred to last quarter, coupled with elevated costs to procure these. Clearly, our demand is outstripping our supply this year. While we hope the supply chain will ease in the next year or 2, the Arista operations team has been diligently engaging with our vendors in strengthening supply agreements and engaging in multiyear purchase commitments. We anticipate gross margin pressure due to mix and trade-offs we are making to pay more to assure supply continuity to our customers. Nevertheless, it gives us confidence to increase our forecasted growth slightly to 27.7%, aiming now for $11.5 billion for 2026. We also increased our AI target now to $3.5 billion this year, thereby more than doubling our AI sales annually…

…Our purchase commitments at the end of the quarter were $8.9 billion, up from $6.8 billion at the end of Q4. As mentioned in prior quarters, this expected activity mostly represents purchases for chips related to new products and AI deployments…

…We see multiyear demand, and we are going to do everything, including hurt our gross margins to supply to that demand this year and next year because we believe that we certainly don’t want to keep GPUs idle and AI infrastructures underutilized because Arista didn’t supply the network…

…The other thing we’re seeing with a lot of these use cases is the lack of power in sites, and the ability and demand to distribute and get a more multitenant scale-across is very high in these 2 use cases…

….One thing to clarify also on gross margins. So we view this as a partnership with our customers. So while we would consider and have raised prices a little bit, unlike our competitors, we haven’t done 2 price increases. We haven’t done major price increases. And the price increases really come into play once our backlog starts to reduce, right? So you won’t see the impact of that. So our gross margins are a strong factor of cost going up and are still eating a lot of the costs and giving our customers the benefit and promise of the pricing we said we would give to them…

…I would just say our purchase commitments are multiyears because we’re having to deal with forecasts that are out multiple years so that we get them in time because the lead time of these chips is so long. So I think that’s the biggest hole, lead times.

Arista Networks continues to have a great relationship with its 2 largest customers, Microsoft and Meta Platforms, in both cloud and AI; management sees the potential for 1-2 new large customers for Arista Networks that use all 3 AI networking use cases – scale-up, scale-out, and scale-across

Microsoft and Meta, they’re our all-time favorites. They’ve been our 10% and greater customers for over a decade. And the partnership could never be stronger, and it continues to get better both in cloud and in AI. In terms of the new entrants, we still expect at least 1, maybe 2 — and maybe I should caveat this by saying, certainly, in demand, we see 1 or 2. We shall see, Todd, how we do on shipments to see if we can achieve the greater than 10%. The 2 of them have very interesting characteristics. They exhibit what I would call the 3 use cases I just alluded to, scale-up, scale-out and scale-across where we really have a fabric notion of creating — so far, we’ve been working with them a lot on the front end, and now we get to complement that on the back end, definitely for scale-out and scale-across and maybe even a little bit of scale-up in some of these use cases.

The biggest use case that Arista Networks’ management sees right now in agentic AI is training, but it will move to distributed inference; management thinks agentic AI will be moving into plenty of enterprise use cases; agentic AI has caused Arista Networks to see a lot more back-end activity now because the hyperscalers have to deal with billions of parameters and tokens, to the extent that the hyperscalers are ignoring the front end refresh; the rise of agentic AI has changed management’s view on the ratio of front-end deployments to back-end deployments from 2:1 to 1:1 or even less; Arista Networks has the same set of products in the same common operating system  across the front-end and back-end, which management sees as lowering costs for customers; Arista Networks is the only vendor that has the same set of products in the same common operating system across the front-end and back-end

The biggest killer application we see in agentic AI right now is still training. And indeed, it’s going to move to more distributed inference. And we’d also like to see agentic AI move into a lot of enterprise use cases, all of which we’re seeing, by the way, but I would say large, medium, small. The largest killer agentic AI application is training, the medium is enterprise and the smallest — medium is inference, and the small is obviously enterprise. The — in terms of back end versus front end, we are now seeing way more back-end activity, particularly with our large AI titans and cloud titans because there is just so much scale they need to prepare for the billions of parameters and tokens, and this is where a lot of — so much so that I think the front end, they might come back and refresh, but they’re almost ignoring right now in favor of the back end…

…By virtue of the back-end deployments, I don’t know if we any more see a 2:1 to the front end, but we at least see a 1:1. And the 1:1 can be wide area, CPU, and storage. Those are probably the 3 common use cases. Not all the customers are up and lifting everything and doing all 3, although we’ve had cases where some of them did an upgrade at the front end before they went into the back end. But usually, they will have to come back to that because the minute you put that kind of performance pressure and scale on the back end, you almost have to do something in the front end. But at the moment, I would say it’s more one-to-one…

…The other thing I have to mention here is just how good it feels to be — have the same set of products in the same common operating system management suite and operating model across the front end and back end. This lowers cost for the customer, simplifies their design process to get that leverage, and we’re one of the few vendors who can do that…

…I think only.

When it comes to greenfield deployments of AI data centers, Arista Networks’ management has observed that customers think of both scale-out and scale-across solutions concurrently; Arista Networks has strong market share in both scale-out and scale-across in greenfield deployments; when it comes to brownfield deployments of AI data centers, Arista Networks now has the opportunity to offer scale-across solutions; the lack of power supply has resulted in data center operators having to distribute the centers, which gives Arista Networks the opportunity to participate in the build out

[Question] You said most of the cloud revenue near-term is going to be scale-out and scale-across as we wait for scale-up to ramp. How are you thinking about your market share when it comes to scale-out versus scale-across in the early days of scale-across? What are you seeing in terms of market share? And are you seeing customer decisions being led in scale-across by sort of the incumbent in scale-out? Or is it a different decision altogether in terms of how they’re designing vendors for scale-across?

[Answer] If it’s greenfield deployment, then they tend to think of it together because they’re not only building the sites, but they’re thinking of the interconnect across them. And therefore, market share is generally strong in both. In some cases, where Arista has not been a historical participant within the data center, we now have an opportunity to offer the scale-across multitenant even in a nongreenfield situation and let’s say, in a brownfield, where now they’ve got disparate data centers or AI clusters that we now have to bring in. And so once again, I think Arista is really fitting example to be in scale-across for both those use cases, but has the additional opportunity in a brand-new data center to be in all use cases, if that makes sense. So it’s giving us a chance to participate with different types of accelerators and different types of models because people aren’t getting the power and they’re having to distribute the data centers. And as a result of distribution, you need more traffic engineering, routing, multitenancy. So I would say scale-across is the common denominator in all our use cases and scale-up and scale-out maybe nice options in brand-new greenfields.

Arista Networks’ management currently sees AI training workloads dominating, but they also see an inference paradigm coming, where CPUs will become more important than GPUs; management is seeing customers wanting to deploy small-ish clusters, in the thousands of GPUs, for inference

While today we are in a training fever, that a more distributed AI — generative AI paradigm with inferences, which means you don’t always need the GPU. You’re going to have high-end CPUs and you’re going to have a smaller set of parameters and tokens to manage, and you’re going to have specific agentic AI use cases and applications. We’re seeing very, very early trials and stages. Nothing super big yet. But we are seeing — I mean, they’re not in the hundreds of thousands of GPUs like you see on the AI titans. But we are frequently seeing our customers in certain high-tech sectors want to deploy clusters that are 1,000 — few thousand, definitely not 10,000, but in hundreds of thousands. And they tend to be exactly, as you said, not training, but more inference based — more agentic AI edge inference based as well. So I think we’ll see more of that. This is the calm before the storm, if you will. And as we — as the AI gets more distributed, I think it doesn’t need GPUs alone, it’s going to need more high-performance compute.

Cloudflare (NYSE: NET)

A rapidly-growing technology company in the Asia Pacific region is experiencing explosive growth, driven by AI coding, and expanded its relationship with Cloudflare; the technology company chose Cloudflare over a hyperscaler

A rapidly growing technology company in APAC expanded their relationship with Cloudflare, signing a two-year $8.7 million contract for application services and our Workers developer platform. Driven by the boom in AI-powered live coding, this company has seen explosive growth, and Cloudflare has become core to their infrastructure, intelligently routing billions of daily requests across the globe. This customer chose Cloudflare over a competitive bid from a hyperscaler due to the strength of our unified platform and our seamless low-latency security. 

A Fortune 100 technology company expanded its relationship with Cloudflare after facing an urgent need to handle massive user-initiated agentic traffic; the technology company was up and running with Cloudflare within a week

A Fortune 100 technology company expanded their relationship with Cloudflare, signing a two-year $8 million contract for our privacy proxy solution, the fifth privacy engagement with this customer, solidifying Cloudflare as their go-to privacy partner. They approached us with an urgent need to handle massive scale with precise geolocation accuracy for user-initiated agentic traffic. We delivered a fully operational solution within one week, demonstrating the speed, trust and engineering depth that continues to set us apart.

A leading AI company expanded its relationship with Cloudflare despite having a strong build-over–buy mentality; the AI company is a massive target for cyberattacks and needed a strong security layer to protect its infrastructure; the AI company is already testing Cloudflare’s AI gateway for AI workloads

A leading AI company expanded their relationship with Cloudflare, signing a one-year $4.1 million contract for application services. As one of the most visible targets for cyberattacks globally, this customer needed a security layer to protect their massive infrastructure build-out. Despite a strong build-over-buy mentality, they chose Cloudflare, trusting a battle-tested network that has proven its resilience against the largest attacks. This is a customer that moves fast and pushes boundaries, and they’re already testing our AI gateway for their AI workloads.

A leading AI company expanded its relationship with Cloudflare with a contract for Argo Smart Routing just one quarter after inking a Workers Developer Platform deal; the AI company used Cloudflare to lower its average global latency by 30%; the hyperscalers could not match Cloudflare’s speed

Another leading AI company expanded their relationship with Cloudflare, signing a 10-month $2 million contract for Argo Smart Routing coming just one quarter after signing a Workers developer platform deal. This customer wants to be the fastest and most reliable AI provider in the market, and Cloudflare is delivering. After deploying Argo, they immediately reduced their average global latency by 30%. In the AI space, that kind of speed is a real advantage that our hyperscaler competitors simply can’t match.

Cloudflare’s management is seeing agentic AI reshape how companies are structured, operate, and create value; Cloudflare itself is the first and most demanding customer of its own AI tools; prior to November 2025, management was cautious about deploying AI internally because management was unclear about the ROI from AI investments; from November 2025 onwards, Cloudflare started experiencing massive gains in productivity from the use of AI; in 2026 Q1, Cloudflare’s usage of AI increased by 600%; nearly all of Cloudflare’s R&D team are using AI coding tools powered by the company’s Workers Developer Platform; 100% of the code submitted by Cloudflare’s R&D team for production are now reviewed by autonomous AI agents; management thinks there will soon be a huge uptick in reliability in software development across the technology industry because AI can now be used to check code; in 2026 Q1, Cloudflare experienced an unprecedented increase in new code generated, solved bugs, and burn-down of technical backlog; employees across Cloudflare are running thousands of AI sessions daily, and these workflows rely on dozens of MCP (Model Context Protocol) servers; Cloudflare has built an agentic harness called Cloudflare OS for teams to get started quickly with agentic AI; Cloudflare’s newfound productivity from AI use has led management to reduce headcount by 20%, but growth is expected in the company’s sales team; Cloudflare has been able to keep the costs of internal AI deployment manageable by running the models on its own infrastructure when appropriate, instead of the model providers’ infrastructure; Cloudflare has been able to achieve significantly higher utilisation of GPU resources than hyperscalers and AI labs; Cloudflare’s AI Gateway enables it to route workloads to the right models, thereby achieving cost-efficency 

In nearly every customer conversation, it’s clear. The emergence of generative and agentic AI is not just redefining the economics of the Internet and software companies, they’re redefining the business models of all companies, fundamentally reshaping how organizations are structured, operate and create value.

At Cloudflare, we don’t just build and sell AI tools and platforms. We are our own most demanding customer. AI and agents are no longer pilot projects at Cloudflare. They are now core parts of our workforce. It’s been an interesting journey. We’ve been selling picks and shovels in the AI gold rush for the last four years, but we ourselves were cautious users wanting to ensure there was real ROI before making significant investment. We avoided a lot of the performative AI some companies engaged in. Internally, the tipping point was last November. At that point, across our teams, we began to see massive productivity gains, team members that were 2x, 10x, even 100x more productive than they had been before. It was like going from a manual to an electric screwdriver. Cloudflare’s usage of AI has increased by more than 600% in the last three months alone. For team members in R&D, 97% use AI coding tools powered by the same Workers Developer Platform we ship to our customers and 100% of their contributions to our production code bases are now reviewed by autonomous AI agents.

I think across the industry, you’re about to see a massive uptick in reliability as every code or configuration change can now have a tireless and uncorrelated set of eyes trained on every incident from the last 10 years, checking to avoid problems. At the same time, the impact on developer velocity is clear. We’ve never seen a quarter-to-quarter increase in new code generated, bugs squashed and technical backlog burn down like we did last quarter…

…Employees across Cloudflare from HR to marketing run thousands of AI sessions each day to get their work done. Those agentic workflows rely on dozens of MCP servers to reach data in systems of record and use hundreds of centrally managed skill files as well as many more that have been created and shared within individual teams. The harness that we’ve built, which we call Cloudflare OS allows teams across the company to quickly get up and running…

…By fully embracing an agentic AI-first organizational structure and operating model as Cloudflare’s revenue scales, our efficiency and productivity will scale even faster. Unfortunately, this decision means parting ways with colleagues who have helped build the strong foundation Cloudflare stands on today, resulting in a reduction of the size of our team by approximately 20%. These reductions are across all functions and geographies and reflect how broadly AI is accelerating our operational velocity. Importantly, however, we continue to expect growth in the net capacity of our quota-carrying sales force to accelerate in 2026 with today’s actions compounding productivity to fuel our growth…

…[Question] How do you think about balancing R&D agentic coding adoption with the cost?

[Answer] We have seen as usage has gone up 600% in the last quarter, we have seen costs go up. But I don’t think it’s gone up nearly as much as some others. And that’s driven by a number of things… more importantly, though, is a lot of times, we’re able to run those models instead of on their infrastructure, on our own infrastructure. And so we have a fleet of GPUs, and we have all of the tools with Cloudflare Workers and Workers AI to be able to build and use those tools themselves. And so most of the use of various AI coding tools isn’t even leaving our network. It’s running on our infrastructure because we’re very good at routing to wherever there’s capacity, we’re able to get a lot out of that. And so I think that’s one of the reasons why we see significantly higher utilization across our GPU resources than some of — than any of the hyperscalers and then — than any even of the AI labs are able to drive. 

And then when we’ve built what we call Cloudflare OS, we’ve paired that with our AI Gateway product. And that AI Gateway product allows you to route different requests based on what’s the right model for the right task. And so that means that if we have a task which we can evaluate as being relatively simple, then we can route that to a model that might be running on our own infrastructure and be able to be delivered at essentially no marginal cost to us. Whereas if we have something that is more important, we might send that off to one of the frontier models and pay more for that…

…Across most of the hyperscalers, you’re seeing utilization rates of their GPUs that are in the single digits, whereas we’re slowly getting our GPU utilization to approach what our CPU utilization is, which is up in the 70% to 80% range.

Cloudflare’s management thinks AI is the biggest tailwind for the company’s network and Workers Developer Platform in its history; management thinks Cloudflare got lucky by already having the right set of tools for agentic AI 

AI is driving a fundamental replatforming of the Internet as well as a paradigm shift in how software is created and consumed, and it’s shaping up to be the biggest tailwind for both our network and our Workers Developer Platform that we’ve ever seen in Cloudflare’s history…

…In our workers platform, we have built a platform that allows you to build agents that are just significantly more efficient than anyone has before. And so across all of the parts of our business because even in the Zero Trust and SASE space, it turns out that having more fine-grained controls about data is exactly what you need if you have kind of these somewhat new agents running around doing things, you want to make sure that they only have access to the things they should. It’s — I wish I could say that we saw all this years ago and built Cloudflare for it. But I think that the reality is that we happen to have built exactly the right set of tools for this moment.

Cloudflare’s management is seeing hundreds of billions of agentic requests monthly, and the requests are growing

So today, literally, we’re seeing hundreds of billions of agentic requests per month, and that number is growing exponentially.

Cloudflare’s management thinks the predominant business model of the internet will be changing dramatically over the next 5 years because of AI, but the end-state is still an open question; management thinks Cloudflare could help define the new business model(s) for the internet; management thinks micro transactions for agentic traffic to websites will be one of the new business models, because agentic internet traffic could surpass human internet traffic in 2027; management thinks that nobody currently has the appropriate infrastructure to handle the potential volume of agentic micro transactions; because of unwanted agentic traffic on advertising-supported media websites, Cloudflare has gone from low penetration in the space to dominating it; media companies have been able to sign better deals with AI companies because of the tools Cloudflare has built; management is focused on making substantial progress with the internet’s new business models, but they are unsure when these will become meaningful 

The business model of the Internet, which has historically been advertising and subscriptions, is about to change dramatically over the next five years. And exactly what it changes to I think it’s still an open question. And I think it might not be one thing. I think it might be several things. Because of how much of the Internet sits behind Cloudflare, we have a seat at the table of defining that…

…Some part of this is going to be some kind of micro transactions for any request that agents are making to website. It might be fractions of fractions of pennies. But if you think about the — I don’t know, about 500 billion requests that pass through Cloudflare in any given second, that some percentage of those we think that there’s going to be some ability to have some micro payment that is made for that because somebody has to pay for the infrastructure. And if you look at the growth in agentic traffic, if you look at the growth in sort of non-human traffic on the Internet, somewhere in 2027, we think it’s going to surpass human traffic, and it’s not going to slow down. And so we’ve got to figure out something else to build it…

…The challenge is like nobody can handle the volumes right now. And so we’re looking around to partner with people. We’re looking around for everything. But right now, the sort of transaction volumes that people are excited about like one million transactions per second, we need something that’s significantly larger than that…

…If you’re an ad-supported business, then your content being crawled is actually a threat. So I think we’re trying to provide tools on both sides of that. The side that you focused on is the folks that want to block it, the ad-supported folks that are out there. And I would say that the first milestone that we’ve seen is that we went from being relatively low in terms of our penetration in the media space to today dominating that space. And so I think that’s the first sign. And what I hear from media company execs is they are signing better deals with AI companies because we’ve given them the tools to be able to control who has their content…

…I don’t know exactly when that will come. But I do — I will say that when we listed what our top six priorities were for 2026, one of the six was making sure that we make real progress and see the first revenue that we can then pass back to that long tail of the Internet in order to help make sure that we continue to create a healthy ecosystem for content creators. And I’m pretty confident we’ll make that goal.

Cloudflare’s management thinks the company’s business is very different from that of the hyperscalers when it comes to providing AI compute infrastructure

The hyperscalers business is to buy a server and then to lease that server back ideally for 5x or more of what they paid for it. And so if they don’t have servers to lease, then they can’t grow their revenue. And so their CapEx has to invest ahead of whatever that demand is that’s out there. We focus on very different things. So the thing to watch for us is when you see us publish a blog post about how we figured out how to get more utilization across our fleet of GPUs or how to get more models loaded quickly across GPUs. That’s real IP that we’re inventing internally and the metaphor to think about is once upon a time when I was in college, I remember a new thing called the web was starting, and so we needed to have a web server. And so we literally — from Gateway, I remember ordering a box that came with cow prints on the outside of it. We bought a gateway server and we plugged it in because there was no idea of virtualization. And then VMware came along and then after that, you had Docker and containers and that was sort of the journey that everyone went on. We’re still at the stage with GPUs of buying the physical server and needing to use that for most of the industry.

Cloudflare has a recent product called Dynamic Workers which allow a company to stand up an AI workflow rapidly; a large AI studio went from zero Dynamic Workers to 1 million in 15 days

We launched something called Dynamic Workers, which allows you to very, very quickly stand up something which is significantly more efficient than a container. Containers are too slow and too heavy to actually be able to respond to these incredibly fast agentic workloads. And so what AI studios are doing is they’re looking at this and they’re seeing the opportunity. And so to give you a sense of — with the — I’m naming them, one of the large AI studios in just the last 15 days went from essentially zero Dynamic Workers to over one million Dynamic Workers running across the platform.

Cloudflare’s management thinks agentic AI will provide tailwinds for its legacy businesses

Every time an agent does something, like if you think about it, you just — if you type something into ChatGPT or any of the things, like to search — the number of sites that get searched, the amount of traffic that gets generated, if I’m looking for a digital camera as a human, I might visit five websites if I really care about it. My agent is going to visit 5,000. And so that’s going to just drive significantly more usage, which is the biggest driver of kind of our Act 1 revenue…

…For Act 2, again, as we talked about already, I think being able to very narrowly define what data an agent has access to and what data they don’t. We’re just seeing more and more of that usage, especially in the self-service category, which there really isn’t another sort of SASE, Zero Trust, self-serve competitor out there with any sort of scale. And so that’s with things like OpenClaw driving a lot of usage there. And what we found time and time again is as hobbyists or individuals adopt technology, they inevitably start to bring that technology more and more to work. And that’s what we’re seeing as we win more of the enterprise accounts across Act 2. 

Coupang (NYSE: CPNG)

Automation and AI is improving Coupang’s service levels and lowering its cost to serve; management expects automation and AI to help Coupang improve its customer experience and margins in the years ahead

Automation and AI across our services, including our Fulfillment and Logistics network, continue to improve service levels and lower cost to serve in parallel, and we expect them to be meaningful contributors to both the customer experience and margin expansion in the years ahead.

Datadog (NASDAQ: DDOG)

Datadog engineers are equipped with the latest AI coding tools and they are building rapidly; management sees the company’s AI initiatives as being split into 2 buckets, namely (1) AI for Datadog, and (2) Datadog for AI; AI for Datadog is about making Datadog’s platform better with AI products and capabilities while Datadog for AI is about Datadog’s end-to-end observability and security capabilities across the AI stack; in AI for Datadog, the company launched MCP (model context protocol) Server for general availability recently and it allows developers to debug applications directly in their AI coding agents; in AI for Datadog, the company launched Bits AI Security Agent recently and it reduces investigations from hours to as little as 30 seconds; in AI for Datadog, the company launched Bits Assistant in preview recently and it allows users to search and act across Datadog with natural language; in Datadog for AI, the company recently launched GPU Monitoring for users to understand their GPU fleets’ performance and drive higher GPU ROI (return on investment)

Our engineers enabled with the latest AI coding tools are building rapidly to help our customers confidently and securely deploy their applications…

…As a reminder, we’re talking about our AI efforts in 2 buckets: AI for Datadog and Datadog for AI. 

So first, AI for Datadog. These are AI products and capabilities that make the Datadog platform better and more useful for our customers. In March, we launched our MCP Server for general availability. With MCP Server, developers access live production data to debug their applications directly in their AI coding agent or IDE. We delivered Bits AI Security Agent, which autonomously triages Datadog Cloud SIEM signals, conduct in-depth investigations of potential threats and delivers actionable recommendations. We’ve seen Bits AI Security Agent reduce investigations that could take hours to as little as 30 seconds. We also shipped Bits Assistant now in preview, which helps customers search and act across Datadog using natural language prompts.

Moving on to Datadog for AI. This includes Datadog capabilities that deliver end-to-end observability and security across the AI stack. We launched GPU Monitoring, enabling teams to understand GPU fleet utilization, workload efficiency, thermal and power behavior and interconnect performance. This drives higher GPU ROI and operational reliability.

Datadog now has 6,500 customers sending data for their AI integrations (was 5,500 in 2025 Q4); these 6,500 customers are only 20% of Datadog’s total customer count, but represent 80% of the company’s ARR; customers’ usage of AI within Datadog is growing rapidly; Bits AI SRE agent investigations have increased by more than 100% from December 2025 to March 2026; the number of LLM spans customers are sending to Datadog is up 3x sequentially in 2026 Q1; the number of Datadog MCP Server tool calls is up 4x sequentially in 2026 Q1; the number of Bits Assistant messages is up 12x sequentially in 2026 Q1; some of the growing AI-related volume that Datadog is processing is because of enterprises’ adoption of AI coding tools; management is seeing an inflection point in AI consumption from customers, driven by a real move towards production-level AI workloads from both AI native and non-AI companies; management is seeing a massive increase in agent usage

We now have over 6,500 customers sending data for one or more of our AI integrations. Though this is only 20% of total customers, they represent about 80% of our ARR. And our customers’ usage of AI within Datadog platform continues to grow rapidly. Bits AI SRE agent investigations have more than doubled from December to March. The number of spans sent to our LLM observability product nearly tripled quarter-over-quarter. The number of Datadog MCP server tool calls quadrupled quarter-over-quarter and the number of Bits Assistant messages increased by a factor of 12 in that period…

…[Question] Is there any way to conceptualize the growth in the sheer raw volume of code that’s being produced in the world today due to adoption of code generators such as Claude Code and Codex and Cursor because they seem to be developing the capability to take on full projects?

[Answer] We definitely think and see that there’s many more applications being created. There’s going to be way more complexity in production. We see some of that happening already today. Some of those new applications are getting into production. They’re finding users. We see some signs of that at every layer of our platform. We quoted a few stats on the increasing data volumes we see in our AI products. That’s definitely a reflection of that. So we see an inflection point there in consumption from customers. We see a move to production that is very real, and we see that across both AI native and non-AI companies…

…We see both a stratospheric increase of agent usage. So we have a ton of usage on our MCP Server. We see customers trying to automate a lot with their own agents, using our agents, using a combination of those.

Example of a 7-figure and 8-figure land deals with the AI research divisions of 2 of the world’s largest technology companies (likely to be 2 of Meta Platforms, Microsoft, and Alphabet, with a likelier pairing of Meta and Microsoft because the deals included GPU monitoring for training workloads, and Alphabet trains on TPUs); the 2 technology companies are training advanced AI models and are relying on Datadog to reduce engineering friction and increase training velocity; the 2 technology companies will be using GPU Monitoring on large parallel GPU grids; the hyperscalers are the companies that make the most sense to pursue observability tools themselves, but they still choose Datadog to be efficient with their own resources; the hyperscalers are using Datadog for both traditional observability and GPU monitoring; it’s still early days for the hyperscalers in terms of their usage of Datadog, but Datadog’s management is optimistic that the 2 hyperscalers can be an example for other AI model builders in the future

We landed 2 large deals, a 7-figure and an 8-figure annualized deals with the AI research divisions at 2 of the world’s largest technology companies. These organizations are building and training the most advanced AI models in the world. It is critical for them to reduce engineering friction and increase training velocity, but fragmented internal and open source tooling made it harder to identify and solve issues and reduce engineering and research productivity. By using Datadog, both companies are accelerating their pace of innovation on their hyperscale AI training workloads. And this includes optimizing their workflows using GPU Monitoring on large parallel GPU grids…

…The thing that’s also interesting, in particular this quarter is that we also landed some large parts of hyperscalers. And hyperscalers typically have a culture of building everything themselves, and they certainly have the balance sheet and the human capital to support some of that build-out. Like if there was ever a set of companies for whom it makes sense to do it themselves, that would be those companies. And yet, we see that they have the same issues. When it comes to going as fast as they can and being as efficient as they can with their resources, like they come to us to replace some of the things that we were using before…

…[Question] About the hyperscalers because I thought that was particularly interesting. And the reason why is I don’t think you called them out previously before, and they are so prevalent in the modern tech stack. To your point, they could do this themselves. So I guess how are they using Datadog? Is it for more kind of traditional observability? Or is it for these newer areas like GPU monitoring that Datadog has performed so well of late?

[Answer] It’s both actually. When you look in general at the large AI customers, they use Datadog the way other companies are largely with a fairly broad set of our products to cover the full surface of observability. What’s new is we now have a product for GPU monitoring. It’s a very new product. And we see the hyperscalers that are coming to us for training workloads in particular, being very interested in that. So again, it’s too early in the product life cycle and the customer life cycle for these specific customers to call definitive victory there, but we see that as a very encouraging sign of where the market might go in the future because we think this might be a bellwether of what the next 10, 100, 500 companies that are going to start training workloads are going to want to do. We have some signs that go beyond the customers we signed this quarter that point that way too.

Datadog’s management continues to believe that digital transformation, cloud migration, and AI adoption are long-term growth drivers of Datadog’s business; management is seeing democratisation of AI training and a growing variety of AI accelerators being used (in management’s words, “the heterogeneity of silicon”), and management thinks both trends are positive for Datadog; the heterogeneity of silicon currently applies to only a very small handful of companies, but management sees a growing opportunity; management was historically more optimistic for AI inference as a growth market for Datadog, but they are increasingly seeing AI training as also a growth market for the company too, driven by growing adoption by the hyperscalers; management is agnostic about the source of usage on Datadog, whether it’s humans or agents; AI training is becoming a growth market for Datadog because it has changed from something artisanal to something in production-mode that has scaled by orders of magnitude and that needs to be incredibly reliable; management is investing heavily into security for AI agents; management thinks there’s a chance a good portion of the market leans towards on-premise observability products

There is no change to our overall view that digital transformation and cloud migration are long-term secular growth drivers for our business. But we now have an additional secular growth driver with AI as we help our customers deliver more value with this transformative new technology. Now more than ever, we feel ideally positioned to help customers of every size and every industry as well as all types of users, whether humans or AI agents, so they can transform, innovate and drive value through AI and cloud adoption…

…The broader market that’s interesting here is training, the training used to be something only 2 or 3 companies were doing or maybe 4 or 5 at a large scale. And it looks like training actually might democratize quite a bit more, and many companies will train models on a regular basis. So it becomes more of a viable category for service providers like us basically. I think the heterogeneity of the silicon is definitely a trend that plays in our favor there. The more heterogeneous, the more you need someone else to make sense of everything for you and tie it all together and also tie it all with the non-GPU aspects and the rest of the infrastructure and the applications and the users and the developers like basically everything we do for living…

…When you think of who is actually — who actually has heterogeneous environments today, that is still a very small number of companies, Google, barely just started selling their TPUs to the outside. So I think it’s still a small number of companies that are there, but we see a growing opportunity there.

Interestingly, last year, when we reported earnings, we said we’re mostly interested in inference workloads and training is not really a market for us yet. Now we actually see training becoming a market. We started landing customers that are actually hyperscalers that have a whole host of homegrown technologies and that are using us specifically in their super intelligence labs to help monitor their workloads, accelerate the training runs, monitor the GPUs also. So we see that as a point of validation that there’s going to be a great market for us…

…We don’t care whether most of the usage is humans, most of the usage is agents. Our business model lends itself to it pretty well, like we’re usage-based, and it doesn’t really matter where the usage is coming from, from that perspective…

…Training was very new a couple of years ago. It was something that was only done by very few companies, and it was, in a way, very artisanal. Like, it was not a production workload. It was something that researchers were building and that was very one-off and homegrown in ways. And now it’s turning into production. It’s turning into something that many more companies are doing. It’s scaling by orders of magnitude. And it’s becoming something that has to be on all the time, reliable and every minute you lose is — or rather every failure you have in your training runs is a week you give away to the competition. And so as a result, it becomes way more interesting as a market for us. And we see some signs of that. Again, we didn’t have a lot of it. We didn’t see a lot of it last year. Now all of a sudden, we’re starting to see quite a bit of activity there and demand…

…On the security of agents, we interface with that in 2 ways. So first, there’s the agents we build ourselves because we are building a lot of automation inside of our product for our customers and agents that automatically identify but also resolve issues without you having to do anything. And there, a lot of it has to do with understanding what permissions to apply, what kind of guardrails to apply, what kind of — how to interface with the humans and how to make that trustworthy and visible in the right way. And so that’s pretty much the whole product surface is to [indiscernible] data. The automation itself actually kind of works already. So you should expect to hear more about that at our conference. This is definitely one big area of investment for us…

…There was a question earlier on data residency and living in customers’ environments. We definitely see a great opportunity there. There is a chance that a good portion of the market leans this way in the future. Today, it’s not the largest part of the market, but we definitely see a potential for that. So we’re investing heavily in that sort of our product.

Datadog experienced adoption growth in AI native customers in 2026 Q1 that significantly outpaced non-AI customers; the AI native cohort continues to diversify and grow; 22 customers in the AI native cohort now spend more than $1 million annually, with 5 spending more than $10 million annually

Our AI native customer growth continues to significantly outpace the rest of the business. This group continues to diversify and grow, including 22 customers spending more than $1 million annually and 5 spending more than $10 million annually. This group includes the leading companies in foundational models, code-gen tools and vertical-specific AI solutions.

MercadoLibre (NASDAQ: MELI)

MercadoLibre’s management rolled out the company’s 1st AI-powered search experience in the marketplace business in 2026 Q1; the new search experience, which involved LLMs (large language models), has led to uplifts in conversion and click-through rates for sponsored listings in Brazil and Mexico; daily active users of MercadoLibre’s Seller Assistant grew 40% month-on-month in March 2026; an AI assistant has increased the productivity of MercadoLibre’s fulfillment network; the new search experience is able to better understand users’ intent

We rolled out our first AI-powered search experience in our marketplace in Q1’26, shifting the architecture away from keywords and rebuilding it around LLMs. In Brazil and Mexico, the improvement in product relevance led to uplifts in conversion and click-through-rate for sponsored listings, both of which represent incremental revenue. These are early results, which we believe have the potential to transform how our customers search and discover products on our platform. Engagement with our Seller Assistant is strengthening, with daily active users growing more than 40% MoM in March. In shipping, an AI-powered assistant that provides reps with real-time process information and performance challenges has increased productivity across our fulfillment network…

…I think it’s worth highlighting the fact that we deployed LLMs in search in commerce for the first time this quarter. And basically, that is live in Brazil, Mexico and Argentina. So now we are using this technology to better understand users’ intent, combining both knowledge on the user behind the query and better interpretation of the query itself.

MercadoLibre’s AI Assistant in MercadoPago is now automatically alert users about negative balances and also identifying opportunities for users to earn higher yields on their savings; AI tools are helping MercadoLibre’s sales force for the Acquiring business to be more productive

In Fintech, our AI Assistant is becoming more proactive. In Brazil, it now alerts users to negative balances in accounts connected via Open Finance and identifies funds held elsewhere that could be earning a higher yield with Mercado Pago — and crucially, it can act on these opportunities instantly, moving balances between accounts within seconds. This is a meaningful step beyond a traditional assistant: it is not just surfacing information, it is helping users take action. In Acquiring, AI tools continue to drive significant improvements in sales force productivity, contributing to the strong market share gains we are seeing across the region. 

Through AI, MercadoLibre’s productivity KPIs were up 56%-80% year-on-year in 2026 Q1 even though headcount was up by just 8%; senior engineers now spend  time building code instead of reviewing code; MercadoLibre is rolling out Claude CoWork to its 31,000 employees

Headcount grew 8% YoY in Q1’26 – a carryover effect of 2025 hiring – but productivity KPIs are growing 7-10x faster. Many of our most senior engineers that were previously spending most of their time reviewing code are now also building code because of the productivity gains enabled by AI tools. Rollbacks – code that is returned to its developer due to errors – are materially lower YoY. More broadly, we have rolled out Claude Cowork to 31,000 employees, making Mercado Libre one of the earliest, large-scale enterprise adopters globally. 

Shopify (NASDAQ: SHOP)

Shopify’s management had bet early on AI and now AI is embedded in everything the company does; Shopify shipped 300 new products and features in 2025 while keeping headcount flat; Shopify has an AI coding partner built right into Slack

In 2026, AI is now Shopify’s native language. We bet early on AI and forced its adoption. It’s embedded in everything we do, the products we build, the channels we power, the way every single person on the team operates. AI has become an exoskeleton for everyone at Shopify, giving them a virtual team of agents and that makes room for rapid experimentation. It allows them to pursue multiple ideas at the same time and then double down on the winners…

…We shipped over 300 new products and features last year alone. We kept our flat head count, which we’re very proud of. And that’s only possible because something has changed fundamentally. And I know Tobi has been talking a bit about river, which is a perfect example of it, but it’s this AI coding partner built right into Slack for the entire team where they can pull into any threat, any conversation and do, frankly, a remarkable amount of the engineering work. And we built it because we needed it, and now it’s deeply embedded in how we operate.

Shopify’s management believes that entrepreneurs will benefit deeply from AI because AI-powered shopping democratises discovery, and this in turn benefits Shopify; each time the world gets more complex, Shopify becomes more valuable for merchants because the company absorbs the complexity into its systems; management sees 3 reasons why Shopify is in a very strong position in the AI age, namely, the company’s (1) data on millions of merchants, hundreds of millions of buyers and billions of products, that enables it to build products informed by the insights developed from the data, such as Sidekick mentioned, (2) demand conversion flywheel, and (3) ability to absorb complexity for merchants; Shopify’s structural advantage is that it gives merchants everything they need, and the company is shipping products even faster now through AI

No group benefits more from AI than entrepreneurs. The logic is simple. AI is making entrepreneurship dramatically more accessible and in fact accelerated. That means we’re going to see more entrepreneurs, and they’re going to scale more easily. AI-powered shopping democratizes discovery. Reach is not just influenced by budget anymore, it is influenced by relevance, which benefits both merchant and buyer. And the right products find the right shopper at the right moment. And this is enormous potential for new and scaling merchants. And because we win when they win, it also has enormous potential for Shopify…

…Every single time the world gets more complex, Shopify gets more valuable. We absorb more of that complexity into our systems and become more valuable to merchants. So when we look at this new era of commerce that we’re in, there are really 3 core principles that explain why Shopify is in such a strong position…

…The first principle, Shopify has a huge advantage that is about to compound. We have 20 years of commerce data. We have data on purchase intent across millions of merchants, hundreds of millions of buyers and billions of products. And in a world where real-time information is now table stakes, the edge is the insight beneath it. And that requires depth, not just access, but experience. We’ve seen merchants start, stall, pivot and scale millions of times across every category and geography. It allows us to build on the real behavior of commerce and to keep shipping products grounded in insights only we have, deep experience applied at speed. That is very hard to replicate and it compounds…

…The second principle, which is the demand conversion flywheel. It should be getting more obvious that every quarter that Shopify is no longer just the platform to convert demand, we are becoming the platform to create it too. And that end-to-end position is a major advantage for merchants…

…The third principle I’ll leave you with is what I call invisible complexity. Here’s the thing. The hardest parts of commerce are the parts that nobody sees, and this is where Shopify thrives…

…That’s the structural advantage of Shopify. We give you everything you need by operating across the entire commerce stack. It’s not the power of any one element of the platform. It’s how they all work together to help merchants accelerate their success. It’s the knowledge and expertise readily available through Sidekick. It’s the speed, context and simplified complexity behind checkout. It’s the ability to sell across every channel, every surface and every geography from day 1. Internally, we are making every function faster, sharper and more productive, and output per employee is improving through deliberate AI usage. The result is that we are building more, shipping more and serving more merchants.

Sidekick is Shopify’s intelligent assistant for merchants that is trained on the company’s knowledge base; the number of weekly active shops using Sidekick grew 385% year-on-year in 2026 Q1; 12,000 custom apps were created with Sidekick in 2026 Q1, up 200% sequentially; half of all Shopify Flows (Shopify’s workflow builder) generated in 2026 Q1 were built with Sidekick; theme edits with Sidekick was in the multimillions in 2026 Q1, up 1,000%; Sidekick has a smart suggestions feature called Pulse; Pulse recently suggested to an accessory brand to create a social proof page and when the accessory brand agreed, the page was created in minutes at no incremental cost to the accessory brand; in the past, the accessory brand would have required a team and several weeks to build the page; merchants that use Sidekick become power users very quickly; Sidekick is used internally at Shopify; management sees Sidekick as a complement to Shopify’s App Store, not a replacement; Sidekick is enabling merchants to build individualised apps rapidly, and thus, move much faster

Sidekick is the perfect example of this. As a reminder, this is our intelligent assistant, which is trained on our knowledge base, paired with completely personalized intel, it has about each merchant’s particular business…

…The number of weekly active shops using Sidekick in Q1 was up 4x year-over-year. We saw over 12,000 custom apps created in Q1 alone using Sidekick. And nearly half of all Shopify flows generated in Q1 were built with Sidekick. And theme edits just from last quarter are in the multimillions, growing over 1,000% in a single quarter. And theme edits just from last quarter are in the multimillions, growing over 1,000% in a single quarter…

…And then there’s Pulse. Sidekick’s smart suggestions feature, which proactively delivers personalized recommendations for merchants using market trends and data from their store, which Sidekick then executes on the merchant’s behalf. And I’ll give you a great example that I just saw the other day. It was an accessory brand, and Pulse noticed that this brand was getting attention in the right places. Its products were being endorsed by fashion publications and showing up on celebrities’ Instagram profiles. So it proactively suggested that the merchant create a social proof page on their website to build trust and validation. And once the merchant agreed, Sidekick created that page on the merchant’s behalf, and it was already all within minutes. Now just a few months ago, that process multiple specialists, marketing, UX design, copywriting and often an incremental cost to the merchant and likely several weeks from start to finish. And now it is happening autonomously in minutes at 0 incremental costs to the merchant. And that is just one of the smart recommendations being served up to that merchant as part of their daily operations…

…Weekly active shops are up 385% using Sidekick. We saw 12,000 custom apps built in Q1, which is up like over 200% quarter-over-quarter…

…Merchants that are just starting to play with it really become power users very, very quickly…

…The impact that we’re seeing not only in terms of how our merchants are using Sidekick, but how we’re using it internally has been super impactful…

…Some of them have actually discovered this incredible tooling, they’re building for their own business and then put in the App Store as well. But in terms of what Sidekick is doing, like Sidekick actually, we see as a real supplement to the App Store, not a replacement…

…The applications that are being built by Sidekick are really very specific nuanced feature sets for particular merchant businesses. And so for most of them, it really is just for the individual merchant. We see them — we see those — the opportunity for the app developers just to continue. That being said, though, what is happening that is super interesting is that now merchants who may have had to spend weeks or even months building a feature either internally or hiring an agency to do so, they’re able to do so much more work themselves using Sidekick, and that means they’re able to go much faster.

Shopify’s management thinks that emerging AI channels for shopping, such as ChatGPT, Microsoft Copilot, and Google and Meta’s AI services, will be a tailwind for e-commerce; Shopify is the only platform enabling discovery and selling inside ChatGPT, Copilot, and Google from a single system of record; AI-driven traffic to Shopify stores is up 8x year-on-year in 2026 Q1; orders from AI-powered searches are up 13x year-on-year in 2026 Q1; new buyer orders from AI-channels are happening at 2x the rate of other channels; Shopify’s Catalog feature provides the necessary information on 1 billion products for AI agents to surface the most relevant products in seconds; traffic from Catalog-powered AI searches converts 2x more traffic than general AI searches; usage of Shopify’s Sign In With Shop user verification tool is up 3x year-on-year in 2026 Q1; Sign In With Shop is important for agentic commerce because it enables agents to know who they are buying for; agents are not bypassing Shopify; Shopify is the storefront within ChatGPT’s recent move to having in-app browsers for checkouts; Shopify recently introduced an agentic plan that allowed brands to sell in AI channels through Shopify Catalog with no Shopify stores required; non-Shopify merchants are realising that Shopify Catalog is enabling their products to surface on agentic surfaces much better than web-scraping, and it is leading these merchants to join the Shopify ecosystem; OpenAI and Microsoft are already using Catalog

We believe that new and emerging AI channels, places like ChatGPT, Microsoft Copilot, Google AI Services and Meta will be a tailwind to driving e-commerce growth and penetration over time…

…We are the only platform that enables discovery and selling inside ChatGPT, Copilot and Google, all from one single system of record. And the early signals on AI channels are really compelling. And in the first quarter, AI-driven traffic to Shopify stores has grown 8x year-over-year, while orders from AI-powered searches have increased nearly 13x. And within this, new buyer orders are occurring at nearly twice the rate of other channels…

…Let’s talk about Shopify’s catalog because this really, really matters. To date, we’ve structured more than 1 billion products with clean attributes, real-time pricing and accurate inventory so AI agents can surface the most relevant products in seconds, and the results speak for themselves. Traffic from catalog-powered AI searches converts 2x more than traffic from general AI searches where the agent is working from scraped or often outdated information from across the web…

…Sign in with Shop is our user verification tool, which recognizes buyers across devices, stores and surfaces with no sign-in friction. And usage is growing steadily. We are up 3x year-over-year, and it is now enabled across nearly our entire merchant storefront base. In an agentic world, this really matters. Agents need to know who they are buying for and we are ready…

…Agents do not bypass Shopify, just the opposite. In fact, they write right into Shopify. I mean, I think you saw in sort of recent headlines that merchant storefronts really matter. You saw ChatGPT move to in-app browsers for their checkouts. So it’s literally the Shopify storefront within the chat. And again, when a buyer is shopping in ChatGPT, they’re browsing Shopify’s incredible catalog. So the momentum on agentic has been amazing…

…In terms of some of the stuff we’re doing with the agentic plan, for example, again, that rolled out early March. That means that any brand on any platform can now sell across AI channels via Shopify Catalog and no Shopify stores required…

…The big thing, though, with catalog is that I think a lot of non-Shopify merchants are seeing that catalog is actually doing a much better job of organizing and syndicating their products across every agentic surface versus sort of the old scraping thing that was happening prior to catalog. So it’s doing 2 things. One, it is unequivocally getting Shopify connected with a lot more non-Shopify merchants per se and beginning those conversations, which, again, may lead to them joining the agentic plan or ultimately may lead them to come into Shopify for their entire migration, which obviously is our plan and our hope. But even if they just want to be part of catalog and just be part of the agentic plan on its own, that already is a massive lift to them relative to everything else…

…OpenAI and Microsoft are already using the Catalog power discovery.

Shopify co-developed the open Universal Commerce Protocol (UCP) with Google; UCP enables the full commerce journey from product discovery to post-purchase support; management built UCP because they believe that agentic commerce should be based on open standards; management has created the UCP Tech Council, which recently saw Amazon, Meta, Microsoft, Salesforce, and Stripe become members

You might have seen with the latest news on the Universal Commerce Protocol, or UCP, which we co-developed with Google. UCP is an open protocol that makes Agentic commerce work at scale. It enables the full commerce journey, product discovery, checkout, payment, post purchase across any platform with any payment processor.

We co-developed UCP because we believe the future of commerce runs in open standards, not closed systems. And then we created the UCP Tech Council, the technical body that steers the protocol’s direction to ensure it evolves to meet the needs of businesses, platforms, developers and consumers. We are now seeing the biggest and most innovative companies across essentially the entire industry coming together around UCP to help push Agentic commerce forward. And last month, Amazon, Meta, Microsoft, Salesforce and Stripe all joined the council, committing their expertise in Internet scale transaction processing to build one universal protocol for commerce.

Gross margin for Subscription Solutions was similar to a year ago, as economies of scale and efficiencies in support were partially offset by increased LLM costs from growing usage of Shopify’s AI products; management expects pressure on the gross margin from usage of Shopify’s AI products to continue

Gross profit for Subscription Solutions grew 21%, with gross margin coming in at 80%, in line with Q1 2025. Economies of scale and efficiencies in support were partially offset by increased LLM costs, driven by growing merchant usage of our AI products, most notably Sidekick. We expect this dynamic to continue.

AI is writing about 50% of Shopify’s code today; there are more app developers building for Shopify’s ecosystem than ever before, and Shopify is using AI to speed up the app approval process

AI right now writes well over 50% of our code today, and that number is going up significantly, not down…

…You’re seeing more app developers build for Shopify’s ecosystem than ever before. In fact, we’ve now put the app approval process on rails using incredible AI testing so that we can get more apps into the app store faster.

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, Coupang, Datadog, MercadoLibre, Meta Platforms, Microsoft, Salesforce, and Shopify. Holdings are subject to change at any time.

What We’re Reading (Week Ending 10 May 2026)

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 10 May 2026:

1. Corporate dark arts: when incentives tell you what might be coming $GME $EKSO $VAC $RPD – Andrew Walker

EKSO is a tiny little company; its market cap for most of last year was <$10m. But it’s a perfect case study in the dark arts and why paying attention to them can be profitable. In late November they gave all of their executives’ PSUs that vested only if the company underwent a change of control and the stock was “at least $7.50” per share within the next five years. The stock was trading in the mid-$4s at the time.

I’m not sure I’ve ever seen a single PSU grant that flashes “we are for sale” harder than that grant.

Sure enough, at the end of December EKSO announced a deal to merge with APLD’s cloud business spinoff. A few weeks later, EKSO did a private placement; it will shock you to learn the placement was priced at $8.22/share, above the mark that vested EKSO’s PSUs.

It wasn’t guaranteed that the market would respond positively to EKSO’s merger…. but I’d suggest EKSO’s board and management knew something was in the pipes when they made those grants, and that whatever was coming was likely to excite the market.

As I write this, EKSO is trading at $12/share.

2. OpenAI’s AI Chip Deal With Broadcom Hits $18 Billion Financing Snag – Anissa Gardizy

When OpenAI and chip designer Broadcom announced last fall that they would make custom artificial intelligence chips together, they positioned it as a done deal.

The companies said the deal would bring enough chips online before 2030 to consume 10 gigawatts of power, equivalent to five Hoover Dams’ worth of electricity, in a bid to lessen OpenAI’s costly dependence on Nvidia hardware.

What they didn’t say was that they hadn’t figured out how OpenAI would pay for the project.

Months later, the companies are negotiating an agreement for Broadcom to finance the first phase of chip production, which would consume 1.3 GW of data center capacity and would cost around $18 billion, according to an internal memo and two people involved in the talks. At that rate, the full 10 GW program, code-named Nexus, could cost $180 billion in chip production alone before factoring in data center construction and other costs…

…But the negotiations have run into a potential problem. Broadcom has said it would finance the first phase only if Microsoft agrees to buy roughly 40% of the chips, an OpenAI executive told colleagues in a memo last month. Microsoft would install the chips in its data centers and then rent them back to OpenAI.

A purchase commitment from Microsoft, one of the world’s most creditworthy companies with decades of data center experience, would give Broadcom confidence it would get its money back, said a person involved in the talks.

But Microsoft could choose not to buy OpenAI’s chips, which would change the financing terms for the project, the memo said…

…OpenAI has made a habit of announcing landmark partnerships without ironing out the details. A month before the Broadcom announcement, for instance, OpenAI said Nvidia would provide up to $100 billion in funding, allowing OpenAI to build its own data centers and use Nvidia’s chips to power them. The headline-making deal eventually fizzled, though Nvidia later made a $30 billion equity investment in OpenAI…

…And in January 2025, OpenAI announced Stargate, a joint venture with SoftBank and Oracle to spend $500 billion developing data centers. But the effort floundered as the three sides disagreed over details and lenders balked at backing multibillion-dollar projects tied directly to a company with an unproven business model…

…Despite the risks from Microsoft’s sway, talks between Broadcom and OpenAI have been progressing. Broadcom had long insisted that OpenAI put up one dollar of its own for every dollar Broadcom provided in financing, a typical arrangement to limit the chip vendor’s risks. That requirement had become a sticking point in the talks, according to the memo and an executive involved in the talks.

But Broadcom recently decided to relax that demand and invest more capital up-front than OpenAI, breaking from Broadcom’’s “long-held hard-line requirement,” the OpenAI memo said.

3. The Fertilizer, the Bond Market, and the End of the Country Banker – Dirt Cheap Banks

Chapter 12 farm bankruptcy filings rose 46% in 2025. That followed a 55% rise in 2024. That is the third consecutive annual increase. The Midwest jumped 70%. The Southeast jumped 69%. Montana, of all places, jumped 200%. Pennsylvania jumped 160%. Arkansas led the country in absolute filings — the most this century for the nation’s top rice-producing state. Total farm debt is projected to hit a record $624.7 billion in 2026. The American Farm Bureau Federation surveyed 5,700 farmers and 70% of them said they could not afford all the fertilizer they needed for the spring. The U.S. Department of Agriculture itself — not some doom-pusher on the internet, the actual government department whose job is to make this look fine — projects that 2026 corn will cost roughly $5.00 a bushel to grow and sell for $4.20. Soybeans, $12.27 to grow and $10.30 to sell…

…Nearly 40% more new farm operating loans were opened in Q4 2025 than in Q4 2024. The average operating loan in 2025 was 30% larger than 2024, with maturities running three months longer. Farmers are not borrowing more because they are growing. They are borrowing more because they are bleeding. And the only reason aggregate farm income looks anything like solvent is that the federal government will spend roughly $55 billion this year — $44.3 billion in direct payments, plus crop insurance subsidies, plus the $11 billion Farmer Bridge Assistance Program — propping up an industry that is, in market terms, no longer functional. Strip the subsidies out and 2026 net farm income falls off a cliff that nobody in Washington wants to look over. Agricultural lenders surveyed by the American Bankers Association expect only about 58% of farm borrowers to remain profitable in 2025, down sharply from 78% in 2023. NDSU’s Agricultural Risk Policy Center projects $44 billion in net cash income losses on the 2025-26 crops alone…

…The North Dakota State University agricultural trade modeling team ran the fertilizer scenarios and they are worth your attention because they are the most rigorous public modeling that exists.

Under their “Quick Reopening” case, urea peaks at $782/short ton in June 2026 and eases gradually. Under their central “Contested Transit” case, peak urea hits $784/st in July with prices staying above $700/st through November; fall 2026 prepay urea averages $733/st (56% above pre-crisis); winter fill at $643/st; spring 2027 top-off at $590/st. Add another fifty to eighty dollars per ton for freight and dealer margin to get the actual interior Corn Belt retail price. Under their “Extended Conflict” case, fall prepay climbs to $989/st; winter fill to $945/st; spring 2027 spot prices remain near $791/st. The World Bank’s Commodity Markets Outlook, released April 28, expects global fertilizer prices to rise more than 30% in 2026, with urea closing the year at $675 per ton — nearly 60% above 2025 levels.

For the farmer, this means 2026 is the easy year. Most spring 2026 nitrogen had already been contracted before the Strait closed in February. The real budgeting concern is 2027. American Farm Bureau Federation survey data shows that for every farmer more concerned about fertilizer for 2026, nearly two are more concerned about 2027. Damage to liquefied natural gas production and sulfur output in the Persian Gulf could take years to repair, even if shipping normalizes tomorrow. The infrastructure does not just turn back on.

If the central NDSU scenario plays out, the 2027 crop year sees farmers face fertilizer costs roughly 50% above pre-war levels at exactly the moment their working capital — the cushion that lets them absorb a bad year — has been exhausted by 2025 and 2026. This would be the fourth consecutive year of negative crop margins. Operating loans would grow even larger, even longer. Chapter 12 filings would push past 600 a year. Agricultural bank delinquency rates, currently 1.09% as of July 2025, would climb to 2.5% to 3.5%. Still well below 1985’s peak of 6.7% at agricultural banks, but moving in the wrong direction at speed.

If the extended conflict case plays out — Strait remains contested through 2027, fertilizer at near-1980-level real prices, fifth consecutive year of negative margins — the trajectory accelerates. 2028 starts to look uncomfortably similar to 1984. The structural buffers begin to fail in sequence, not in parallel…

…The American Enterprise Institute has been making the case openly: most farm households receive over half their income from non-farm sources; the agricultural sector’s debt-to-asset ratio is 13.75%; the system can absorb shocks without the level of subsidization currently in place. That argument is not winning yet. But it is being made by serious people in Washington, and it is being made at a moment when every other federal spending priority is under similar pressure. If a debt-ceiling fight or a continuing-resolution fight produces a sequester or a freeze, agricultural subsidies are not exempt. They are politically vulnerable in a way they have not been in a generation.

If subsidies are cut even modestly — say, a 30% reduction from the projected $55 billion to roughly $38 billion — the market-based losses that currently get masked by federal payments become visible all at once. Farm income drops by an amount equivalent to roughly 11% of total receipts. The farms that are barely solvent stop being solvent. The farms that depend most heavily on subsidies — the commercial row-crop operations in the Midwest and Plains, the largest borrowers, the ones holding the biggest loans at the most concentrated agricultural banks — fail in clusters.

If subsidies are cut substantially — back toward the 2024 level of roughly $10 billion — the math becomes cataclysmic. Net farm income outside government payments would fall by roughly $40 billion. The structural protection that has kept the current stress from becoming a 1980s-style crisis disappears. Farmland values, which have so far held in part because farmers can still service their debts, begin to crack. The 220 community banks that the FDIC identifies as having agricultural loan concentrations above 300% of capital become acutely vulnerable.

This scenario is the dark mirror of 1985. In 1985, there were no subsidies of this scale to remove. The crisis happened anyway. In 2027 or 2028, removing the subsidies would be the trigger that closes a system that is currently holding together by their grace alone…

…The 1980s farm crisis killed 205 agricultural banks between 1984 and 1987 — 37.4% of the 548 total bank failures during that window. There were 14,483 FDIC-insured commercial banks in 1984; by 2023 that number had fallen to 4,027 — a 72.19% decline. At the end of 2024 there were approximately 4,050 community banks left in the United States. Roughly 220 of them carry agricultural loan concentrations above 300% of capital, clustered in eight states: Illinois, Iowa, Kansas, Minnesota, Missouri, Nebraska, North Dakota, and South Dakota. Most have under $200 million in assets. Most are not publicly traded.

4. Warren Buffett Case Study – East Sullivan Mines 1962 – Dirt Cheap Stocks

At yearend 1962, the Buffett partnership was managing $9.8 million.

East Sullivan was a $106,000 position.

East Sullivan was a mining business that produced copper, gold, silver and zinc.

It was headquartered in Quebec and formed in 1944.

East Sullivan had profitable operations. In 1962, it produced millions of pounds of zinc and copper along with 4,600 ounces of gold and 168,000 ounces of silver.

In 1962 the business had 33% EBIT margins. 1961 had 20% EBIT margins.

It was a nice little business. Of course, margins would swing wildly in this kind of operation, but still, it was doing well when Buffett owned it.

The business had cash and investments in excess of its market cap. It was profitable and paying a sizable dividend.

East Sullivan’s investments were largely made up of ownership in affiliated companies.

Members of the Beauchamin family made up the majority of the management team and the board.

Then there were a bunch of related businesses that were also interconnected and controlled by the Beauchamin family…

…East Sullivan was doing $1.2mm of EBIT from its own operations.

Let’s assume that the $9.6mm of marketable securities and affiliated businesses could produce a 7% return. That’s probably conservative.

7% on $9.6mm is an additional $672k of look-through ebit.

The market cap was $8.9mm. EV would’ve been $7.8mm if only giving credit for East Sullivan’s cash account.

The look through EBIT is $1.9mm (1.2mm + 672k).

That’s ~4x EV/EBIT…

…We don’t know how long Buffett held. But the investment was likely a good one for him.

Shares touched $3.00 in 1963. By 1964 they were $5.70. And they peaked at $9.40 in 1965.

If Buffett had held to the top in 1965 he would’ve earned a 73% IRR.

If he held through the end of his partnership in 1969, he would’ve earned a 34% IRR.

5. Iran war is crushing Asia’s farmers, threatening global food supply – Rebecca Tan and Wilawan Watcharasakwej

Saithong Jamjai has just finished harvesting the rice on the 19 hectares of farmland she owns in central Thailand and now is the time to sow again. But she won’t, she said, because of the U.S.-Israeli war against Iran.

She has gone over the math for weeks. Because of surging prices, driven by the war, of fuel, fertilizer, plastics and other necessities, planting and harvesting will cost her at least $33,000, she said. The grain that she’ll produce, she estimates, will sell in August for only $22,000.

“A confirmed loss,” Saithong, 53, concluded. She’d rather let her land bake under the yellowing husks from last season…

…Addressing world leaders in Rome on Thursday, Dongyu Qu, the director general of the U.N. Food and Agriculture Organization, said the war had created not only a geopolitical crisis but “a disruption at the core of the global agrifood system.”

Iran’s destruction of gas infrastructure in the Gulf and the dueling U.S.-Iran efforts to choke the Strait of Hormuz have prevented crucial supplies of fuel and its derivatives like urea — a potent source of nitrogen that enhances harvests — from leaving the Middle East. Because fuel infrastructure takes years to build, there is no ready replacement for these supplies.

In effect, 30 percent of the world’s urea has been “wiped out,” said Pranshi Goyal, senior analyst at the market intelligence firm CRU Group. China, a major fertilizer producer, has restricted exports to ensure its farmers have enough. Russia, another big manufacturer, is seeing demand soar, potentially boosting its economy and aiding its war in Ukraine. On what is known as the spot market, urea prices are up 40 percent since February…

…The longer the production plants in the Middle East stay closed, the longer they will take to restart. “This problem builds in a nonlinear fashion,” Goyal said.

So do its repercussions.

In Thailand, the Philippines, Bangladesh and Australia, which are the first since the war to enter key sowing periods, farmers are choosing to skip or reduce planting, or cut fertilizer use, which will lower yield.

As the war stretches deeper into the crop calendar, farmers from more countries will be forced to make similar choices, said Maximo Torero, chief economist for the FAO. “Right now, the impacts are more severe in Asia,” Torero said. “But clearly, this is moving east to west and south to north.”

In June, India and Brazil, two of the world’s biggest agricultural producers, will ramp up orders for urea. If, by then, vessels carrying urea are not sailing, there will be “significant yield loss” across many countries, Torero said…

…Thailand’s Commerce Ministry, for example, said in April the country still has 343,000 tons of urea fertilizer, sufficient to support the upcoming planting season. Driving through the vast flatlands surrounding Thailand’s Chao Phraya River basin, however, reveals a different picture.

Across Ayutthaya and Suphan Buri provinces, fertilizer shops large and small were completely out of urea — and said they had been for weeks. Distributors are offering only Russian compounds that farmers are wary to use, shop owners said. Seansdee Teerasattayaporn, 62, who runs a fertilizer wholesale business, sent a truck to a marketplace frequented by large dealers to try to procure urea but after waiting four days, he said, the truck returned empty.

Heading into planting season, many farmers said they are facing the worst conditions in their lifetimes. Not during the outbreak of the Russia-Ukraine war were shortages or costs this dire, they said. Nor during the pandemic…

…In an interview, Foreign Minister Sihasak Phuangketkeow asserted that Thailand still has sufficient farming supplies and Thai leaders are jetting across the world to procure more. But he acknowledged the country is competing against bigger nations with deeper pockets, amid extraordinary logistical challenges. “We have not faced such a crisis before,” he said.

On Tuesday, two weeks after a trip to Moscow, Thailand’s agricultural minister said an attempt to secure urea from Russia is likely to fall through. Because of shipping disruptions, it would take at least two months for Russian urea to arrive in Thailand — far too late for the current planting season.

Agricultural experts say the Iran war has underlined the need for farmers to become more self-reliant, for example, weaning themselves of diesel by switching to solar power or swapping out chemical fertilizer for organic alternatives that can be produced locally. But to make these switches, farmers need government subsidies and time, both of which are in short supply, said Esther Penunia, secretary general of the Asian Farmers Association…

…Thai farmers have been doubly hurt because the Middle East is also one of their biggest export markets. The region accounted for 17 percent of Thailand’s rice exports in 2025, according to customs data. Iraq was the single largest destination for Thai rice.

The day U.S. and Israeli forces bombed Iran, ship operators at a Bangkok port told sellers to lift containers of rice bound for Gulf countries off ships and back into warehouses, said Chookiat Ophaswongse, president of the Thai Rice Exporters Association. Since then, there have been no shipments of rice to the Gulf. Malaysia and the Philippines have absorbed some of Thailand’s excess supply but not all of it, leaving a glut that has kept rice prices low, Chookiat said.

Even before the war, many Thai farmers were in financially precarious situations, relying on loans to survive from one season to the next. Now, the squeeze of higher planting costs and lower projected rice sales could drive millions of farmers into spiraling debt that will take years to clear, said Pramote Charoensilp, 64, president of the Thai Farmers and Agriculturists Association. 

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 Microsoft. Holdings are subject to change at any time. 

The Latest Thoughts From American Technology Companies On AI (2026 Q1)

A collection of quotes on artificial intelligence, or AI, from the management teams of US-listed technology companies in the 2026 Q1 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. Since then, developments in AI have progressed at a breathtaking pace.

We’re thick in the action of the latest earnings season for the US stock market – for the first quarter of 2026 – and 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:

With that, here are the latest commentary, in no particular order:

Alphabet (NASDAQ: GOOG)

Gemini Enterprise has 40% sequential growth in paid monthly active users in 2026 Q1; Gemini 3.1 Pro is pushing the frontier in reasoning, multimodal understanding, and cost; there are now a wide variety of models in the Gemini 3.1 family to meet different developer needs; Gemini 3.1 Flash Live is powering conversational features in search and the Gemini app, and speech-to-text is now available in 70 languages; Gemini 3.1 Pro has delivered a big upgrade to Alphabet’s Deep Research product; the Lyria 3 model has generated over 150 million songs since its launch in the Gemini app; Nano Banana 2 has generated 1 billion images in half the time of Nano Banana 1; management recently launched Gemma 4, Alphabet’s best open model to date, and it has been downloaded more than 50 million times in a few weeks; Nano Banana 2 was recently integrated into the Gemini app to enable personalised image creation; Gemini is now integrated with Google Maps, so users can converse with Google Maps via chat

Gemini Enterprise is seeing tremendous momentum with 40% growth quarter-over-quarter in paid monthly active users…

…Gemini 3.1 Pro continues to push the frontier in reasoning, multimodal understanding and cost. We have quickly expanded the Gemini 3.1 series of models to offer more choices for developers, including our cost-efficient Flash models. 3.1 Flash Live, our latest audio model, has improved precision and reasoning, making voice interactions more natural and intuitive. It’s now powering conversational features in search and the Gemini app. Speech-to-text is now available in 70 languages. And with 3.1 Pro, our Deep Research agent got a big upgrade, including MCP support and native visualizations.

Our generative media models are incredibly popular. Lyria 3 has generated over 150 million songs since launching on the Gemini app. Nano Banana 2 reached 1 billion images in nearly half the time of Nano Banana 1. And Veo 3.1 Lite is our most cost-efficient video model to date.

On top of this, we launched Gemma 4, our most intelligent open model. It’s been downloaded over 50 million times in just a few weeks. In fact, our open models have now been downloaded over 500 million times…

…This month, we integrated Nano Banana 2 to make personalized image creation possible in the Gemini app. Maps recently got its most significant upgrade in over a decade with Gemini. Users can now have a conversation with Maps and get more personalized suggestions and intuitive directions.

Alphabet’s management thinks Google Cloud has the widest variety of compute options with Alphabet’s custom TPUs and Axion CPUs, and NVIDIA GPUs; Google Cloud will be among the first cloud providers to offer NVIDIA’s Vera Rubin NVL72 systems; Alphabet recently introduced the 8th generation of TPUs that has a training variety and an inference variety; TPU 8t, the training variety, offers 3x the processing power and 2x the performance of the previous generation; TPU 8i, the inference variety, has 80% better performance per dollar in inference compared to the previous generation; Alphabet’s TPUs are powering the company’s AI research in both training and tooling; management will begin to deliver TPUs to select customers in their own data centers to expand the TPU opportunity; management expects to recognise most of the revenue of external TPU shipments in 2027; management does think about the ROIC of external TPU shipments compared to internal deployment

Our custom TPUs, Axion CPUs and the latest NVIDIA GPUs continue to form the industry’s widest variety of compute options. NVIDIA GPUs are a core part of our AI accelerator portfolio and will be among the first to offer NVIDIA Vera Rubin NVL72 in addition to the Blackwell and Hopper-based instances already available.

At Cloud Next, we introduced our 8-generation TPUs, individually specialized for training and serving and able to take on the most demanding agentic workloads. TPU 8t provides high-performance model training with 3x the processing power of Ironwood and 2x the performance. TPU 8i delivers cost-effective, low-latency inference with 80% better performance per dollar than the prior generation. This exceptional infrastructure powers our world-class AI research that includes models and tooling, which continue to progress really well.

Our TPUs continue our leadership in performance, cost and power efficiency for customers like Thinking Machines Lab, Hudson River Trading and Boston Dynamics. As TPU demand grows from AI labs, capital markets firms and high-performance computing applications, we’ll begin to deliver TPUs to a select group of customers in their own data centers in the hardware configuration to expand our addressable market opportunity…

…We expect to begin recognizing a small percent of the revenues from these agreements later this year with the vast majority of revenues to be realized in 2027. It is important to keep in mind that revenues from TPU hardware sales will fluctuate from quarter-to-quarter, depending on when TPUs are shipped to customers…

…On the second question around TPUs, obviously, I would — we do think about it as what are we doing through Google Cloud to help our customers? And that’s the framework with which we think about it. In that context, there are situations where it makes sense. For example, you take customers like capital markets where they are running this highly performant AI workloads. They wanted TPUs in their data centers. So there are — and those trends are true across a diverse set of industries and in certain cases, frontier AI labs, too. And so we are opportunistic about it. But I do think we step back and think about it overall as the opportunity for Google Cloud. A lot of it is providing infrastructure through cloud. At times, it is direct sales of TPU hardwares to a select group of customers. But again, we do take ROIC approach. And some of it helps us get more economies of scale, scale in our overall compute environment as well. And so helps us invest in the cutting edge, which we need to do in the next generation as well.

Alphabet is using Antigravity, the company’s 1st-party agentic coding solution, to manage fully autonomous digital task forces

With Antigravity, we are shifting to truly agentic workflows. Our engineers are now orchestrating fully autonomous digital task forces and building at a faster velocity. Much more to come here. 

Google Search queries are at an all-time high, driven by AI; AI Overviews is driving overall search growth; AI Mode is seeing strong growth in both users and usage globally; management recently shipped agentic experiences in Google Search, such as restaurant booking, to new countries; management recently shipped the multi-modal capability, Search Live (where users can have voice conversations AI while sharing their phone’s camera feed to study surroundings), globally; search latency has been reduced by 35% in the past 5 years despite the new AI features introduced in Google Search; management has reduced the cost of responses by AI Overviews and AI Mode by 30% since they were upgraded to Gemini 3

I continues to drive search usage and queries are at an all-time high. We continue to invest in improvements to AI Overviews, which are driving overall search growth and we are also seeing strong growth in both users and usage of AI Mode globally…

…We also shipped agentic experiences like restaurant booking to new countries and new multimodal capabilities like Search Live globally…

…Even as we have brought new AI features into our results page, we have reduced search latency by more than 35% over the past 5 years. And since upgrading AI Overviews and AI Mode to Gemini 3, we have reduced the cost of core AI responses by more than 30%, thanks to continued hardware and engineering breakthroughs.

Alphabet’s management thinks a key point of Google Cloud’s differentiation is its 1st-party solutions across the enterprise AI stack; Google Cloud’s enterprise AI solutions became Google Cloud’s primary growth driver for the first time in 2026 Q1; revenue from products built on Alphabet’s GenAI models was up 800% year-on-year in 2026 Q1; new customer acquisition doubled in 2026 Q1 from a year ago; the number of $100 million to $1 billion deals doubled year-on-year in 2026 Q1; Google Cloud customers outpaced initial commitments by 45% in 2026 Q1, accelerating from 2025 Q4; Google Cloud recently introduced new capabilities across its vertical AI stack, including a new Gemini Enterprise AI Platform that helps users build and manage agents; Gemini Enterprise paid monthly active users was up 40% sequentially in 2026 Q1; the partner ecosystem for Gemini Enterprise had 9x year-on-year growth in 2026 Q1 in seats sold by partners and number of partners using Gemini Enterprise internally; 330 Google Cloud customers processed over 1 trillion tokens each over the last 12 months, with 35 processing over 10 trillion tokens each

Google Cloud is differentiated because we are the only provider to offer first-party solutions across the entire enterprise AI stack…

…Our enterprise AI solutions have become our primary growth driver for cloud for the first time. In Q1, revenue from products built on our GenAI models grew nearly 800% year-over-year. We are winning new customers faster with new customer acquisition doubling compared to the same period last year. We are seeing strong deal momentum, doubling the number of $100 million to $1 billion deals year-on-year and signing multiple $1 billion-plus deals…

…Customers outpaced their initial commitments by 45%, accelerating over last quarter.

At Cloud Next last week, we introduced hundreds of new capabilities across our vertically optimized AI stack that are designed to work together for our enterprise customers. We introduced a new Gemini Enterprise Agent Platform that empowers users to build, orchestrate, govern and optimize agents with the controls that enterprise customers need. Along with new capabilities in Gemini Enterprise app like Projects, Canvas, Long-Running agents and Skills, every employee can build agents.

In Q1, Gemini Enterprise paid monthly active users grew 40% quarter-over-quarter. That includes major global brands like Bosch, Citi Wealth, Merck and Mars Inc. Our partner ecosystem plays an increasingly critical role in driving Gemini Enterprise adoption. We saw 9x year-over-year growth, both in seats sold with partners and in the number of partners adopting it for internal use…

…Over the past 12 months, 330 Google Cloud customers each processed over 1 trillion tokens. 35 reached the 10 trillion token milestone.

Gemini is applied in Youtube for better matching and discovery between brands and creators; Gemini now powers YouTube Creator Partnerships; management has made it easier for advertisers to buy premium advertising space on Youtube; Supergoop! partnered with a YouTube creator for a Shorts and CTV campaign and it led to a 93% lift for a product and a 55% overall brand lift.

We are applying Gemini to drive better matching and discovery between brands and creators of all sizes. And Gemini now powers YouTube Creator Partnerships, a centralized platform integrated directly into YouTube Studio for creators and Google Ads for advertisers. 

We’ve also made it easier to buy premium ad space in top-tier podcast shows by curating the most watched podcasts into popular genres. For example, Supergoop! partnered with YouTube creator, Liza Koshy on a multi-format shorts and long-form CTV campaign, resulting in a 93% lift for their Glowscreen product and a 55% overall brand lift.

Waymo has so far launched in 6 new cities in 2026 and is currently in 11 major US cities; Waymo is now providing 500,000 rides per week (was 400,000 in 2025 Q4)

Waymo is on a great trajectory. It launched in Nashville a few weeks ago, that makes 6 new cities so far in 2026 and operations in 11 major U.S. cities in total. Waymo also surpassed 500,000 fully autonomous rides per week, doubling in less than a year.

Alphabet’s management is accelerating the deployment of Gemini across the company’s entire advertising infrastructure; the deployment of Gemini has led to new performance breakthroughs in advertising quality, advertiser tools, and new AI user experiences; Alphabet is making significant strides in improving relevance even when there isn’t a direct user query; advertising in Discover is getting better aligned with unique user interests; promoted pins in Maps are deeply relevant to user surroundings, location of interest, history and intent; Alphabet’s advertising relevance has increased by nearly 10%; Gemini is now powering Smart Bidding to more accurately match user intent to an advertiser’s product; management launched AI Max to help advertisers adapt to a new conversational way of searching by consumers; AI Max was moved out of beta earlier in April 2026; Hilton EMA used AI Max to capture 33% more clicks at 20% of the spend, and to increase average booking value by 55%; Etsy used AI Max to increase search volume by 10% with 15% of queries being net new; more than 30% of customer search spend now uses AI Max or Performance Max, and advertisers using the tools enjoy more conversions for the same spend; management is reinventing advertising formats for AI-native experiences; direct offers in AI mode are resonating with users; management is testing a new advertising format in AI Mode that displays retailers who sell recommended products in the AI Mode’s answer to a query; management launched Universal Commerce Protocol (UCP) in January 2026; UCP has new members consisting of major technology companies; brands such as Sephora, Macy’s and Ulta Beauty have already rolled out UCP; Ulta Beauty recently launched agentic commerce experiences in AI Mode and the Gemini App; management has received great feedback on UCP and they think UCP will power a new checkout experience in AI Mode, Search, and the Gemini app

We are accelerating the deployment of Gemini across our entire ads infrastructure to help businesses reach more customers in more places than ever before. This is driving significant improvements across all areas of marketing and continues to fuel new performance breakthroughs across 3 areas critical for our customers’ success, ads quality, advertiser tools and new AI user experiences.

First, ads quality. AI is boosting our ability to deeply understand user intent for a given search query and to find the most relevant ad. Even when we don’t have a direct user query, we’re making significant strides in improving relevance. In Discover, new AI models and classifiers are driving higher relevance by better aligning ads with unique user interests. In Maps, we’re using Gemini to ensure promoted pins are deeply relevant to user surroundings, location of interest, history and intent. This work is improving ads relevance by nearly 10%, leading to significant increase in user engagement. We’re pairing this strengthened prediction-driven relevance with bottom-of-funnel precision. Over the past year, we’ve made over 20 improvements to search and shopping bid strategies. Smart Bidding now uses Gemini to match user intent to an advertiser’s product and services more accurately and further drive performance. This level of granularity was previously impossible to achieve at scale.

Second, on advertiser tools, where Gemini helps advertisers drive more efficient and effective campaigns. People no longer search in fragments. They search conversationally and share more context. We launched AI Max to help advertisers adapt to this new way of searching. And earlier this month, it moved out of beta with improved performance quality across targeting and creative capabilities. Take Hilton EMA, they captured 1/3 more clicks for 1/5 of the spend while simultaneously increasing the average booking value by 55%. And Etsy saw a 10% search volume uplift with 15% of those queries being net new to their business. We see significant opportunity as advertisers continue to make good progress on AI readiness and the adoption of AI tools. For instance, more than 30% of our customer search spend now uses AI-enabled campaigns, AI Max or Performance Max. And these advertisers are seeing more conversion for the same spend.

Third, how we monetize new AI user experiences in search? We aren’t just bringing existing ad formats into AI experiences. We are reinventing ads for this new era. Direct offers in AI Mode are resonating with users and continue to receive positive customer feedback. Gap, L’Oreal and Chewy are just some of the latest partners who have now signed up to test this Google Ads pilot.

We’re also exploring new formats for retailers. AI Mode already surfaces organic product recommendations based on the user’s query and we’re now testing a new ad format that displays retailers who sell those recommended products. In addition, the retail industry is rapidly coalescing around the open source Universal Commerce Protocol, or UCP, we launched in January in partnership with the ecosystem. Last week, we welcomed Amazon, Meta, Microsoft, Salesforce and Stripe as new members to the UCP Tech Council. They joined founding members, Shopify, Etsy, Target, Wayfair and Google to further accelerate the transition towards an agentic future. Partners like Sephora and Macy’s have joined companies like Ulta Beauty, who are already rolling out UCP and can now redefine consumer journeys from discovery to checkout. Ulta Beauty just last week launched agentic commerce within AI Mode and Search and the Gemini app. Shoppers can now review product recommendations, compare options and complete streamlined checkout for eligible purchases directly within AI Mode and Gemini…

…We’ve received tremendous feedback so far from hundreds of top tech companies, payments partners, retailers, really interested in integrating. And it will help power a new checkout experience in AI Mode, in Search and the Gemini app and allowing shoppers to actually check out from select merchants, right as they’re researching on Google and going through this journey.

Google Cloud had 63% revenue growth in 2026 Q1 (was 48% in 2025 Q4) driven by growth in GCP; GCP grew at a much higher rate than Google Cloud’s overall growth; Google Cloud’s growth was driven by AI solutions and AI infrastructure; Google Cloud operating margin was 32.9% (was 30.1% in 2025 Q4 and was 17.8% in 2025 Q1); Google Cloud backlog grew nearly 100% sequentially to $462 billion in 2025 Q4 (was $240 billion in 2025 Q4); most of Google Cloud’s backlog are GCP contracts, and just over 50% of the backlog is expected to be recognised as revenue in the next 2 years; Google Cloud’s impressive margin improvement was driven by leverage from revenue growth, and management’s insistence on running an efficient organisation

 Cloud revenues accelerated across all key areas and were up 63% to $20 billion. Revenue growth was driven by strong performance in GCP, which continued to grow at a rate that was much higher than cloud’s overall revenue growth rate. The largest contributor to cloud’s growth this quarter was AI solutions, driven by strong demand for industry-leading models, including Gemini 3. In addition, we had strong growth in AI infrastructure due to continued deployment of TPUs and GPUs and core GCP continues to be a sizable contributor driven by demand for infrastructure and other services such as cybersecurity and data analytics. Workspace again delivered strong double-digit revenue growth, driven by an increase in the number of seats and the average revenue per seat. Cloud operating income was $6.6 billion, tripling year-over-year and operating margin increased from 17.8% in the first quarter of last year to 32.9%.

Google Cloud’s backlog nearly doubled sequentially, reaching $462 billion at the end of the first quarter. The increase was driven by strong demand for enterprise AI offerings and the inclusion of TPU hardware sales that Sundar referenced earlier. The majority of the backlog is related to typical GCP contracts and we expect to recognize just over 50% of the backlog as revenue over the next 24 months…

…[Question] There’s a thesis out there that AI revenues are a lower margin in general but we are seeing margins improve. So more insights on just the cloud business and what’s driving that margin expansion.

[Answer] There are pushes and pulls across the business, including within cloud specifically. And I would start with the top line. When we see this robust strong revenue growth, both in Cloud and Google Services, it does provide leverage all the way down to the bottom line within the income statement. And you know we’ve been working hard to ensure we have — we’re running a productive and efficient organization. And it’s not just how we operate the business but even in areas such as our technical infrastructure, where we are investing the significant CapEx investments in our data centers and servers, we are looking at how we drive scientific process innovation within that organization. And that is reflected both in Cloud and Google Services as we allocate costs based on based on consumption. In the past, I did talk about the depreciation associated with these investments that is hitting both Google Cloud and Google Services. Google Cloud expanded margin quite significantly from a year ago, as you’ve seen in our numbers that we’ve just previewed. And a lot of it, again, is the top line growth that Google Cloud is providing or producing as well as an incredibly efficient way of running the business.

Alphabet’s management has raised capex guidance for 2026 to $180 billion to $190 billion (was previously $175 billion to $185 billion; 2025’s capex was $91.4 billion, which was itself up 65% from $55.4 billion in 2024, and 2024’s capex was up 69% from 2023); management is seeing unprecedented demand for AI compute; Alphabet’s investments in AI compute are delivering strong growth; management expects 2027’s capex to be much higher than 2026’s; management is investing in capex based on tangible demand signals and a ROIC framework; Google Cloud remains constrained by supply and would have grown faster in 2026 Q1 if supply was higher

…We will begin to deliver TPU hardware to a select group of customers in their own data centers. We expect to begin recognizing a small percent of the revenues from these agreements later this year with the vast majority of revenues to be realized in 2027. It is important to keep in mind that revenues from TPU hardware sales will fluctuate from quarter-to-quarter, depending on when TPUs are shipped to customers…

…Wiz will be reported in the Google Cloud segment. And second, we expect a low single-digit percentage point headwind to cloud’s operating margin for the remainder of 2026 related to the acquisition…

…We are updating our full year 2026 CapEx guidance range to $180 billion to $190 billion, up from our previous estimate of $175 billion to $185 billion to now include investment related to the acquisition of Intersect, which closed in March.

We are seeing unprecedented internal and external demand for AI compute resources. The investments we are making in AI is delivering strong growth as evidenced by the record revenue and backlog growth in Google Cloud and strong performance in Google Services. Looking ahead, these strong results reinforce our conviction to invest the capital required to continue to capture the AI opportunity. As a result, we expect our 2027 CapEx to significantly increase compared to 2026. In terms of expenses, as we’ve discussed previously, the significant increase in our investment in technical infrastructure will continue to put pressure on the P&L in the form of higher depreciation expense and related data center operations costs such as energy. We also expect to continue hiring in key investment areas such as AI and cloud and are investing in marketing to support our AI products…

…You’ve seen us over the past several years increase CapEx every year. And we have done it very thoughtfully to meet the demand that we are seeing, both from external customers as well as demands across the organization. And you’re seeing the proof point, the ROIC on that in terms of just the growth rate we’re seeing, whether it’s growth rate within search or certainly the cloud business and the opportunity we have within the cloud backlog…

…I do think looking ahead, our ability to invest in this moment and stay at the frontier, I think puts us in a strong position. And I think we are doing it based on tangible demand signals we are seeing. And it’s not just on the revenue side but I’m talking from a ROIC framework and that’s what is helping us navigate this moment responsibly…

…We are compute constraint in the near term. And as an example, our cloud revenue would have been higher if we were able to meet the demand.

Amazon (NASDAQ: AMZN)

AWS grew 28% year-on-year in 2026 Q1 (was 24% in 2025 Q4) and is now growing at its fastest pace in 15 quarters; AWS’s run rate has reached $150 billion (was $142 billion in 2025 Q4); the last time AWS grew at a similar rate, it was half its current size; AI’s growth is unprecedented; the 1st 3 years of AWS’s AI revenue run rate was $15 billion, 260x larger than AWS’s run rate in its 1st 3 years; management thinks customers are choosing AWS for AI for 4 reasons, namely, (1) AWS’s broader capabilities, (2) customers want their AI inference to be at where their other applications and data reside, and this happens to be in AWS, (3) customers want to consume non-AI services as they grow their AI usage, and AWS has a broad set of offerings, and (4) AWS has the strongest security and operational performance; AWS has won many new enterprise customers since 2025 Q4’s earnings call, including OpenAI, Anthropic, Meta Platforms, and NVIDIA; AWS continues to see strong growth in non-AI workloads as enterprises focus on cloud migrations; management is seeing customers who want to benefit from AI accelerate their migration to the cloud; management is seeing a strong correlation in customers’ AI spend and core growth in AWS; AWS’s AI revenue is growing triple digits year-on-year; AWS operating income in 2026 Q1 was $14.2 billion, reflecting 37.7% operating margin (was 35.0% in 2025 Q4 and 39.5% in 2025 Q1); AWS’s backlog is $364 billion in 2026 Q1 with significant sequential growth (was $244 billion in 2025 Q4), and the backlog has reasonable breadth and does not include a recent $100 billion deal with Anthropic

AWS growth continued to accelerate, up 28% year-over-year, the fastest growth rate in 15 quarters, up $2 billion quarter-over-quarter, the largest Q4 to Q1 AWS revenue increase ever. AWS is now a $150 billion annualized revenue run rate business. It’s very unusual for a business to grow this fast on a base this large. And the last time we saw growth at this clip, AWS was roughly half the size. We’ve never seen a technology grow as rapidly as AI…

…3 years after AWS launched, it had a $58 million revenue run rate. In the first 3 years of this AI wave, AWS’ AI revenue run rate is over $15 billion, nearly 260x larger.

There are several reasons customers are choosing AWS for AI. First, we’ve built broader capabilities than others…

…Second and another reason customers continue choosing AWS is that as they expand their use of AI, they want their inference to reside near their other applications and data and much more of it resides in AWS than any place else. Third, as customers expand their AI usage, they also want to consume additional non-AI services, and they’re choosing AWS because we’ve built the broadest and most capable core offerings by a wide margin. We offer thousands of features across compute, storage, databases, analytics, security and more, and Gartner consistently recognizes AWS’ leadership across their major cloud evaluation areas. Fourth, AWS is the strongest security and operational performance of any AI and infrastructure provider and start-ups, enterprises and governments continue to choose AWS as the foundation for their most critical workloads…

…Since last quarter’s call, we’ve announced new agreements with OpenAI, Anthropic, Meta, NVIDIA, Uber, U.S. Bank, Fox, Southwest Airlines, U.S. Army, Bloomberg, Cerebras, AT&T, Nokia, Fundamental, The National Geographic Society, PGA TOUR and many more…

…Moving to our AWS segment. Revenue was $37.6 billion and growth accelerated 480 basis points to 28% year-over-year, driven by both core and AI services. We continue to see customers increase cloud migrations and scale their use of AWS core services. Customers seeking the full benefit of AI are accelerating their transition to the cloud. We also see a strong correlation between AI spend and core growth. As customers spend more on AI, we see a corresponding demand increase in core. We expect this to increase over time as customers move more AI workloads into production, strengthening demand for our core services…

…Our AI revenue is growing triple digits year-over-year…

…AWS operating income was $14.2 billion and reflects our strong growth, coupled with our focus on driving efficiencies across the business…

…The backlog for Q1 is $364 billion. That does not include the recent deal that we announced with Anthropic for over $100 billion. There’s reasonable breadth in that as well. It’s not just 1 customer or 2 customers.

AWS’s chips business, including Graviton and Trainium, grew 40% sequentially in 2026 Q1; the chips business is now at a $20 billion annual revenue rate (was $10 billion in 2025 Q4), and growing triple-digits; if AWS sold its chips as a stand-alone business, its annual revenue run rate would be $50 billion; AWS’s custom silicon business is now 1of the top 3 data center chip businesses in the world; Anthropic and OpenAI both recently signed very large multi-year commitments for Trainium; Trainium now has $225 billion in revenue commitments; Trainium 2 has 30% better price-performance than competitor GPUs and is largely sold out; Trainium 3, which only started shipping at the start of 2026, is 30%-40% more price-performant than Trainium 2 and is nearly fully subscribed; Trainium 4 is already been reserved despite being 18 months from broad availability; Amazon Bedrock runs most of its inference on Trainium; Meta Platforms has committed to using tens of millions of AWS’s Graviton CPUs; Amazon management sees massive demand for CPUs as agentic AI, post-training, and inference scales up; Graviton has 40% better price-performance than other x86 CPUs; Graviton is used by 98% of the top 1,000 AWS EC2 customers; AWS is bringing in more Trainium chips than NVIDIA GPUs, but NVIDIA remains an important partner; management expects Trainium to eventually save AWS tens of billions of dollars of capex annually and provide several hundred basis points of operating margin; management believes that people will always want choice in models and chips; management is currently not interested in selling Trainium racks to 3rd party data centers, but thinks AWS could do so in the next few years

Our chips business continues to grow rapidly and is larger than what a lot of folks thought. We saw nearly 40% quarter-over-quarter growth in Q1, and our annual revenue run rate is now over $20 billion and growing triple-digit percentages year-over-year…

…If our chips business was a stand-alone business and sold chips produced this year to AWS and other third parties as other leading chip companies do, our annual revenue run rate would be $50 billion. As best as we can tell, our custom silicon business is now one of the top 3 data center chip businesses in the world, the speed at which we’ve gotten here is extraordinary…

…We’ve recently shared very large multiyear, multi-gigawatt Trainium commitments from the 2 leading AI labs in the world in Anthropic and OpenAI as well as an increasing number of companies like Uber betting on Trainium. And we now have over $225 billion in revenue commitments for Trainium. Our Trainium2 chip has about 30% better price performance than comparable GPUs and is largely sold out. Trainium3, which just started shipping at the start of 2026 and is 30% to 40% more price performance than Trainium2 is nearly fully subscribed. And much of Trainium4, which is still about 18 months from broad availability has already been reserved. Amazon Bedrock, which is used expansively by over 125,000 customers, runs most of its inference on Trainium and almost 80% of the Fortune 100 companies are using Bedrock.

We also just announced that Meta is committed to using tens of millions of Graviton cores. Graviton is our industry-leading CPU chip, which allows Meta to run the CPU-intensive workloads behind agentic AI with the performance and efficiency they need at their scale. AI is commonly seen as a GPU story, but the rise of agentic workloads, real-time reasoning, code generation, reinforcement learning and multistep task orchestration is driving massive CPU demand as well. As AI systems shift from answering questions to taking actions and as post-training and inference scale up, the compute required pulls heavily on CPUs. That’s why Meta chose Graviton, which delivers up to 40% better price performance than any other x86 processors and now used by 98% of the top 1,000 EC2 customers…

…While the largest number of AI chips we’re bringing in are Trainium, we continue to have a deep partnership with NVIDIA. We have immense respect for them, continue to order substantial quantities. We’ll be partners for as long as I can foresee, and we’ll always have customers who want to run NVIDIA on AWS, and we will also have a very large chips business ourselves. Customers always want choice. It’s always been true and always will be true…

…At scale, we expect Trainium will save us tens of billions of dollars of CapEx each year and provide several hundred basis points of operating margin advantage versus relying on others’ chips for inference…

…But the one thing you learn over and over again with every technology, it was true in databases, it was true in analytics. It was true in models. It’s true in chips, too, by the way, is that customers want choice. There is not one tool to rule the world, and they want choice…

…On the question about Trainium and the notion of our selling racks over time, I do think that’s very much a possibility. Always, we have to balance — we have such demand right now for Trainium, and we have such demand from various companies who will consume as much as we make that we have to decide how much we’re going to allocate to the existing demand and customers and how much we’re going to save to sell as racks. And for our existing customers that we sell Trainium to, how many will be Trainium plus running on our cloud infrastructure versus just the chips themselves. But I expect over time, there’s a good chance we’re going to sell racks over the next couple of years.

Amazon’s management remain confident in the returns generated by the company’s capex; much of the capex spent in 2026 will be installed in future years; customers have already committed to substantial portions of the 2026 capex; management sees attractive margins and ROIC (return on invested capital) for the 2026 capex; AWS has to spend more short-term capex the faster it grows, since AWS needs to spend on land, power, chips etc 6-24 months in advance of monetisation; AWS’s capex often fund assets with years and decades of useful lives; AWS’s capex generate attractive cumulative free cash flow and ROIC a few years after being in service; Amazon’s free cash flow in the early years of high-growth periods for AWS is limited until the early capacity is monetized and revenue growth outpaces capex growth, and management has seen this cycle in AWS’s first big growth wave and expects similar positive outcomes from the current wave; management expects to continue making significant investments in AI; management has no change on Amazon’s 2026 capex plan (original guidance for 2026 was for $200 billion, and this is up from $128 billion in 2025, and $83 billion in 2024); management first saw the trend of rising input prices for capex in 2025 H2 and has been working with suppliers to get supply; management is seeing rising memory prices be a push-factor for companies to shift from on-premise to the cloud

We continue to be confident in the long-term CapEx investments we’re making. Of the AWS CapEx we intend to spend in 2026, much of which will be installed in future years, we have high confidence this will be monetized well as we already have customer commitments for a substantial portion of it and that it will yield compelling operating margins and ROIC…

…The faster AWS grows, the more short-term CapEx we will spend. AWS is to lay out cash for land, power, buildings, chips, servers and networking gear in advance of when we can monetize it, typically 6 to 24 months before we start billing customers depending on the component. However, these CapEx investments fund assets with many year useful lives, 30-plus years for data centers, 5 to 6 years for chips, servers and networking gear. The free cash flow and ROIC for these investments are cumulatively quite attractive a couple of years after being in service. However, in times of very high growth like now, where the CapEx growth meaningfully outpaces the revenue growth, the early years free cash flow is challenged until these initial tranches of capacity are being monetized and revenue growth outpaces CapEx growth. We’ve been through this cycle with the first big AWS growth wave and like the results. We expect to feel similarly about this next wave with much larger potential downstream revenue and free cash flow…

…We will continue to make significant investments, especially in AI, as we believe it to be a massive opportunity with the potential to drive long-term revenue and free cash flow…

…I don’t have an update on — a new update on capital. Our plan is largely the same…

…Everybody knows that the cost of these components, particularly memory has skyrocketed. And we’re just in a stage where there’s just not enough capacity for the amount of demand. We have worked very closely with our strategic partners. We saw this trend happening early in the kind of the middle of the latter part of last year, and we’ve worked with our strategic suppliers here to get a significant amount of supply. And so we’re working very closely with them. I think the team has been very scrappy. I think we’ve done a good job in making sure that we’re not capacity constrained there, but we’re watching that very closely.

One of the interesting things that we see right now with the change in price and in supply on things like memory is that it is a further impetus pushing companies who have on-premises infrastructure into the cloud. And it’s because a meaningful part, these suppliers are prioritizing their very largest customers which cloud providers are. And so we have seen a number of conversations we’ve been having with enterprises for many months where it’s just been slower in getting the transformation plan to move to the cloud accelerate rapidly just because we have a lot more supply than what others have.

SageMaker, AWS’s model-building service, reduces training time of models by up to 40%; Bedrock, AWS’s fully-managed service for companies to build upon frontier models, had 170% sequential growth in customer spend in 2026 Q1; Bedrock processed more tokens in 2026 Q1 than all prior years combined; OpenAI’s latest models are already, or will soon be, available on Bedrock; Amazon management recently added the Amazon Bedrock Managed Agents feature, which helps organizations build generative AI applications and agents at production scale;  Amazon Bedrock Managed Agents is powered by OpenAI, and OpenAI is seeing unprecedented demand for the product; Amazon management believes companies will derive the most value from AI from agents; Strands, AWS’s open source AI agents SDK (software development kit) has been downloaded more than 25 million times, with downloads up 3x sequentially in 2026 Q1; AgentCore is used to deploy an agent every 10 seconds; AWS has turnkey agentic solutions, including Kiro and Quick; Kiro, AWS’s coding agent, saw users double sequentially in 2026 Q1 and enterprise usage 10x; Quick, AWS’s AI assistant, has seen new customers grow 4x sequentially in 2026 Q1; management recently launched the Quick desktop app, which helps improve productivity of users; Amazon Bedrock now has 125,000 customers; 80% of the Fortune 100 are using Amazon Bedrock; AWS delivered 4x improvement in Trainium 2’s token throughput for Bedrock, leading to more capacity to serve customers; management thinks having OpenAI’s models on Bedrock is a big deal; Bedrock is already serving 3rd-party models from all the non-OpenAI key players; management believes that people will always want choice in models and chips; management believes that most of the work being done with models in the future will be of the stateful variety; Bedrock Managed Agents is a feature unique to AWS 

We’ve built broader capabilities than others. That includes model building with SageMaker, which reduces training time by up to 40%, high-performance inference with the leading selection of frontier models in Bedrock, which saw 170% growth in customer spend quarter-over-quarter and processed more tokens in Q1 than all prior years combined.

We’re excited to make OpenAI’s models available in Bedrock. Yesterday, we added OpenAI’s GPT-5.4 model with 5.5 coming soon. Yesterday, we also started the preview of Amazon Bedrock Managed Agents powered by OpenAI, the Stateful Runtime Environment that enables any organization to build generative AI applications and agents at production scale. We believe that modern agentic applications will be stateful, and this new technology will rapidly accelerate agentic AI adoption. OpenAI has said they’re already seeing unprecedented demand for this new product, and we’re seeing heavy customer interest as well.

Most of the value companies derive from AI will be through agents. In AWS customers can build agents with their proprietary data and Strands, which has been downloaded more than 25 million times and saw 3x more downloads quarter-over-quarter. Customers can deploy agents with enterprise scale, security and reliability with AgentCore, which is being used to deploy an agent as frequently as every 10 seconds. We also offer turnkey agents for coding, software migrations, business operations and knowledge workers in Kiro, Transform, Connect and Quick, and they continue to resonate with customers. The number of developers using Kiro more than doubled quarter-over-quarter and enterprise customer usage increased nearly 10x. Customers have used Transform to save over 1.56 million hours of manual effort when migrating and modernizing their workloads. The number of new customers using Quick has grown more than 4x quarter-over-quarter, and we just announced our Quick desktop app yesterday. It’s very compelling as it can query your e-mail, calendar, Slack, local files and several other applications you use every day to flag important communications, retrieve and summarize information, make recommendations, compose and send communications to others and create agents that highlight or automatically do work that you used to have to do yourself. You can easily keep refining your preferences and Quick’s advanced knowledge graph enables its AI agents to automatically learn from your interactions to become more personalized over time…

…Amazon Bedrock, which is used expansively by over 125,000 customers, runs most of its inference on Trainium and almost 80% of the Fortune 100 companies are using Bedrock…

…Bedrock has been a significant growth driver. In 2025, we delivered 4x improvements in Trainium2’s token throughput. And since the majority of Bedrock’s workloads run on Trainium, these efficiency gains directly translate into more capacity to serve customers…

…The fact that we’re going to have all of the OpenAI models available in Bedrock is a big deal. It’s a big deal for customers. And we have — we obviously have a very large amount of AI being done in Bedrock today on the models we have and this is Anthropic and Llama and Mistral and a host of others. But the one thing you learn over and over again with every technology, it was true in databases, it was true in analytics. It was true in models. It’s true in chips, too, by the way, is that customers want choice. There is not one tool to rule the world, and they want choice…

…Most of the model work and most of the AI has been done in these stateless models, kind of tokens in and tokens out. And while I think there will continue to be lots of work done that way, I think the future of using these models is a stateful model, a stateful API. And that’s because when you’re building agents, you’re building AI applications, you don’t want to start a new every time you interact with the model. You want to store state. You want to store identity, you want to store what the conversation or the actions have been, you want to reach out and do a little bit of compute here. You want to have the tools to be able to reach — the models reach out to the different tools to accomplish different tasks. And that only happens if you’re able to store state. And so the Bedrock Managed Agents that we collaborated with and invented with OpenAI that we just announced a preview of yesterday is also — I think that’s the future of how these agents are going to be built. It’s something that nobody else has, and I think it’s very exciting to our customers.

Amazon is able to deliver items faster while lowering its cost to serve, and management sees meaningful opportunities to further improve the fulfillment network’s productivity; Amazon’s latest generation of robotics offers a step change in efficiency; management is deploying the latest generation of robotics in both new and existing fulfillment facilities, and early results are positive

Overall unit growth of 15% continues to outpace our cost to operate the fulfillment network as outbound shipping costs grew 12% year-over-year and fulfillment expense grew 9% year-over-year, both on an FX-neutral basis. As our network efficiency improves, we’re able to deliver items faster and improve the customer experience while at the same time lowering our cost to serve. Looking ahead, we see meaningful opportunities to further enhance productivity across our global fulfillment network, all while continuing to raise the bar in delivery speed. We will keep optimizing inventory placement to shorten distance traveled, reduce touches per package and improve consolidation rates.

Alongside these efforts, we deploy robotics and automation, which have been integral to our operations for decades. Our latest generation technologies offer a step change in efficiency, which we’re deploying in both new and existing facilities. All of our U.S. large-format fulfillment center launches in 2026 will have this latest generation technology. We’re seeing early positive results with improved site safety, higher productivity and lower cost to serve.

Amazon management recently launched Health AI, a personal health agent

We launched Health AI, a 24/7 AI-powered personal health agent backed by One Medical clinicians that gives U.S. customers instant clinical guidance and takes action with their permission from booking appointments to managing prescriptions to facilitating medical treatment with a real One Medical provider.

Rufus, Amazon’s AI shopping assistant, saw monthly active users grow 115% year-on-year in 2026 Q1, and engagement increase by 400%; Rufus has improved a lot over the past year

Rufus, our agentic AI shopping assistant continues to resonate with customers. Rufus can research products, track prices and auto buy products in our store when they reach a set price. Monthly active users are up over 115% and engagement is up nearly 400% year-over-year…

…If you haven’t checked out Rufus in a while, it’s really substantially improved over the last year.

Amazon management recently launched Seller Central, an AI-powered insights-hub for sellers on Amazon; the initial response to Seller Central has been very strong

We recently introduced a new AI experience for sellers in Seller Central that dynamically generates a custom, personalized visualization of data, key insights and scenarios tailored to the sellers’ goals. It’s early, but the initial response and feedback are very strong.

Amazon’s management recently expanded Creative Agent to more countries; Creative Agent is Amazon’s agentic offering that helps advertisers plan and execute the entire advertising creative process; management recently launched sponsored products and brand prompts in Rufus; 20% of shoppers interacting with brand prompts in Rufus carry on the conversation

Our Ads team also continues to invent and deliver for advertisers with AI. For example, we expanded Creative Agent, an agentic partner that plans and executes the entire ad creative process to Canada, France, Germany, India, Italy, Spain and the U.K. And we recently introduced Sponsored Products and Brand Prompts in Rufus that help brands showcase products and customers make more informed buying decisions. It’s early, but we’re seeing nearly 20% of shoppers who interact with the Brand Prompts in Rufus continue the conversation about that brand.

Amazon’s management recently expanded early access to Alexa+ to Mexico, UK, Italy, and Spain; compared to the previous Alexa, users are completing 3x more purchases on device, streaming 25% more music, and using smart home functionality 50% more 

Alexa+ early access expanded to millions more Prime members in Mexico, the U.K., Italy and Spain. Customers are loving Alexa+, talking to Alexa twice as much and for longer durations across a wider breadth of topics, completing purchases on devices 3x more, streaming music 25% more and using smart home functionality 50% more than Alexa classic.

Amazon’s management continues to be very bullish on agentic commerce; management thinks agentic commerce will be very good for customers and Amazon in the long run; agentic commerce is currently only a small fraction of referrals from search engines; management thinks the user-experience with agentic commerce from 3rd-party agents is still poor, as pricing and product information are often wrong, and the agents don’t have personalization data and shopping history; management is working with 3rd-party agent providers to improve the experience; management continues to think that the agentic shopping assistant that will prevail will come from existing retailers that customers already have a good relationship with, and management is attempting to build Rufus to be the prevailing agentic shopping assistant; management thinks agentic commerce will be a great thing for Amazon’s advertising business because of 2 reasons, namely, (1) agentic AI will drive greater volume of advertising, and (2) agentic commerce provides multiple opportunities to surface relevant products to customers

We are very bullish on what agentic commerce will look like. I think it’s going to be very good for customers in the long term. I think it will be good for us, too…

…We’ll do a lot of work with third-party horizontal agents to try and make that customer experience better. And by the way, I do think today, it reminds me in some ways the stage we’re in of what we saw in the early days of search engines and they’re trying to refer business to e-commerce. It’s never been a giant part of the referrals to our e-commerce business. But over the years, the experience got better. And what you see with agentic commerce is it’s a small fraction of what we see with the search engine referrals, but the experience just hasn’t gotten great with these third-party horizontal agents yet. They’re not often able to get the pricing right or the product information right. They don’t have any personalization data or any shopping history. And so we do want to see that get better with third-party horizontal agents. We’re having conversations with all those folks to try and make that better and find something that works for customers and all the companies.

And then it will be interesting over time which agents customers choose to use. I happen to think that if you’re going to a particular retailer that you’d like to do business with and you like to shop from, if they have a great agentic shopping assistant, you’re going to often start there because it’s where you’re doing your shopping, it’s easier to — they have better product information. They have better information about what other customers like you are buying. You can make all sorts of changes to how your account and your shipping information is working there. And so that’s what we’re aiming to make Rufus be is we’re aiming to have it be the best shopping assistant anywhere, and I think we’re on that path…

…On the Agentic Commerce and how that impacts advertising, I actually believe that we’re going to like this for advertising. I think it’s going to be good for customers, and it’s going to be good for our business. And I think, first of all, the first thing to remember is the way that our ads team has built tools and agents themselves is making it so much easier to do advertising. If you look at small and medium-sized businesses that had to take weeks and months to do creative and to pick the right audience, all of that is just — it’s so much faster and so much easier because of our advertising agentic tools. And you no longer have to take as much time or spend as much money building the creative.

So I think there are going to be a lot more advertising — advertisers with the rise of what’s happening in AI. And then if you look at the Agentic Commerce experiences, if you look at any of these agentic experiences, they tend to be multi-turn conversations where you’re not interacting with one search and getting an answer. You tend to find that you’re asking questions, you’re narrowing questions, it’s asking you questions on what you want. And in that process of having multi turns, there are multiple opportunities to surface relevant products to customers, many of which will be organic and some of which will be sponsored. And it also gives rise to opportunities like sponsored prompts.

In the 2025 Q4 earnings call, Amazon’s management said market demand for AI compute looked like a barbell with AI labs on one end spending a lot on compute for just a handful of applications, and with enterprises on the other end using AI for productivity purposes; now, management is starting to see enterprises using AI for brand-new experiences

The AI labs are spending an incredible amount of money on compute at this point and in compute, both on the AI side as well as on the core side. And the models that they’re building and the companies that have successful generative AI applications are certainly spending a lot. And there are several of those labs. But we also see quite a bit of enterprise adoption and usage of AI. As I’ve said before, the largest absolute place that we see enterprises having success is in projects that are around cost avoidance and productivities. These are things like automating customer service or business process automation or fraud or things of that sort. But the number of projects that we’re working with across enterprises and that we’re now starting to see to come to production around brand-new experiences, trying to figure out how to reinvent their current experiences, but using inference and AI to be smarter, also very significant. So we’re seeing the adoption in both of those segments.

Amazon’s management sees a giant impact on how AI will shape Amazon’s business internally; management believes AI will completely reinvent Amazon’s current customer experiences in the fullness of time; management is aware of the innovator’s dilemma that can trap Amazon in reinventing AI-native customer experiences, and is actively avoiding the trap; Amazon swapped the engine of a service running at full tilt with a team of just 5 people who used agentic coding tools to build the new engine in 65 days; the engine would previously have taken 40-50 people a year to rebuild

On the use of AI internally and for our current businesses, I think that the shortest first summary I could give you, Colin, is that I do not see a place in any of our businesses or any of the ways that we do work where we’re not going to have giant impact on what we do. I think I’ve long had this belief that while you can add incrementally to a lot of your existing customer experiences, different agentic and AI experiences, I really believe that in the fullness of time, and I don’t know if that’s 3 years from now or 5 years from now or it could be sooner, too, that all of these customer experiences we know are going to be completely reinvented…

…It’s tricky for — if you have an existing business that’s doing well. But you have to look at every single one of your customer experiences and you have to be able to carve off resource for that team to think anew about what would the future customer experience look like if you started from scratch today, and if you had all the technologies like AI available to you when you started. And that is what we’re doing in every single one of our experiences…

…If you look at one of our services, we swapped out the engine of the service while we are also running the service full tilt. And normally, that would have taken 40 or 50 people about a year to do, and we took 5 really smart people, AI forward-thinking people building on agentic coding tools and those 5 people rebuilt it in 65 days. Like that is a very different world of operating. And that’s the world I think we’re heading to over the next few years.

Apple (NASDAQ: AAPL)

The iPhone 17 family contains the A19 and/or the A19 Pro chips, which include neural accelerators to deliver strong AI capabilities

During the quarter, we welcomed iPhone 17E, the newest addition to what is already the strongest iPhone lineup we’ve ever had. It brings outstanding performance and core iPhone experiences at a remarkable value for everyone from enterprise teams to consumers. Across the lineup, this is the most powerful, capable and versatile iPhone family we’ve ever created. That starts with the latest in Apple silicon for iPhone, A19 and A19 Pro, which include neural accelerators in the GPU to deliver a huge boost to AI performance

Apple’s management thinks the Mac is the best platform for AI, with Apple’s in-house chips giving Macs the ability to run advanced AI models on-device; the MacBook Air now comes with the M5 chip, which enables the product to run AI models on device; the MacBook Pro has even more advanced versions of the M5 chip in M5 Pro and M5 Max

From Mac Mini to MacBook Pro and everything in between, Mac is the best platform for AI with Apple Silicon delivering exceptional performance, industry-leading efficiency and the ability to run advanced models locally in ways that simply weren’t possible before…

…We’ve also further improved MacBook Air, already the world’s most popular laptop with M5, making everyday tasks faster and more responsive than ever. MacBook Pro reaches new heights with M5 Pro and M5 Max, delivering extraordinary performance and dramatically advancing what users can do with AI on a portable system…

Apple’s new AirPods Max 2 has Apple’s most advanced active noise cancellation technology; AirPods can now do live translation, thanks to Apple Intelligence

During the quarter, we introduced customers to a new level of audio experience with AirPods Max 2, delivering stunning sound quality and our most advanced active noise cancellation yet…

…AirPods can bridge languages too, thanks to Live Translation powered by Apple Intelligence.

Apple Intelligence now has more powerful capabilities such as visual intelligence for cleanup; management is looking to launch a more personalised Siri later in 2026 ; Apple Intelligence is powered by Apple’s self-designed chips; management is not treating AI as a standalone feature but is instead treating AI as an essential experience

In addition to live translation, Apple Intelligence brings together dozens of powerful capabilities from visual intelligence to cleanup and photos that are seamlessly integrated into the moments that matter most to our users every day. And we look forward to bringing a more personalized Siri to users coming this year. What truly sets Apple apart is how Apple Intelligence is woven into the core of our platforms, powered by Apple Silicon and designed from the ground up to deliver intelligence that is fast, personal, and private. This is not AI as a stand-alone feature, but AI as an essential intuitive part of the experience across our devices. It builds on years of innovation from the neural engine to advanced on-device processing, enabling capabilities that are not only incredibly powerful, but also respectful of user privacy.

Reminder that in 2025, management committed to invest $600 billion over 4 years (was a $500 billion commitment in 2025 Q2; Apple has around $190 billion in gross profit per year, for perspective) in the USA in areas such as advanced manufacturing, silicon engineering and artificial intelligence; Apple now has Mac mini production in the USA; in March 2026, management brought 4 new companies to Apple’s American manufacturing program; Apple is on track to buy over 100 million advanced chips from TSMC’s Arizona fab; later in 2026, Apple will open its advanced manufacturing center in Houston to provide hands-on training for students, supplier employees and American businesses

We’re also making great progress in advancing American supply chain innovation. As part of our $600 billion commitment to the U.S., we were pleased to share recently that Mac mini production is coming to America later this year, expanding our factory operations in Houston with a brand-new facility. In March, we were thrilled to welcome 4 new companies to our American manufacturing program to help manufacture essential materials and components for Apple products sold worldwide. These include sensors that support key iPhone features like camera stabilization and integrated circuits essential for features like crash detection and activity tracking. These efforts build on the progress we’ve made in the American manufacturing program, including the work we’re doing to advance an end-to-end silicon supply chain across the U.S. At TSMC’s Arizona facility, for example, Apple is on track to purchase well over 100 million advanced chips.

As we’re accelerating our long-standing support for U.S. innovation, we’re also investing in America’s workforce. We’re looking forward to opening the doors to an all-new advanced manufacturing center in Houston later this year, which will provide hands-on training led by Apple experts and tailor-made for students, supplier employees and American businesses.

The Mac Mini and Mac Studio models are great devices for AI and agentic AI, and so demand from consumers was greater than management expected; management thinks the supply constraints with the Mac Mini and Mac Studio will take a few months to resolve; management’s guidance for 2026 Q2 already embeds significantly higher memory costs; management thinks memory costs will have an increasing impact on Apple’s business

You look forward to the June quarter, the majority of our supply constraints will be on several Mac models given the continued high levels of demand that we’re seeing. And we have less flexibility in the supply chain than we normally would. For Mac, in the June quarter, there’s 2 factors that are driving the constraints. One is that on the Mac Mini and the Mac Studio, both of these are amazing platforms for AI and Agentic tools. And the customer recognition of that is happening faster than what we had predicted. And so we saw higher-than-expected demand. The second reason is that the customer response to Mac Neo has just been off the charts, with higher-than-expected demand…

…We think looking forward that the Mini and the Mac Studio may take several months to reach supply-demand balance…

…I’ll go back to December for a moment and just walk you through the chronology. In the December quarter, we really had a minimal impact due to memory, and you can kind of see that in the gross margin results. We said it would be a bit more in the March quarter, and we did see higher memory costs in the March quarter, and they were partially offset by benefits from carry-in inventory that we had. For the June quarter and what’s embedded in the guidance that Kevan went through earlier, we expect significantly higher memory costs. They are also partly offset by the benefit of carry-in inventory. And then where we don’t give color beyond June, I can tell you that beyond the June quarter, we believe memory costs will drive an increasing impact on our business.

Apple’s management has been investing more in AI in both products and services, and this shows up in the company’s operating expenses, specifically in R&D (research and development); the increased investments in AI include building Apple’s own foundation models, and in the collaboration with Google; Apple’s collaboration with Google on foundation models is going well

[Question] As we think longer term, do you think Apple will invest more? Where will Apple invest more heavily over the next several years? And is this at all related to your net cash comments in terms of perhaps building out more infrastructure as we enter an AI-centric world?

[Answer] We are clearly investing more. You can see that in the OpEx numbers. And if you click down on those a step deeper and look at the R&D area separate than SG&A, you’ll find that R&D is even accelerating much higher than the company is. And so we are clearly investing. We’re investing in products and services, and we see opportunities in both of those…

…We believe AI is a really important investment area for Apple, and we’re going to be doing that incrementally on top of what we normally invest in our product road map…

…[Question] Last quarter, you did talk about Apple foundational models and sort of the two-pronged strategy there of the collaboration with Google as well as continuing to internally sort of work on your own models. Hoping you can sort of give us an update in terms of how you’re able to balance those 2 priorities as well as do you feel like you need to double down and invest more to be able to balance those 2 priorities side by side?

[Answer] We are investing more. You can see that in the OpEx numbers. And as I’ve mentioned before, the R&D, in particular, is — has scaled rather significantly on a year-over-year basis. The collaboration with Google is going well. We’re happy with where things are, and we’re happy with the work that we’re doing independently as well.

ASML (NASDAQ: ASML)

ASML’s management is seeing the semiconductor industry’s growth continue to solidify, driven by AI investments, and this applies to both advanced Memory and advanced Logic; management thinks semiconductor supply will not meet demand for the foreseeable future, and this is creating constraints in end markets, including AI; management is seeing ASML’s Memory customers being asked to ramp supply; ASML’s memory customers are sold out for 2026, with supply constraints extending beyond the year; management is seeing ASML’s Logic customers building capacity, including for the 2nm node to meet AI demand and mobile demand; management is seeing ASML’s customers increasing their capital expenditure to ramp up their capacity, and this capacity is supported by long-term commitments from their customers; management is seeing ASML’s Memory customers and Logic customers increase their adoption of EUV and DUV immersion lithography; the level of demand for ASML’s DUV immersion lithography systems in 2025 was significantly lower what’s currently seen; besides DUV immersion, management is also seeing health in the DUV dry lithography business; management has seen major adoption of EUV by ASML’s DRAM customers in 2025 because EUV provides better performance; DRAM has been a really good story for lithography intensity in 2025; ASML’s customers have been very open with the company on their expansion plans

We see that the semiconductor industry growth continues to solidify. This is still very much driven by investments in AI infrastructure. So, this translates into a lot of demand for advanced Memory, for advanced Logic. We expect in fact that the supply will not meet the demand for the foreseeable future. So, this is creating a strong constraint in the end markets from AI to mobile and PC. As a result our customers are strongly invited to create more capacity. So if we look at Memory, what our customers tell us is that they are sold out for 2026. And their supply constraints will last beyond 2026. For advanced Logic, we see our customers building capacity for several nodes, while they also continue to ramp 2 nm in order to address the AI products…

…We see our Memory and Logic customers increasing their capital expenditure and trying to accelerate basically their capacity ramp in 2026 and beyond. What’s also very interesting is that a lot of this demand is supported by long-term commitment from their customers. On top of that, we see both Memory customers, DRAM customers and advanced Logic customers continuing to increase their adoption of EUV, but also immersion. So this translates basically into higher lithointensity and a higher litho demand for ASML…

…When it comes to immersion DUV, we actually had a bit of a slow start because in the course of last year, we were looking at a significantly lower demand for immersion. That has now reversed itself…

…I already mentioned what we’re doing on immersion, but also the dry business is doing quite nicely…

… In the Logic business, our customers are adding capacity across multiple advanced nodes to support demand while continuing to ramp the 2-nanometer node in support of next-generation HPC and mobile application…

…We have seen a major adoption of EUV in DRAM in 2025. And you may have noticed that our, I will say, U.S. DRAM customer also made this announcement that they were shifting also pretty strongly on EUV. And the reason for that is, of course, performance, but it’s also capacity because if you are going to use more EUV layers, you are going to need less multi-patterning and multi-patterning takes a lot of space also in the fab. So I think this is also definitely another argument in favor of EUV. I think this was mentioned, by the way, by this U.S. customer in their call. So I would say the first results of that is, first, more adoption of Low NA EUV…

…DRAM has been really a good story when it comes to litho intensity in ’25…

…Customers are very, very open. By the way, that’s also the case on the Logic side. But very — customers are very open to us, and they’re very openly discussing with us also their expansion plans for this year, but also beyond.

ASML’s management does not want EUV systems to be the bottleneck in building compute capacity for AI; EUV systems are not the bottleneck today

We do not want EUV to be the bottleneck. So I think I’d like to say that very, very strongly…

…I know the question of bottleneck comes back very often. I think we don’t feel at all that we are the bottleneck today.

Intel (NASDAQ: INTC)

Intel’s management expects sustained momentum for the company’s Xeon server CPU products in 2026 and 2027, with the Xeon 6 being Intel’s fastest new product ramp in 5 years alongside the Core Series 3 products; Xeon’s momentum is powered by the reinsertion of CPUs as a foundation for AI where the CPU-to-GPU (accelerators) ratio is swinging back to the CPU’s favour; management thinks the CPU’s resurgence in AI is great news for Intel’s x86 CPU ecosystem; Intel saw strong ASIC growth in 2026 Q1 sequentially and year-on-year; Intel’s DCAI (Data Center and AI) segment, signed multiple long-term agreements in 2026 Q1; Xeon 6 was recently selected as the host CPU for NVIDIA’s DGX Rubin NVL8 systems; Xeon remains the most deployed host CPU for AI systems; DCAI recently started a multiyear collaboration with SambaNova to design a next-generation AI inference architecture; management’s confidence in the sustained growth of CPUs for AI is growing; management’s outlook for server CPU demand has improved in 2026 Q1; management expects the server CPU industry to have a strong year of double-digit unit growth in 2026, extending to 2027; the long-term agreements signed by DCAI have volume and pricing terms, and last 3-5 years; Intel’s customers are telling the company that CPUs are more important in AI inferencing and agentic AI than AI training, with the ratio of GPUs-to-CPUs flipping from 8:1 to possibly 1:more-than-1; management believes Intel’s CPUs will be very effective competitors to the likes of ARM, AMD, and the hyperscalers

Demand continues to run ahead of supply for all our businesses, especially for Xeon server CPUs, where we expect sustained momentum this year and next. Intel 3-based Xeon 6 and Intel 18A based Core Series 3 products are now in full volume production ramp and each represents the fastest new product ramp in 5 years…

…For the last few years, the story around high-performance computing was almost exclusively about GPU and other accelerators. In recent months, we have seen clear signs that the CPU is reinserting itself as the indispensable foundation of the AI era. CPU now serves as the orchestration layer and critical control plane for the entire AI stack. This is not just our wishful thinking, it is what we hear from our customers, and it is evident in the demand profile for our products. Xeon server demand is seeing strong and sustained momentum. Customers are deploying server CPUs along accelerators in the ratio that is moving back towards CPU. The accelerator remains central to Frontier AI, and we will continue to participate, innovate and partner in that category. Our recent announcement with SambaNova Systems is an example of such partnership on heterogeneous compute architectures. But the backbone of AI computing in production remain a CPU anchored architecture. That is good news for the x86 ecosystem. It is great news for Intel…

…We also saw strong ASIC growth with revenue up more than 30% sequentially and nearly doubling year-over-year…

…Within the quarter, DCAI signed multiple long-term agreements, including Google, supporting our view that the current business momentum is sustainable. In addition, Xeon 6 was selected as the host CPU for NVIDIA’s DGX Rubin NVL8 systems, and Xeon remains the most deployed host CPU due to its industry-leading memory, security and networking orchestration. Lastly, DCAI also established a multiyear collaboration with SambaNova to design a next-generation heterogeneous AI inference architecture combining SambaNova’s RDUs and Intel Xeon 6 processors…

…Our confidence in the sustained growth of CPUs driven by the AI infrastructure build-out is growing. Our outlook for server CPU demand has improved over the last 90 days, and we expect a strong year of double-digit unit growth for the industry and for us with momentum extending into 2027…

…Most of these agreements are structured with volume and pricing, and they are usually somewhere between 3 and 5 years…

…The feedback from the customer, CPU is very important when you move from training to inference. Inference side, I think in terms of orchestration, control plane and also managing all the different agent with data, CPU is much more efficient. So I think the ratio of CPU to GPU used to be 1 and 8, and now it’s 1:4 and I think towards parity or even better…

…One statistic that we look at is the ratio of CPUs to GPUs. And if you look at training solutions, they’re generally running in the kind of 7 to 8 GPUs to 1 CPU. As we look into inference, it’s probably getting into like the 3 to 4:1 kind of level. And as you get into agentic and multi-agent, it’s one potentially even flip in the other direction a little bit…

…[Question] On server CPU competition. So both when we look at competition versus x86 against AMD, do you think you are gaining share? Do you expect to gain share against them? And then broader, I think the competition against Arm because NVIDIA is planning to launch a stand-alone Vera CPU Rack. Recently, we heard Amazon talk up their Graviton option. I think Google yesterday said they would launch Axion and connect it with every TPU. So just kind of near term, how do you look at competition versus AMD and x86?

[Answer] The CPU is a great demand right now. I think we all enjoy that. And then in terms of our product road map, we have been fine-tuning the last year… We are laser-focused on execution. Multithreading, I think we are putting in. So we’re going to have Coral Rapid, have the multithreading that we can compete effectively with AMD. And we try to accelerate that Coral Rapid ahead. And then the other part is we’re also looking at some of the architecture, CPU and GPU architecture… In all, I think we have the team, we have the technology road map. I think we’re going to be — over time, going to be a very effective competitors to them.

Intel’s management sees the semiconductor industry’s addressable market approaching $1 trillion, driven by AI demand, and the company is well positioned to benefit

Driven by tremendous demand for AI, the semiconductor industry TAM is now approaching $1 trillion. Intel is well positioned to benefit from this demand with 3 strategically important assets: our x86 CPU franchise, our advanced packaging technology and our vast manufacturing network.

Intel’s management sees AI moving into the real world, with more distributed inference

Artificial intelligence is now moving into the real world towards a more distributed inference and reinforced learning workloads like agentic, physical AI and robots and edge AI.

Intel’s management is pleased with the progress of the company’s foundry technology development, but it will be a long journey; the manufacturing yields of the Intel 3 and Intel 18A process technologies are now running ahead of management’s projects; Intel continues to make progress in advanced packaging technologies, with additional customer backlog growth in 2026 Q1; Intel’s 14A process technology is now at a higher level of yield compared to 18A at a similar point in time, and the company is developing PDKs (process design kits) with multiple customers; management expects to see design commitments for 14A in 2026 H2 and 2027 H1; the progress of Intel Foundry has driven the company to land more of its own future product tiles on the Intel 14A process; Intel Foundry will be supporting TeraFab, the huge semiconductor project undertaken by Elon Musk’s companies; management wants to work with TeraFab to improve the manufacturing efficiency of semiconductors; rising prices for memory chips and other materials are a headwind for Intel Foundry’s gross margin in 2026 H2; management will continue to utilise a multi-foundry approach for Intel; Intel Foundry’s advanced packaging business is seeing demand in the billions of dollars; Intel Foundry’s advanced packaging is a differentiated offering – it allows customers to use larger reticles – and so it’s getting attractive pricing; Intel Foundry’s 18A yields are going to hit management’s end-2026 targets by the middle of the year; most of Intel Foundry’s supply is for internal demand at the moment, but management expects it to win customers over time

The accelerating deployment of AI infrastructure creates a meaningful opportunity for us as we continue to build our external foundry business. I’m pleased with the progress we have made in foundry technology development over the last year, even though I will continue to remind you this will be a long journey for us. We have made steady progress with Intel 4 and Intel 3 and 18A yields are now running ahead of the internal projections, representing a meaningful inflection in our execution and our factory finished good output.

We also continue to make steady progress on our advanced packaging technologies, including additional growth in customer backlog in the quarter.

Intel 14A maturity yield and performance are outpacing Intel 18A at a similar point in time, and we continue to develop PDKs with multiple customers actively evaluating the technology…

…We expect to see earlier design commitments emerge beginning in the second half of 2026 and expanding into the first half of 2027…

…I’m particularly pleased that our progress today has driven us to land more of our own future product tiles on Intel 14A as well. At a time when advanced wafer capacity is in the short supply, this enables us to have better control over our supply chain…

…As we look to continue challenging the status quo, I can think of no better partners than Elon Musk. We recently announced our partnership with SpaceX, xAI and Tesla to support Terafab. Elon and I share a strong conviction that global semiconductor supply is not keeping pace with the rapid acceleration in demand. We are excited to explore innovative ways to refactor silicon process technology, looking for unconventional ways to improve manufacturing efficiency that will eventually lead to a dynamic improvement in the economics of semiconductor manufacturing…

…Our foundry team is delivering consistent yield and throughput improvements across all process nodes, which will help gross margins. With that said, Intel 18A is still early in its ramp and rising input costs, especially in memory, present growing headwinds in the second half that we need to overcome…

…I’d say the one cautionary concern I have on gross margin in the back half of the year is just some of the materials have gone up in terms of cost, substrates are going up, T glass. We’ve got memory going up, as you know. So those things offset some of the improvements that we’re having through the year…

…TSMC is a very important partner for us. Morris and C.C. have decades of friendship. And then clearly, with our product group will decide which is the best foundry. So I think we’re going to use a multi-foundry approach, our own internal and also external. And so we really have good relationship, continue to build from both sides to benefit the customer…

…[Question] I would love to kind of level set where we are on the advanced packaging front. You talked about rising backlog. Anything you can share in terms of what that number looks like?

[Answer] We have been really pleased with our traction there. And I think maybe naively, I had thought that these opportunities would come in the hundreds of millions of dollars level. But so far, what we’re seeing is that their demand is more in the billions of dollars per year kind of level. So this is going to be a big part of the foundry revenue as we get through this decade. And the good news is advanced packaging really is a differentiated offering for us, and it does a lot for the customer in terms of allowing them to use larger reticles. So there’s real value to the customer. And as a result, we get very attractive pricing relative to some of the other areas of the foundry business…

…18A yields are somewhat a closely guarded proprietary piece of information for us. So we don’t typically — I would just say Lip-Bu had a target as we came into the year for the end of this year, and we’re probably going to hit that probably the middle of this year…

…[Question] As we think about your capacity tightness, the leading edge foundries are also quite tight as well. Has this driven any near- to medium-term share gains?

[Answer] All the supply right now or the lion’s share of the supply is all internal, but we do expect, obviously, to win customers over time.

Intel’s AI-driven businesses are now 60% of revenue, and was up 40% year-on-year in 2026 Q1

AI-driven businesses now represent 60% of revenue and grew 40% year-over-year.

Intel’s management now expects capital expenditures to be flat in 2026, but the actual dollar-amounts spent on tools will be up 25% in 2026, as management is seeing a lot of demand and wants to catch up on supply

We forecast capital expenditures in 2026 to be flat to last year versus our prior expectation of flat to down, reflecting increased capacity investments to support committed demand and a continued emphasis on improving fab productivity and output. We now expect expenditures to be roughly equal across the year and still to be heavily weighted towards the equipment that directly grows wafer outs to support growth this year and next…

…In the last few years, a lot of our CapEx spending was space. And I think we’re actually in a pretty good position in space. We wanted to have white space available to move into when needed. And I think Lip-Bu and I both feel like we’re in a good place. So we actually will be bringing the space spend down pretty materially, even though the total is flat. And so what that means is the tool spend is actually increasing pretty significantly. In fact, tool spending will be up year-over-year 25% or so. And so that’s, I think, a function of the fact that we just see a lot of demand, and we want to make sure we’re catching up on the supply front.

Intel’s management thinks the ASIC business will be a fast-growing one for the company in the next 5 years; the ASIC business is already at a run rate of more than $1 billion

[Question] On the ASIC business, Dave, I think you said it doubled year-on-year. If you could maybe help us with what is included in that? I believe it’s IPUs, but I just want to get a better sense how big it is.

[Answer] Stay tuned on that one, the next 5 years is going to be a fast growing for us…

…One thing that people have been surprised about is how big the business is already. It’s at a run rate that’s north of $1 billion already.

Intuitive Surgical (NASDAQ: ISRG)

The da Vinci 5 captures real-world surgical data at greater scale and fidelity, enabling deeper surgical insights; the surgical insights captured by da Vinci 5 will be used by Intuitive Surgical for AI-enabled capabilities; management expects to add telesurgery and more automation to Intuitive Surgical’s robotic surgery platforms over the long term; management believes that AI will help Intuitive Surgical to move its Quintuple Aim forward; the data captured by da Vinci 5 includes video, kinematic, and force data; the AI-powered insights that management wants to deliver to customers can be in the form of operational guidance, learning of a surgeon/care team, and in the operating theatre

da Vinci 5 captures real-world surgical data at greater scale and fidelity, enabling deeper insight into how procedures are performed in practice. That insight paired with clinical context from connected electronic medical records, provides better understanding of variation, workflow and outcomes, and informs current and planned digital and AI-enabled capabilities…

…Collectively, these efforts are foundational to our long-term digital and AI road map where we expect to add telesurgery, deeper decision support and augmented dexterity, including aspects of future automation, all in pursuit of advancing the Quintuple Aim…

…We believe, yes, that AI will be a contributor to moving the Quintuple Aim forward…

…It starts with high-quality data, and that data will exist in video data from surgeries. It will exist in robotic data streams like kinematic data and force data. It will exist in connected electronic medical records, where we’re working with customers to do so. And once we have that high-quality data set, then the job of our AI and our data scientists is to turn that into meaningful insights…

…So there are, I think, ways in which this will show up to the customer. Some will be as operational guidance and assistance as they look at their hospital robotic program and want to increase efficiencies or understand costs. Some of it may show up in the learning of a surgeon and/or a care team. But a lot of it will show up in the operating room and I think show up in the surgery itself. And an example of this kind of first phase might be AI-enabled anatomy identification where you can see AI showing critical structures in the surgical field, showing tissue planes to help assist the surgeon. Then, over time, what we expect is that many of those same foundations that are being established and built in kind of that first phase, if you will, will support more advanced assistance around augmented dexterity and it will include — likely include aspects of automation. There, an example might be helping to control the camera as the surgeon is focused on the procedure.

Intuitive Surgical’s management thinks the company’s differentiation in AI comes from its installed base of da Vinci 5 systems, and the number of procedures performed by the systems annually which generates unique data

How do we sit, how do we exist within the AI ecosystem and how are we differentiated? I think part of that differentiation is around the installed base of systems that we have out there, including about the 1,500 da Vinci 5 systems, the 3 million and more procedures that are being done on an annual basis. And I believe that gives us the foundation to strengthen the differentiation over the next 3 to 5 years. If you look at the industry and you say, what is broadly available, broadly available to everyone, it’s things like edge and cloud compute, the math that underscores much of this, some of the training algorithms. Our advantage, we believe, lies is in the unique data sets that are available to us today through something like Force Feedback and will be increasingly available to us as we add capability to da Vinci 5.

Mastercard (NYSE: MA)

Mastercard is working with key players in the agentic commerce ecosystem, including Google, Microsoft, and OpenAI; Mastercard is partnering with OpenAI on Mastercard Agent Pay, which enables agent-to-agent payments; nearly all Mastercards globally are enabled for Mastercard Agent Pay; Mastercard’s management launched Verifiable Intent in 2026 Q1; Verifiable Intent is a temper-resistent record of authorisations a user has given to his/her agent; the FIDO Alliance is using Verifiable Intent as a foundation for security standards in agentic commerce; Crossmint, a leading blockchain infrastructure provider, will integrate Mastercard Agent Pay and Verifiable Intent so that it can enable secure Mastercard transactions for agents; Crossmint’s integrations will be launched initially on OpenClaw; management thinks Mastercard’s network will serve agentic commerce with tokenised credentials; management thinks agentic commerce will bring even more incremental opportunity in transactions and services over time; volumes with Mastercard Agent Pay are still low

On Agentic, the ecosystem continues to evolve. Our payment solutions are ready, and we are engaged, shaping what comes next with key players, including Google, Microsoft, OpenAI, and other partners across the ecosystem. We’re deepening our partnership with OpenAI, reinforcing their use of Mastercard Agent Pay, working to enable agent-to-agent payments and collaborating to embed our services across their solutions while using their tools as an enterprise customer. I’m also happy to share that nearly all Mastercards around the world are now enabled for Mastercard Agent Pay…

…In quarter 1, we launched Verifiable Intent, a tamper-resistant record of what a user authorized when an AI agent acts on their behalf. In fact, the FIDO Alliance is now using it as a foundation for setting security standards in this space. And earlier this month, we announced a partnership with Crossmint, a leading blockchain infrastructure platform. Crossmint will integrate Mastercard Agent Pay and Verifiable Intent to enable secure Mastercard transactions for AI agents in its ecosystem. This will initially launch on the OpenClaw platform with plans to expand…

…But as agent-driven commerce gains traction, our network is there with tokenized credentials, powering the payments, bringing the security, and trust, and reach that everyone is looking for. It’s very clear there is even more incremental opportunity in transactions and in services over time…

…[Question] In Mastercard Agent Pay. Michael, you talked about some of the partners and some of the activity on the ground, but can you just give us a little bit more detail on volumes or any surprises with respect to actual activity or actual demand?

[Answer] In terms of where volumes are, we’re still at early stage. So that is also true because a few things were not quite in place yet.

More than 500 customers are already engaged with Mastercard Threat Intelligence, which was launched in 2025 and powered by Recorded Future’s capabilities (Recorded Future was acquired by Mastercard in 2024 Q4 and it provides AI-powered solutions for real-time visibility into potential threats related to fraud); Mastercard Threat Intelligence have helped customers take down malicious domains responsible for the payment card test impacting over 10,000 e-commerce sites; Recorded Future puts Mastercard in a unique position to provide insights on threats faced by states 

Last year, we launched Mastercard Threat Intelligence, bringing Mastercard and Recorded Future capabilities together. In a short period of time, more than 500 customers are already engaged. Using the product, partners have taken down malicious domains responsible for the payment card test impacting over 10,000 e-commerce sites. That’s tangible value…

…Asymmetrical warfare, state actors, all of that is going on, and Recorded Future puts Mastercard in a very unique position to be a trusted partner to provide those kind of insights.

Mastercard has started to launch Mastercard Agent Suite, where Mastercard will design and deploy AI agents within customer environments; management thinks Agent Suite could be a much bigger opportunity than on the consumer side

You heard us talk about Agent Suite, which we started to launch, where we’re going to get into the business of building agents with our customers in the B2B space, et cetera. So early-stage on B2B earlier than on the consumer side, but I would think this is a much bigger opportunity, and it fits right into our focus on commercial payments. So early-stage ecosystem building, covering your basis, that’s what we’re doing.

Meta Platforms (NASDAQ: META)

Meta’s AI research lab, Meta Superintelligence Labs (MSL), has released the first model, MuseSpark, in its Muse family of models; MSL has built what management thinks is the strongest research team in the industry; MSL is already training even more advanced models than Muse; management thinks MuseSpark has already made Meta AI a world class assistant for users in many areas; management has heard very positive feedback on MuseSpark; management thinks Meta’s product team is now able to build products on top of the company’s models because the models are now strong, unlike in the past; management thinks models in the future will have to be able to improve themselves in order for them to be considered leading models; management is not focused on building coding capabilities with Meta’s AI models; coding is not the only ingredient needed for models to be self-improving

Our biggest milestone so far this year has been the release of our Muse family of models and our first model MuSpark along with a significantly upgraded new version of Meta AI. This was the first release from Meta Super Intelligence Labs, and it shows that our work is on track to build a leading lab. Over the past 10 months, we have built the strongest research team in the industry and established the scientific and technical foundations to scale very advanced models. Spark is just one step on that scaling ladder, and we are already training even more advanced models…

…Spark has already made Meta AI, a world-class assistant that leads in several areas related to our vision of personal super intelligence, including visual understanding, health, shopping, social content, local, creating games and more. We’re hearing very positive feedback on it so far…

…We have our product team, and that team is now really unlocked to be able to build things on top of our models because we now have a very strong model. So before this, we have been prototyping a bunch of things using other different models, whether it was our previous older models or kind of using the APIs from other companies. And now we’re unlocked to be able to go build things and get them to scale on top of our own models…

…You’re not going to have leading models in the future if your models can’t improve themselves, right? So you’re getting to a point where today, the models are still able to learn from people — and then I think at some point, the models will have to improve themselves. And that’s how the growth is going to — an improvement in the models is going to happen…

…Does that make us a developer tools company? Not necessarily. I mean, I’m not against having an API or coding tools or anything like that. But it’s not our primary focus. But I actually think people conflate coding with self-improvement more than they should. Coding is one ingredient for the model self improving. It’s not the only thing. And we are focused on all of the parts that are going to be necessary for self-improvement in service of the personal super intelligence vision that we have for people and businesses.

Meta AI has seen large increases in usage since MuseSpark was introduced, with double-digit percent increases in Meta AI sessions per user; the Meta AI app has consistently been near the top in app stores; MuseSpark is now powering Meta AI in chat threads in Facebook, Instagram, WhatsApp, and Messenger, as well as in the standalone Meta AI app

We’ve seen large increases in Meta AI use since releasing the updates, and the Meta AI app has consistently been near the top of the app stores as well…

…We’re seeing encouraging results within Meta AI since we began powering responses with the first model from MSL, Muhspark. In tests we ran leading up to the launch, we saw meaningful engagement gains that accelerated week-over-week with each new iteration of the model. We’re seeing similar games within Meta AI following the broad rollout of our new model with double-digit percent increases in Meta AI sessions per user. MuseSpark is now powering Meta AI in direct chat threads across our family of apps as well as the stand-alone Meta AI app and website, giving billions of people globally access to our latest model.

Meta’s management has a very view on AI than others in the industry; management thinks that AI will help people and improve many aspects of their lives; management wants to build AI agents that empower people and businesses; management thinks there are clear monetisation opportunities for personal superintelligence

My view of AI is very different from many others in the industry. I hear a lot of people out there talk about how AI is going to replace people. Instead, I think that AI is going to amplify people’s ability to do what you want, whether that’s to improve your health, your learning, your relationships, your ability to achieve your personal career goals and more. My view is that human progress has always been driven by people pursuing their individual aspirations. And I believe that this will continue to be true in the future. People will be more important in the future, not less. Meta believes in empowering individuals. And those are the kinds of products that we’re going to build, and I believe that they’re going to be some of the most important and valuable products of all time. We are building a personal agent focused on helping people achieve the diverse goals in their lives. We’re also building a business agent focused on helping entrepreneurs and businesses across the world, use our tools and others to grow their efforts, reach new customers and serve existing customers better. These agents will work together to form an ecosystem…

…The focus is on building personal super intelligence, building a consumer agent that can work for you and help you get things done. That right now is a consumer experience that we’re focused on, but we think there will be clear monetization opportunities over time. You can imagine commission structures or a premium offering.

Meta’s management has been testing business AIs and weekly conversations have 10x-ed (from 1 million to 10 million) since the start of 2026; the Meta AI business assistant was recently fully rolled out to all eligible advertisers on supported Meta buying services and performance has been strong, with common account issues being resolved at a 20% higher rate; the business AIs are tested in SMBs across Latin America and Asia Pacific; management will expand access to the business AIs in 2026 Q2; the business AIs are currently free, but management expects to monetise them over time

We’re already testing an early version of business AIs and weekly conversations have grown 10x since the start of this year…

…The Meta AI business assistant has now been fully rolled out to all eligible advertisers on supported Meta buying services, providing personalized recommendations to advertisers, resolving account issues, and servicing campaign insights to help optimize results. Performance has been strong since we began testing the assistant in Q4 with common account issues being resolved at a 20% higher rate…

…In Q1, we expanded business AIs on WhatsApp to SMBs across Latin America and Indonesia as well as on Messenger in Asia Pacific. We now have more than 10 million conversations each week being facilitated through business AIs, up from 1 million at the start of the year. We’ll further expand access to more countries this quarter while adding more capabilities to the AIs…

…Business AIs today are currently free for most businesses on our messaging apps. But as we make more progress, we expect that we will also work towards establishing a longer-term monetization model.

Meta’s management is working to incorporate MuseSpark in the company’s upcoming models used in its recommendation systems, core apps, and advertising products; the upcoming models will enable Meta to understand more of people’s goals for the first time in the company’s history; in the last few years, Meta has seen an increasing return on the amount that it can improve user-engagement, and this has encouraged management to continue investing heavily in this area 

We’re also working on using Spark in our upcoming models to improve our recommendation systems and core business in Facebook, Instagram and ads. Right now, our apps primarily help people accomplish 3 important goals: connecting with people, learning about the world and entertainment. But we’ve always wanted our apps to understand more of people’s goals so we can help improve their lives in all the ways that they want. These new AI models will let us understand this in more detail. So instead of just looking at statistical patterns of what types of people engage with what content, for the first time in Meta’s history, we’re going to be able to develop a first principles understanding of what you care about and what each piece of content in our system is about — is that way we can show you more useful things for what you’re trying to accomplish. And we’ll also be able to create personalized content specifically for people to help you achieve your goals as well. Since our recommendation systems are operating at such a large scale, we’ll phase in this new research and technology over time.

But the trend over the last few years seems clear that we are seeing an increasing return on the amount that we can improve engagement for people and value for advertisers. This encourages us to continue investing heavily in what we expect will provide increasing value over the coming years as well.

Meta will be rolling out more than 1 GW (gigawatt) of its own custom chips; Meta’s AI compute infrastructure will include large amount of its own chips and AMD chips, alongside NVIDIA chips; Meta is investing in more compute, partly through multiyear cloud deals; Meta’s contract commitments increased by $107 billion in 2026 Q1; the multiyear cloud deals support both Meta’s training and inference needs; management has consistently underestimated Meta’s compute needs even as the company has been ramping up compute capacity significantly; management expects compute to be even more central for the business going forward

We are rolling out more than 1 gigawatt of our own custom silicon that we’re developing with Broadcom, as well as significant amount of AMD chips to complement the new NVIDIA systems that we’re rolling out as well…

…We’re also signing cloud deals that will come online over the course of this year and 2027, allowing us to scale more quickly. These multiyear cloud deals and our infrastructure purchase agreements drove a $107 billion step-up in our contractual commitments this quarter. Our investments will support our training needs for future models and most importantly, provide us the inference capacity necessary to deliver personal and business agents to billions of people around the world, along with several other AI product experiences we’re developing…

…Our experience so far has been that we have continued to underestimate our compute needs even as we have been ramping capacity significantly as the advances in AI have continued and our teams continue to identify compelling new projects and initiatives. And now to, there are very compelling internal use cases. So our expectation is that compute will become even more central to the business going forward.

Meta’s AI glasses continue to perform well, with daily users tripling year-on-year in 2026 Q1; the AI glasses continue to be one of the fastest-growing categories of consumer electronics ever; Meta released new glasses for all-day wear in 2026 Q1; Met has new partnerships and styles for AI glasses coming later this year; all of Meta’s AI glasses are designed to easily update to Meta’s newest AI models and features; Meta’s AI glasses are evolving into a personal agent product; the sales of Meta’s AI glasses have shifted from the prior generation to the latest generation; management is seeing strong interest in the Meta Ray-Ban Display product that comes with neural bands; management thinks the Meta Ray-Ban Display product will be the next generation for how AI glasses evolve

Our AI glasses continue to perform well with the number of people using them, daily tripling year-over-year. This continues to be one of the fastest-growing categories of consumer electronics ever. We released Ray-Ban Meta optics this quarter designed for all day wear rather than primarily as sunglasses. And building on our release of Oakley last year, we have some exciting new partnerships and styles that I think are going to have the potential to reach even more people coming later this year. All of our glasses are designed to easily update to use our newest AI models and features. I’m also really excited to see the glasses evolve from being able to answer questions to being able to be a personal agent that’s with you all day long, helping you remember things and achieve your goals…

…We’re seeing sales shift now from the prior generation of Ray-Ban Meta’s to the latest generation, which I think speaks to the value of the improved features like extended battery life and higher features like higher resolution video capture…

…We see strong interest now in the Meta Ray-Ban displays with the Meta neural bands. So that’s an encouraging sign that there is consumer appetite for display glasses, which is kind of the next generation of how this product evolves.

Ranking improvements made in 2026 Q1 drove a 10% increase in time spent on Instagram Reels, an 8% increase in total video time on Facebook globally, and a 9% increase in video watch time on Facebook in the US and Canada; the ranking improvements are driven by a number of things, including (1) the doubling in the length of user interaction sequences for training on Instagram, (2) increasing the speed of indexing new posts by the ranking models, and (3) applying more advanced content understanding techniques; same-day posts are now more than 30% of recommended posts in Instagram and Facebook, up more than 2x from a year ago; management is now using AI to auto translate and dub videos into a viewer’s local language; more than 500 million users are watching translated videos weekly on each of Facebook and Instagram; management continues to invest in Meta’s recommendation capabilities, and the investments include near term ones such as scaling up models in size and complexity and incorporating LLMs, or large language models, to deepen content understanding, and long-term ones such as building foundation models for organic content and ads recommendations, and LLM-based recommendation systems; management thinks there is still a lot of room to continue improving recommendations on both Facebook and Instagram

We’re continuing to see significant gains from our content recommendation initiatives. On Instagram, the ranking improvements that we made in Q1 drove a 10% lift in Reels time spent. On Facebook, total video time increased more than 8% globally in Q1, the largest quarter-over-quarter gain in 4 years. Within the U.S. and Canada, ranking improvements we made drove a 9% increase in video watch time on Facebook in Q1. 

These gains are benefiting from advances we’re making across the full stack. Starting with data, we doubled the length of user interaction sequences we use for training on Instagram in Q1 and increase the richness of how each user interaction is described, enabling our systems to develop a deeper understanding of user interests. Within our models, we’ve significantly increased the speed with which our ranking models index new posts, which is enabling us to recommend them sooner after they are published. We’re also applying more advanced content understanding techniques, which is enabling us to quickly identify posts that may be interesting to someone even if they haven’t engaged with a lot of similar content. These and other improvements have enabled us to increase the diversity and recency of recommended content with same-day posts now representing more than 30% of recommended reels on both Instagram and Facebook more than double the levels 1 year ago.

We’re also using AI to unlock more inventory by auto translating and dubbing videos into a viewer’s local language, enabling us to recommend a more diverse set of content. Over 0.5 billion users on each of Facebook and Instagram are now watching AI translated videos weekly. 

Looking forward, we’re making several investments we expect will deliver more valuable recommendations. This year, we will continue scaling up our models in several dimensions, including their size and complexity, while incorporating LLM to deepen content understanding across our platform. This will enable us to better match people to a wider variety of content aligned to their interests. At the same time, we are executing on our longer-term efforts to develop the next generation of our recommendation systems. This includes building foundation models that power organic content and ads recommendations as well as developing LLM based recommender systems. Our focus this year is validating the model architectures and techniques in these domains before we scale them out in future years…

…There is still a lot of room to continue improving recommendations over the rest of the year, and we expect we’ll be able to do that to drive additional engagement on both Facebook and Instagram.

Meta continues to enhance its systems to show advertising to users at the optimal time and location; improvements made to Lattice and GEM (Generative Ads Model) in 2026 Q1 increased conversion rates for landing page view advertising by more than 6%; management expanded coverage of Meta’s new adaptive ranking model, which was rolled out in 2025 H2, to off-site conversions and this drove a 1.6% increase in conversion rates across Facebook and Instagram’s major surfaces; Meta is introducing Meta Ads AI Connectors in open beta and it allows advertisers to connect their Meta advertising accounts directly to an AI agent; more than 8 million advertisers are now using at least one of Meta’s Gen AI advertising creative tools with very strong adoption among SMB advertisers; advertisers using Meta’s video generation feature are seeing 3% higher conversion rates in tests; Meta’s value optimisation suite, which maximises the return on advertising spend for advertisers by prioritising the highest value conversions, has seen strong adoption with the revenue run rate reaching $20 billion in 2026 Q1, more than double from a year ago; Meta’s new adaptive ranking model enables the company to leverage LLM-scale model complexity when it previously couldn’t

We continue to enhance our systems to show ads at the optimal time and location…

…In Q1, enhancements we made to Lattice’s modeling and learning techniques, along with advances in our GEM model architecture, drove a more than 6% increase in conversion rate for landing page view ads. In addition, we’ve been investing in more performing inference models for 1 more serving ads. In the second half of last year, we began rolling out our new adaptive ranking model, which is an LLM scale adds recommender model that we use for inference. This model improves our inference ROI by routing requests to more compute-intensive inference models when it determines there is a higher probability of conversion. In Q1, we expanded coverage of our adaptive ranking model to support off-site conversions, which drove a 1.6% increase in conversion rates across the major surfaces on Facebook and Instagram…

…This week, we’re also introducing Meta ads AI connectors in open beta, providing advertisers the ability to connect their Meta ad account directly to an AI agent. We’ve always supported advertisers both on our platform and through tools like the marketing API. And now we’re extending that to AI. So businesses and agencies can analyze and optimize campaigns with the tools they’re already using.

Usage of our ad creative tools is also scaling with more than 8 million advertisers using at least one of our Gen AI ad creative tools and particularly strong adoption among small- and medium-sized advertisers. These tools are benefiting performance as well with advertisers using our video generation feature seeing more than 3% higher conversion rates in tests…

…We also continue to invest in the value optimization suite, which helps advertisers maximize their return on ad spend by prioritizing the highest value conversions rather than optimizing solely for the most conversions at the lowest cost. Adoption by businesses has been strong following performance improvements we’ve made over the past year with the annual revenue run rate of our value optimization suite now over $20 billion, more than doubling year-over-year…

…The inference models are bound by strict latency requirements since they need to find the right ad within milliseconds, and that has, again, historically prevented us from meaningfully sizing up — scaling up their size and complexity. But in the second half of last year, we introduced a new adaptive ranking model, which enables us to leverage LLM scale model complexity of 1 trillion parameters, and we made advances in the model architecture and codesign the system with the underlying silicon, so it maintains the sub-second speed that is required to serve ads at scale. We also developed an approach that intelligently routes request more compute-intensive inference models if it determines that there is a higher probability of conversion and that lets us drive both better performance and increased inference ROI.

Microsoft (NASDAQ: MSFT)

Microsoft’s management has 2 priorities to capture the AI opportunity, namely, (1) build the leading cloud and AI infrastructure, and 2) build high-value agentic systems across core domains

We are at the beginning of one of the most consequential platform shifts that will change the entire tech stack as agents proliferate and become the dominant workload. This will drive TAM expansion and change the value creation equation across the entire economy. To capture this opportunity, we are executing against 2 priorities. First, we are building the world’s leading cloud and AI infrastructure for agentic computing era. Second, we are building high-value agentic systems across core domains such as productivity, coding and security

Microsoft’s management is optimising every layer of its technology stack and this is producing operational gains; Microsoft’s dock-to-live times for its data centers has reduced by 20% since the start of 2026; Microsoft has delivered a 40% improvement in inference throughput in Copilot’s most-used models

We’re optimizing every layer of the tech stack, from DC design, to silicon to system software, the model architecture as well as its optimization. This is translating into operational gains. We have reduced dock-to-live times for new GPUs in our biggest regions by nearly 20% since the beginning of the year. Our Fairwater data center in Wisconsin came online earlier this month, 6 weeks ahead of schedule, allowing us to recognize revenue earlier. And we delivered a 40% improvement in inference throughput for our most used models across Copilot, driven by our software and hardware optimization work.

Microsoft added 1 gigawatt of GPU compute capacity in 2026 Q1 (FY2026 Q3); Microsoft is on track to double its overall compute footprint in 2 years; management announced new data center investments across 4 continents in 2026 Q1 (FY2026 Q3)  

All up, we added another gigawatt of capacity this quarter and remain on track to double our overall footprint in just 2 years. We are moving aggressively to add capacity aligned to our demand signals we see and we have announced new data center investments across 4 continents.

Microsoft’s AI infrastructure utilises chips from NVIDIA, AMD, and itself (Maia); Microsoft’s Maia 200 chip has 30% better tokens per dollar compared to other leading AI chips, and is now live in 2 Microsoft data centers; Microsoft’s Cobalt server CPUs are deployed in half of the company’s data center regions; as Microsoft’s customers scale their AI workloads, they are increasingly using other Microsoft cloud services and are choosing Cobalt to run these services; management is expanding Cobalt’s supply significantly to meet demand

We also continue to modernize our fleet with our first-party innovation alongside the latest from NVIDIA and AMD. Across our fleet, millions of servers are powered by our custom networking security and virtualization silicon, including Azure Boost as well as our first-party CPUs and accelerators. Our Maia 200 AI accelerator, which offers over 30% improved tokens per dollar compared to the latest silicon in our fleet, is now live in our Iowa and Arizona data centers. Our Cobalt server CPU is deployed in nearly half of our DC regions running workloads at scale for customers like Databricks, Siemens and Snowflake. As our largest customers scale their AI deployments, they’re increasingly leveraging other services across our platform and choosing to run those workloads on Cobalt. And we are expanding Cobalt supply significantly to meet this demand.

Microsoft’s management thinks Microsoft offers the broadest selection of models among the cloud hyperscalers; over 10,000 customers have used more than 1 model on Foundry; the number of customers who used Anthropic and OpenAI models doubled sequentially in 2026 Q1, or FY2026 Q3 (was 1,500 in 2025 Q4, or FY2026 Q2); Bayer is using multiple models in Foundry to build its in-house agent platform; over 300 Microsoft customers are on track to process 1 trillion tokens each on Foundry in 2026, up 30% sequentially 

We offer the broadest selection of models of any hyperscaler, so customers can choose the right model for the right workload across OpenAI, Anthropic, open source and more. Over 10,000 customers have used more than one model on Foundry. 5,000 have used open source models, and the number who have used Anthropic and OpenAI models increased 2x quarter-over-quarter…

…Bayer is using multiple models in Foundry to create its own in-house agent platform with more than 20,000 active monthly users. All up, over 300 customers are on track to process over 1 trillion tokens on Foundry this year, accelerating 30% quarter-over-quarter.

Microsoft’s management is building a unified IQ layer for organisational intelligence; the IQ layer initiative is driving acceleration in Microsoft’s data businesses, with Cosmos DB revenue up 50% year-on-year in 2026 Q1 (FY2026 Q3), Fabric customers growing 60% year-on-year to 35,000, and Fabric OneLake data up 4x year-on-year; 15,000 customers now use both Fabric and Foundry, up 60% year-on-year; Fabric provides agents with operational, analytical, and unstructured data; Microsoft’s Copilot Studio is helping enterprises build agents; nearly 90% of the Fortune 500 have active agents built with Copilot Studio’s low-code and no-code tools; Copilot’s credit consumptive offer is up 2x sequentially in 2026 Q1 (FY2026 Q3); Agent 365 is a control plane for managing agents’ governance, identity, and security; tens of thousands of companies are already using Agent 365 to manage tens of millions of agents

Across Fabric, Foundry, Microsoft 365 and our Security Graph, we are building a unified IQ layer for organizational intelligence. Thousands of enterprises already are accessing context across these IQ layers. And as AI usage grows, so does the context layer, creating a flywheel that continuously improves the grounding, relevance and effectiveness of every agent they use and build, making our IQ layers an unmatched context engine for organizational intelligence. More broadly, our database business accelerated quarter-over-quarter. Cosmos DB alone saw 50% year-over-year revenue growth driven by AI app workloads. We now have 35,000 paid Fabric customers, up 60% year-over-year. And all up, the amount of data in Fabric OneLake data lake increased nearly 4x year-over-year. Over 15,000 customers now use both Foundry and Fabric, up 60% year-over-year as enterprises connect agents to real-time operational, analytical and unstructured data that Fabric brings together…

…We are also helping knowledge workers build agents with tools like Copilot Studio. Nearly 90% of the Fortune 500 now have active agents built with our low-code/no-code tools. And we are seeing fast growth of our Copilot credit consumptive offer, up nearly 2x quarter-over-quarter as customers increasingly extend Copilot with custom agents tailored to their workflows…

…With Agent 365, we offer a control plane that extends company’s existing governance, identity, security and management frameworks to agents. Tens of thousands of companies are already managing tens of millions of agents in Agent 365, and we expect this momentum to grow significantly as agents will increasingly need tools for identity, governance, security and more.

Microsoft’s management is turning its family of Copilots from synchronous assistance software to asynchronous digital workers; Microsoft 365 Copilot seat adds grew 250% year-on-year in 2026 Q1 (FY2026 Q3), the fastest growth since launch; there are now over 20 million Microsoft 365 Copilot paid seats; the number of companies with over 50,000 Microsoft 365 Copilot seats grew 4x year-on-year in 2026 Q1 (FY2026 Q3); WorkIQ grounds Copilot’s responses with an organisation’s full context; the data residing in WorkIQ now spans 17 exabytes, up 35% year-on-year; users can now access multiple models together in Microsoft 365 Copilot to generate the best responses; monthly active usage of Microsoft’s 1st-party agents in Microsoft 365 Copilot is up 6x year-to-date; Copilot queries per user was up 20% sequentially in 2026 Q1 (FY2026 Q3); weekly engagement of Microsoft 365 Copilot is now on par with Outlook

We are evolving our family of Copilots from synchronous assistance to async coworkers that can execute long-running tasks across key domains. In knowledge work, it was another record quarter for Microsoft 365 Copilot seat adds, which increased 250% year-over-year, representing our fastest growth since launch. Quarter-over-quarter, we continue to see acceleration and now have over 20 million Microsoft 365 Copilot paid seats. The number of customers with over 50,000 seats quadrupled year-over-year and Accenture now has over 740,000 seats, our largest Copilot win to date. And Bayer, Johnson & Johnson, Mercedes and Roche all committed to 90,000 or more seats…

…Work IQ grounds Copilot responses in the full context of an organization, including people, roles, documents and communications, all within the company’s security boundary. The system of work behind Work IQ alone now spans more than 17 exabytes of data growing 35% year-over-year. The liquidity and freshness of that data matters, with billions of e-mails, documents, chats, hundreds of millions of Teams meetings, and millions of SharePoint sites added each day. And that context is getting even richer as Copilot adoption grows, Copilot and Agent conversations and artifacts they create feedback into Work IQ, making it even more context-rich…

…In Microsoft 365 Copilot, you now have access in chat to multiple models by default with intelligent auto routing, in Agents with Critique and Council. You can use multiple models together to generate optimal responses. As of last week, Agent Mode is now default experience across Copilot in Word, Excel and PowerPoint. And with Cowork, you now have a new way to delegate and complete work using Copilot.

All this innovation is driving record usage intensity across Copilot. We have seen a surge in usage of our first-party agents with monthly active usage up 6x year-to-date. Copilot queries per user were up nearly 20% quarter-over-quarter. To put this momentum in perspective, weekly engagement is now at the same level as Outlook, as more and more users make Copilot a habit.

Microsoft’s management is observing a shift in pricing in business software from seat-based models to seat-plus-consumption models because of AI; nearly 60% of Microsoft’s service customers are already buying usage-based credits; HSBC is using pre-built agents to reduce issue resolution time for customer inquiries by 30%; LinkedIn Talent Solutions’ agentic products now have an annualised revenue run rate of more than $450 million; management thinks the pricing model for business software could yet evolve further to include business outcomes into the equation

When it comes to biz apps, we are seeing a new pattern emerge as customers shift from traditional seat model to seats plus consumption. The customer service category is at the forefront of this transformation as nearly 60% of our service customers are already purchasing usage-based credits. For example, HSBC uses prebuilt agents with Dynamics 365 to manage customer inquiries across products, markets, regulatory requirements, reducing issue resolution time by over 30%. And our agentic products in LinkedIn Talent Solutions, which help hirers automate time-consuming tasks like sourcing, screening and drafting messages have already surpassed a $450 million annualized revenue run rate…

…From a customer perspective, they’re going to evaluate it by evals. Where are they seeing the value of tokens, as simple as that. So where they see the outcome, the eval and the token, whether it’s improving revenue, improving efficiency, and that’s what will refine. Like when we talk about IT budgets, IT budgets are going to have to be reshaped by a combination of business outcomes, making their way into IT budgets and maybe reallocation from other line items on the income statement like OpEx.

GitHub is growing rapidly, driven by agentic coding; nearly 140,000 organisations are using GitHub Copilot; GitHub Copilot enterprise subscribers nearly tripled year-on-year in 2026 Q1 (FY2026 Q3); most users in GitHub Copilot use multiple models; usage of GitHub Copilot CLI (command line interface) nearly doubled month-on-month; management has shifted GitHub Copilot to a usage-based pricing model

GitHub itself is seeing unprecedented growth driven by proliferation of agentic coding, and we are hard at work to scale and meet this demand. We see this even with GitHub Copilot. Nearly 140,000 organizations now use GitHub Copilot and enterprise subscribers have nearly tripled year-over-year. The majority of users leverage multiple models. We’re also seeing rapid adoption of GitHub Copilot CLI with usage nearly doubling month-over-month. And earlier this week, we announced our move to usage-based pricing model for GitHub Copilot as we align pricing to actual usage and cost.

1/3 of Microsoft’s cloud and AI-related capex in 2026 Q1 (FY2026 Q3) are for long-lived assets that will support monetisation over the next 15 years and more, while the other 2/3 are for CPUs and GPUs; Azure is still capacity-constrained, and management wants to balance Azure demand for compute with 1st party demand for compute; Azure’s capacity-constrain is expected to last through at least 2026

Capital expenditures were $31.9 billion, down sequentially due to the normal variability from cloud infrastructure buildouts and the timing of delivery of finance leases. And this quarter, roughly 2/3 of our CapEx was for short-lived assets, primarily GPUs and CPUs. The remaining spend was for long-lived assets that will support monetization over the next 15 years and beyond. This quarter, total finance leases were $4.7 billion and were primarily for large data center sites. And cash paid for PP&E was $30.9 billion, roughly in line with capital expenditures as the impact from finance leases was partially offset by differences between the receipt of goods and payment…

…In Azure and other Cloud Services, revenue grew 40% and 39% in constant currency against a prior year that included accelerating growth. Results were ahead of expectations as we delivered capacity earlier in the quarter, enabling increased consumption across both AI and non-AI services. Strong customer demand across workloads, customer segments and geographic regions continues to exceed available capacity…

…Broad and growing customer demand continues to exceed supply, and we continue to balance the incoming supply we can allocate here against our other high ROI priorities, first-party applications, R&D and end-of-life server replacement…

…Even with these additional investments and continued efforts to bring GPU, CPU and storage capacity online faster, we expect to remain constrained at least through 2026.

Azure grew revenue by 40% in 2026 Q1 (FY2026 Q3) (was 39% in 2025 Q4); Azure’s revenue growth was better than expected because capacity was delivered earlier in the quarter; Azure continues to be constrained by capacity and the constraint is expected to last through at least 2026; management wants to balance Azure demand for compute with 1st party demand for compute; as Microsoft’s customers scale their AI workloads, they are increasingly using other Microsoft cloud services; Azure’s margin for its AI business remains better than the non-AI business when it was at a similar age

In Azure and other Cloud Services, revenue grew 40% and 39% in constant currency against a prior year that included accelerating growth. Results were ahead of expectations as we delivered capacity earlier in the quarter, enabling increased consumption across both AI and non-AI services. Strong customer demand across workloads, customer segments and geographic regions continues to exceed available capacity…

…Broad and growing customer demand continues to exceed supply, and we continue to balance the incoming supply we can allocate here against our other high ROI priorities, first-party applications, R&D and end-of-life server replacement. As a reminder, year-over-year Azure growth rates can vary quarter-to-quarter based on capacity, timing and contract mix…

…Even with these additional investments and continued efforts to bring GPU, CPU and storage capacity online faster, we expect to remain constrained at least through 2026…

…. As our largest customers scale their AI deployments, they’re increasingly leveraging other services across our platform and choosing to run those workloads on Cobalt…

…We’ve been talking about sort of where this AI business of ours has been in the cycle compared to even the cycle we saw with the cloud, which now seems very long ago. And how margins were actually better and they remained better in our AI business versus where we saw in the cloud transition, looking back.

Microsoft’s management has gained more confidence over the past 1-2 years that the economics of AI’s addressable market are in areas where the company has structurally strong positions in

One of the things that we have learned even in the last, whatever, 2 years or so in AI and also build more conviction and confidence on is where is the TAM and the category economics of the TAM. And so this, I mean, it’s fascinating that here we are in 2026 and the most exciting things are plug-ins in Word or Excel or CLIs in coding or — and so when you see that, that means we have a structural position in knowledge work, coding, security, which are the big TAMs.

Microsoft’s management continues to feel good about partnering with OpenAI after the recent change to the 2 companies’ agreement; Microsoft has full IP rights to OpenAI’s frontier models all the way to 2032; OpenAI remains a large customer of Microsoft

We feel good about our partnership with OpenAI. I’m always very, very focused on any partnership and ensuring that there’s a win-win construct at all times. I mean that’s how you can remain with partners. In this case, it starts with, quite frankly, IP, Amy referenced this. We have a frontier model, royalty-free with all the IP rights that we will have access to all the way to ’32, and we fully plan to exploit it…

…They’re a large customer of ours, not just on the AI accelerator side, but also on all the other compute side, and so we want to serve them well.

Netflix (NASDAQ: NFLX)

Netflix has been using generative AI to improve content recommendations for members; management is also leveraging generative AI to provide better tools for filmmakers; Netflix acquired InterPositive, a company providing AI-powered filmmaking tools, in March 2026; management thinks Netflix has significant and unique data for applying AI; management thinks even with AI tools, only great artists can make great art; Netflix’s content creation partners have been leveraging AI tools for many purposes, and these tools also help improve on-set safety; InterPositive contains proprietary technology created specifically for filmmakers and for filmmaking, so it’s different compared to other generative AI video apps; management is already seeing momentum around adopting InterPositive’s tools among Netflix’s content creation partners; management has been working on content recommendation and personalisation for many years, but they think generative AI provides plenty of opportunity for Netflix to continue improving in those areas; management thinks AI can be applied in Netflix’s advertising suite to make it easier to create new formats, customise ads, and improve contextual relevance 

We’ve been using machine learning and AI for many years, and as the technology advances with GenAI, we continue to find new opportunities to deliver an even more seamless experience for members and expand possibilities for storytellers. This includes using GenAI to improve recommendations for members through deeper content understanding so we can recommend the right title at the right moment, test conversational discovery experiences, and improve the breadth and quality of our promotional assets. Leveraging GenAI, we are enabling our creative partners with more and better tools to help them tell their stories, with the potential to make our single largest area of spend—content—even more impactful. To accelerate this opportunity, in March we announced our acquisition of InterPositive, the filmmaking technology company founded by Ben Affleck that develops AI‑powered tools built by and for filmmakers…

…Given our technology DNA, we have a significant and unique data assets here. We have tremendous scale. So we see that as all great opportunities to leverage new technical capabilities across every aspect of the business. So I think AI is going to deliver benefits for our members, for creators and for our employees…

…It takes a great artist to make great art and AI won’t change that. But AI will give those artists better tools to bring those visions to life in ways that we’re just scratching the surface on. So today, our talent leverages these tools for things like set references, pre visualization, visual effects, sequence prep, shot planning. All of these things, by the way, also improve on-set safety, which is something that’s not talked about enough…

…With our acquisition of InterPositive, we think it accelerates our GenAI capabilities because it’s a proprietary technology that was created specifically for filmmakers and specifically for filmmaking and that’s different than other GenAI video applications. So while our ownership of InterPositive is very new, we have generated a bunch of interest with our creators who spent time with the tools, and we’re seeing real momentum build around adoption…

…We’ve been in personalization and recommendation for 2 decades, but we still see tremendous room and opportunity to make it even better by leveraging some of these newer technologies. We see that recommendation systems based on these new model architectures, not only improve the current personalization, but it also allows us to iterate and improve more quickly to improve that velocity. Things like adding support for different content types going forward, that’s much more quick, much more efficient…

…We really see an opportunity to leverage AI within our Netflix ad suite. Makes it easier to design new creative formats, custom ads, improved — that improve contextual relevance. And the technology stack just allows us to roll them out more quickly, more effectively and allow partners to leverage those things in an easier manner.

Taiwan Semiconductor Manufacturing Company (NYSE: TSM)

TSMC’s capital expenditure is always in anticipation of growth in future years; management expects capex for 2026 to be near the high end of its previous guidance of US$52 billion to US$56 billion (growth at the high would be 37% from 2025’s capex of US$41 billion); management now expects TSMC to grow revenue by more than 30% in USD terms in 2026 (previous guidance was for growth to be nearly 30%); TSMC’s capex in the last 3 years was ~US$100 billion, and the next 3 years is expected to be much higher, although management does not expect a sudden surge in capital intensity; management thinks the AI accelerators business will have a CAGR for 2024-2029 towards the high end of the previously released growth forecast of mid-to-high-50% CAGR

At TSMC, a higher level of capital expenditures is always correlated with higher growth opportunities in the following years…

…We now expect our 2026 capital budget to be towards the high end of our range of between USD 52 billion and USD 56 billion, as we continue to invest heavily to support our customers’ growth…

…We maintain strong confidence for our full year 2026 revenue to now grow by above 30% in U.S. dollar terms…

…In the past 3 years, our total CapEx was $101 billion. This year, we’re already seeing is towards the high end, which is $56 billion, which is already over 50% of the past 3 years in total. So we have a strong conviction in the AI megatrend. So we expect the CapEx in the next few years, in the next 3 years will be significantly higher than the past few years…

…Now therefore, we do not expect in the next several years, a sudden surge in capital intensity…

…But again, let me say that is toward higher 50s of the CAGR that we observe.

TSMC has been sourcing helium (an element whose supply has been affected by the conflict in the Middle East) from different regions, and it has safety stock in hand; TSMC has been working with Taiwan’s government to secure power, and Taiwan has sufficient LNG supply through at least May; management does not expect any near-term impact to TSMC’s operations from the Middle East conflict in terms of materials and power supply

About the materials and energy supply update given the recent situation in the Middle East. TSMC operates a well-established enterprise risk management system to identify and assess all relevant risks and proactively implement risk mitigation strategies. In terms of material supply, TSMC’s strategy is to continuously develop multi-store supply solutions to build a well-diversified global supplier base and to improve the local supply chain. For specialty chemicals and gases, including helium and hydrogen, we source from multiple suppliers in different regions and we have prepared safety stock inventory on hand. We are also working closely with our suppliers to further strengthen the resiliency and sustainability of our supply chain. Thus, we do not expect any near-term impact on our operations for material supply.

In terms of energy, TSMC worked closely with Thai Power and the Taiwan government to ensure a stable and sufficient energy supply. With the recent situation in the Middle East, the Taiwan government has announced it has secured sufficient LNG supply through at least May. The government has also said it is actively working on securing further LNG supply, diversifying sourcing to other regions and other power backup plans. Therefore, we do not expect any near-term disruption or impact to our operations.

TSMC’s management sees very robust AI-related demand, as the shift from generative AI and queries (chatbots) to agentic AI is leading to a step-up in token consumption; management is seeing very strong signals and positive outlooks from TSMC’s customers’ customers, who are the cloud service providers; management’s conviction in the AI megatrend remains high

AI-related demand continues to be extremely robust. The shift from generative AI and the query mode to agentic AI and command and action mode is leading to another step-up in the amount of token being consumed. This is driving the need for more and more computation, which supports the robust demand for leading edge silicon. Our customers and customers of customers, who are mainly the cloud service providers, continue to provide us with a very strong signal and positive outlook. Thus, our conviction in the multiyear AI megatrend remains high, and we believe the demand for semiconductors will continue to be very fundamental.

TSMC’s management intends to ramp up new technology nodes in Taiwan because of the need for tight integration between production and R&D; TSMC’s N2 node entered high-volume manufacturing in 2025 Q4 in Taiwan with good yield; N2’s ramp is supported by strong demand from both smartphone and HPC AI applications; management believes that N2, N2P, and A16 will lead to the N2 family becoming another large and long-lasting node for TSMC; management has decided to add capacity for N3 even though TSMC has historically not added capacity to a node once it has reached its target capacity, because of the strong demand for N3 in AI applications; management is seeing robust multiyear demand for N3 nodes from end markets such as smartphone, HPC AI, and more; TSMC is adding a new N3 fab to its giga fab cluster in Tainan, with volume production expected in 2027 H1; TSMC is continuing to convert N5 tools to support N3 capacity in Taiwan; management is focusing on flexible capacity support among the N7, N5, and N3 nodes; the upcoming A14 node has 10-15 speed improvement at the same power or 25-30 power improvement at the same speed, and a nearly 20% chip density gain; the A14 node is on track and progressing well; management is seeing a high level of customer interest and engagmeent for A14; volume production of A14 is expected for 2028

Our practice is to prioritize the land in Taiwan to support the fast ramp of our new node due to the need for tight integration with R&D operations. Today, our new node, N2, has already entered high-volume manufacturing in the fourth quarter of 2025 with good yield. N2 is ramping successfully in multi phases at both Hsinchu and Gao Hsiung site supported by strong demand from both smartphone and HPC AI applications. With our strategy of continuous enhancement such as N2P and A16, we expect our N2 family to be another large and long lasting node for TSMC.

Historically, we do not add additional capacity to a node once it reached its targeted capacity. However, as a foundry, our first responsibility is to provide our customers with the most advanced technologies and necessary capacity to unleash their innovations. Based on our assessment, to meet the strong demand in AI application, we are stepping up our CapEx investment to increase our N3 capacity. Thus, we are now executing global capacity plan to support the robust multiyear pipeline of demand for 3-nanometer technologies, which are used by smartphone, HPC AI, including HBM based side, automotive and IoT customers. 

In Taiwan, we are adding a new 3-nanometer fab to our giga fab cluster in Tainan Volume production is scheduled for the first half of 2027…

…In addition to all the new fabs, we continue to convert 5-nanometer tool to support 3-nanometer capacity in Taiwan…

…We are also focusing on capacity optimization across nodes, which including flexible capacity support among the N7, N5 and N3 nodes…

…Figuring our second-generation transistor structure, A14 delivered another 4-node stride from N2, with performance and power benefit across to address the sensible need for high performance and energy efficient computing. Compared with N2, A14 will provide 10 to 15 speed improvement at the same power for 25 to 30 power improvement at the same speed and close to 20% chip density gain. Our A14 technology development is on track and progressing well. We are observing a high level of customer interest and engagement from both smartphone and HPC applications. Volume production is scheduled for 2028. Our A14 technology and its derivatives will further extend our technology leadership position and enable TSMC to capture the growth opportunities well into the future.

TSMC’s 2nd Arizona fab will utilise N3 technologies; the N3 nodes in the 2nd Arizona fab will begin volume production in 2027 H2; management has gained a lot of experience in Arizona, and expects to improve the cost structure of the Arizona fabs

In Arizona, our second fab will also utilize 3-nanometer technologies. Construction is already complete and volume production will begin in the second half of 2027…

…We already gained a lot of experience in Arizona. And so now we have much more confidence in last year that we can make good progress and moving aggressively forward and with, we expect we can improve the cost structure, of course.

TSMC’s management now plans to utilise N3 technology in the company’s 2nd fab in Japan; volume production is scheduled for 2028

In Japan, we now plan to utilize 3-nanometer technology in our second fab and volume production is scheduled in 2028.

TSMC’s management is open to including CPUs into its HPC (high-performance computing) AI calculation, but they will not do it right now, because TSMC is not able to tell where the CPUs it manufactures goes to

[Question] TSMC’s definition of AI revenue includes GPU, AI accelerator, HPM based maybe I up a few others, but it does specifically excludes data center CPU, I think you made that the definition very clear for a couple of years now. But with the CPU, there’s more and more conversation about CPU now becoming part of the AI infrastructure, especially for agentic workflows. Any chance for TSMC to maybe provide us revised numbers for AI revenue and maybe the AI revenue growth take a projection going 2029, 2030 and maybe hopefully give us some sense about the historical AI revenue numbers would have been if some of the data centers CPU numbers, especially for genetic AI workloads are included there.

[Answer] Certainly, CPUs becomes more and more important in today’s AI data center. But actually, let me share with you, this is a good question, by the way. Let me share with you that we are not able to identify which CPU goes to where, right? It’s a PC or it’s desktop or it’s AI data center. So today, we still not include the CPUs in our AI HPC’s calculation. Someday later, we might consider.

TSMC is working with NVIDIA for its next-generation LPU (language processing unit); the LPU comes with NVIDIA’s recent acqui-hire deal with Groq; Groq’s LPUs have historically been manufactured by Samsung

[Question] NVIDIA, of course, they recently added more CPU content to the overall but I think that most people are focusing on that brand-new LPU. They recently added — we understand I appreciate that the TSMC very strong institute and we’ll definitely participate in that upside in CPU. But the LPU business, it’s the acquired business, well, for historical reasons, it’s still at your competitors Samsung Foundry. And I think investors are looking at that and the thing that maybe looks like Samsung foundry finally made the first inroads into AI. So any thoughts from TSMC side, how should we think about whether and how TSMC will win back that LPU business or any future business coming from your customers?

[Answer] We are working with our customers for their next-generation LPU anyway. And we are very confident in our technology position, and we will work hard to capture every piece of business possible.

Tesla (NASDAQ: TSLA)

Tesla’s management is going to increase the company’s capital expenditure significantly, partly for AI-related investments; the increase in capital expenditure will last for a few years; management expects Tesla’s capex to be $25 billion in 2026, and thus cause the company to have negative free cash flow for the year

We’re going to be substantially increasing our investments in the future so you should expect to see significant — a very significant increase in capital expenditures, but I think well justified for a substantially increased future revenue stream…

…We’re investing in and improving our core technologies, battery powertrain, AI software, AI training, chip design, manufacturing — laying the groundwork for significantly increased manufacturing and production. We are also strengthening our supply chain across the board, batteries, energy, AI, silicon, everything, and laying the groundwork, like I said, for what we expect to be a significant increase in vehicle production in the future and, of course, a very significant increase — well, actually releasing Optimus…

…We are in a very big capital investment phase, which is going to start now and would last a couple of years. So based on that, our current expectation for 2025 — 2026 is over $25 billion of CapEx. And just to remind you, we are paying for 6 factories which were going to go into operation. Some have already started, some would go into operation later part of this year. We’re further increasing our investment in AI-related initiatives, including the AI infrastructure to support Robotaxi and the launch of Optimus. We’ve already started placing orders for the research semiconductor fab in Austin and for solar manufacturing equipment. While this may seem a lot and will have the impact of negative free cash flow for the rest of the year, we believe this is the right strategy to position the company for the next era.

Tesla’s management thinks Optimus can be useful outside of Tesla sometime in 2027; management continues to think Optimus will be the biggest ever product made; Tesla is preparing its Fremont factory for production of Optimus later this year; the production S-curve of Optimus will be very slow at the start, before ramping significantly in 2027; Tesla is building a 2nd Optimus factory, with production scheduled for mid-2027; v3 of Optimus (Optimus 3) is almost ready to be demonstrated, but management is hesitant because they have found competitors trying to copy Optimus’s design (in the 2025 Q4 call, management said Optimus 3 would be ready in a few months); management thinks Optimus can start production in July/August 2026, but it will take tremendous work to get there; management does not know what the production rate for Optimus will be in 2026; the production rate for Optimus will be limited by the slowest part in the entire Optimus supply chain; management wants to place a lot of intelligence locally in Optimus in the event that the robot loses wireless data; management thinks Optimus would need an orchestrator-AI and a voice AI, both of which can be Grok (a foundation model from one of Elon Musk’s companies, xAI)

But increasing our internal production for testing and then probably being able to have Optimus be useful outside of Tesla sometime next year. As you’ve heard me say a few times, I think, Optimus will be our biggest product — not just Tesla’s biggest product ever, but probably the biggest product ever. And I remain convinced of that conclusion…

…We’re preparing Fremont for start of production later this year with Optimus. Again, totally new supply chain, totally new technology. So therefore, the production S-curve is always very slow in the beginning, but it will ramp up to significant numbers next year. And we’re constructing a second Optimus factory in — at our Giga Texas location. And that will probably start production around summer next year.

The V3 Optimus design is almost ready to demonstrate. I think we want to just make sure it’s like polished. Like it works functionally, but there’s some aesthetic elements that need to be finalized. And I think probably middle of this year, we should be able to show it off. We’re also a little hesitant to show V3 off because we find our competitors do a frame-by-frame analysis whenever we release something and copy everything they possibly can. So I think there’s some value to not showing new technology until it’s close to production…

…We want to push the Optimus 3 unveil maybe closer to production. Start of production is — we’re assuming is somewhere around the late July, August time frame…

…The last S, X production will be in early May. But you have to look at the entire upstream portion of the production line. So you have to start with sales, battery packs, motor production, all the parts production. And so we’ve been dismantling the S, X production line from the more base-level parts — more basic level parts to — as you get to more larger subassemblies, you start dismantling the line from the small parts first, not from the final assembly first. So the final assembly line will — that will be dismantled next month and after the last of the S, X vehicle is done. You can’t dismantle some gigantic production line like overnight. It takes at least a few months to do so. And then you’ve got to install a new production line, and you’ve got to provide all of the wiring and communication, test out the machines of the new production line for Optimus. So that also takes several months. So frankly, if we’re able to go from stopping production on one line, dismantling that entire line, reinstalling a whole new line and turning that on in a matter of 4 months, that is an insanely fast speed. I don’t think any other company on earth has ever done that before…

…I don’t know what the production rate of Optimus will be this year. It is impossible to predict these things…

…when you have a brand-new product in an entirely new production line and you have 10,000 unique items, all of which have to go right into ramp production, it will move as fast as the least lucky, slowest, dumbest part in the entire 10,000. And this is a — Optimus is a completely new product with completely new production line. So it’s just literally impossible to predict, except that I think it will be quite slow at first as we iron out the 10,000-plus unique items that have to be sold for Optimus to reach volume production…

…We think we can put a lot of intelligence locally in the robot, and it certainly needs to be enough intelligence that if the robot gets disconnected, like if it’s a bad cellular signal or there isn’t WiFi, Optimus can’t just get stuck. It needs to have enough local intelligence that it can still do useful things even if it loses connection, kind of like a car…

…You can think of like Optimus needs kind of a manager to tell it what to do, broadly speaking, like if — otherwise it’s going to keep doing the same thing it did before. So I think you need kind of an orchestration AI, which Grok would be good for orchestration. And then for Optimus’ voice, having a low-latency intelligent voice AI, Grok is actually very good for that. So if you want to talk to Optimus and have kind of a Grok-level conversation, you kind of need to connect to a Grok-level AI for that.

All Tesla cars are autonomy-ready; supervised full self-driving is getting really good; v14.3 (version 14.3) of FSD was a major architectural update; management has a pipeline of improvements for FSD that they think will lead to unsupervised full self-driving being available globally; v15 of FSD is coming by end-2026 or early-2027; v15 of FSD will be a complete software architecture overhaul; v15 of FSD will run on Tesla’s AI4 chip; management thinks v15 of FSD will increase the safety level of FSD to way above human level; FSD now has 1.3 million paid customers globally (1.1 million in 2025 Q4); most of the growth in FSD customers in 2026 Q1 came from subscriptions, as management has removed the upfront-purchase option in some markets during the quarter; FSD recently received approvals in Netherlands; management is looking for EU-wide approval for FSD in 2026 Q2; FSD has received some approvals in China, although broader approval has yet to arrive; management hopes FSD can be fully approved in China by 2026 Q3; management has changed Tesla’s sales strategy to emphasise FSD as the product; management hopes to have unsupervised FSD in a dozen states by end-2026; management thinks unsupervised FSD revenue will not be material in 2026 but will be material in 2027; management thinks unsupervised FSD will reach customer-cars by 2026 Q4, but the release will be gradual; the FSD software deployed in Netherlands has the same exact architecture and the training procedure as the US version, but with more Europe data; management believes that the way Tesla solves full autonomy in the US can be applied to all parts of the world, if Tesla can add data from local regions; the Tesla customer fleet of vehicles is driving close to 10 billion miles on FSD in a few weeks; management thinks v14.3 of FSD is the last piece of the puzzle to enable unsupervised FSD; most Tesla drivers with Hardware 4 are already using FSD; FSD’s churn rate has improved

It’s always, I think, worth noting that a Tesla car is incredibly — incredible value for money, and they’re all autonomy-ready, depending on what part of the world you’re in. The supervised full self-driving is getting extremely good…

…For full self-driving and Robotaxi, version 14.3 was a major architectural update. And we have a whole pipeline of major improvements to full self-driving that, we believe, will lead to unsupervised full self-driving being available anywhere in the world that it is legal to do so. And then there’s a version 15, hopefully later this — hopefully by the end of this year, but certainly by early next year. And that will be a complete overhaul of the software architecture, and will run on AI4. That’s — and at that point, we’re really just increasing the safety level of FSD above human safety level, even more. Meaning, I think, even within version 14, we’re significantly safer than human, but v15 will take that to another level…

…On the FSD adoption front, we continue to see improvement, reaching nearly 1.3 million paid customers globally. The bulk of the growth came from subscriptions, while upfront purchases only increased 7% as we remove the purchase option in some markets in Q1.

We recently received approvals for FSD in Netherlands. This sets up us well for an EU-wide approval later in Q2, and we’re just gated by how the regulators go about it. Additionally, we’ve also received approvals in China. The broader approval is still not there, but we’re working with the regulators in the country, and we’re hoping that we can get approval by Q3…

…We have evolved our vehicle sales strategy, where we now emphasize FSD as a product and vehicle as only the delivery mechanism…

…We certainly hope to be — have unsupervised FSD/Robotaxi operating in, I don’t know, a dozen or so states by the end of this year…

…I think probably unsupervised FSD or Robotaxi revenue would not be super material this year. But I do think it will be material — it will be material probably in a significant way next year…

…[Question] When do you expect FSD unsupervised to reach customer cars?

[Answer] I’m just guessing here, but probably in the fourth quarter. It’s difficult to release this like to everyone everywhere all at once because we do want to make sure that they’re not unique situations in a city that particularly complex intersection or actually, they tend to be places where people get into accidents a lot because they’re just — perhaps there’s — and like I said, an unsafe intersection or bad road markings or a lot of weather challenges. So I think we would release unsupervised gradually to the customer fleet as we feel like a particular geography is confirmed to be safe…

…From a technology standpoint, what we deployed in Netherlands and Europe is the same exact architecture and the training procedure and so on, except it had more Europe data. And I suspect that same thing will be true for unsupervised FSD as well. Whatever we use to solve in the U.S. will work in other places and the rest of the world, too, provided we were able to add the data from the local regions…

…We are simultaneously solving the long tail of safety by monitoring the metrics across the entire Tesla customer vehicle fleet, which is close to driving 10 billion miles on FSD in the next few weeks…

…I think 14.3 is last piece of the puzzle for unsupervised FSD. Now the question is like degrees of safety. Like how — safety and convenience, I suppose…

…[Question] You have 180,000 new users, paying users this quarter, and I compare that to your overall installed base. It might be 15%, but then if I shrink that to the U.S. or to North America where most of them are, it’s probably more like 30%, 35%. And I’m trying to — and I compare that to what you sold, about 100,000 cars in North America in the quarter. So you’re winning twice more FSD users than you’re selling cars. And then if I add to that picture the fact that, I guess, it’s mostly Hardware 4 owners who subscribe to FSD, it sounds like most drivers in North America who have Hardware 4 would already be using FSD. Is that the right way to think about it and the kind of like success FSD is meeting today?

[Answer] You’re thinking about it the right way…

…We are actually seeing churn of subscribers also coming down, which again is a reflection of the product is getting better.

Tesla has started production of Cybercab, which are autonomous vehicles for the company’s Robotaxi fleet; the production of Cybercab will be a stretched-out S curve, ramping up only towards end-2026; the Robotaxi service has been expanded to Dallas and Houston; the expansion of the Robotaxi service is limited by management’s desire for really high safety levels; Robotaxi has, to-date, not had a single accident or injury; management hopes to have unsupervised Robotaxi in a dozen states by end-2026; management thinks Robotaxi revenue will not be material in 2026 but will be material in 2027; Robotaxi is currently running on FSD v14.3; Cybercab is 2-person vehicle; management thinks most of Tesla’s future vehicle production will be Cybercab; Tesla’s vehicles in the Robotaxi fleet sometimes get stuck because it’s programmed for maximum safety; the vehicles in the Robotaxi fleet can sometimes be stuck on infinite loops

We have just started production of Cybercab…

…Whenever you have a new product with a completely new supply chain, new everything, it’s always a stretched out S-curve. So you should expect that initial production of Cybercab and Semi will be very slow, but then ramping up and going kind of exponential towards the end of the year and certainly next year…

…We’ve expanded Robotaxi to Dallas and Houston using the same software source in the Bay Area. And the limiting factor for expansion is really rigorous validation, making sure things are completely safe. We don’t want to have a single accident or injury with the expansion of Robotaxi. And we have, to the credit of the team, not had a single one to date…

…We certainly hope to be — have unsupervised FSD/Robotaxi operating in, I don’t know, a dozen or so states by the end of this year…

…I think probably unsupervised FSD or Robotaxi revenue would not be super material this year. But I do think it will be material — it will be material probably in a significant way next year…

…So far, we have 0 incidents, and that’s what the NHTSA filing also shows…

…The version of Robotaxi that’s running in Austin, Dallas, Houston, et cetera, those are essentially 14.3 variants, and it’s obviously safe that, that’s why we’re able to launch in those cities…

…Cybercab is a compact vehicle. It’s actually — I mean, it’s very roomy, but it’s a 2-person vehicle. And we do think probably most of our production long term will be Cybercab because 90% of miles driven are with 1 or 2 people…

…A lot of what limits wider deployment of Robotaxi are actually not safety issues, but convenience issues or the car basically gets paranoid and gets stuck. Like sometimes it gets — because it’s programmed for maximum safety, so the problem is that then it sometimes just gets scared to do things. So like sometimes it gets scared to cross railroads, for example, or it’ll get stuck at a light or where there’s — the light never changes from red or, I mean, there was one kind of amusing situation where a whole bunch of Robotaxis got stuck in the left turn lane in Austin because, I kid you not, a Waymo had crashed into a bus. And so they could not turn left because the Waymo had crashed into the bus. And so you have this like long line of like, I don’t know, a dozen or more Tesla Robotaxis that were waiting for the bus to move, but the bus was never going to move because the Waymo crashed into the bus…

…We’ve also had literal infinite loops where the car might want to make a turn into a road, but there’s construction, and then it goes around the block, tries to turn into the road with construction, goes around the block, tries to turn into the road, and so you have to stop the infinite looping, the literal infinite looping.

Tesla has taped out its AI5 chip; management thinks the AI5 chip will be the best AI chip for inference at the edge, and will be the best value-for-money AI chip; Tesla is already designing the AI6 chip and is working on Dojo 3; management expects AI5 to go into Optimus and Tesla data centers, because AI4 is currently sufficient to achieve autonomy that is much safer than human drivers, so AI5 is not needed in the vehicle fleet; management thinks it will make sense at some point in the future to put AI5 into Tesla vehicles; management is planning to increase the memory and compute capacity of AI4, but the progress partly depends on Samsung (the fab for the chip)

Congratulations to — again to the Tesla AI chip team for taping out AI5. That’s going to be a great chip. I think probably the best AI inference chip for edge compute that exists. And certainly, I think, unequivocally the best value for money. The team did a great job. And we already have a lot of momentum for designing AI6, and we’ve begun to discuss ideas for Dojo 3…

…I do expect that AI5 will go into Optimus and into the data center because it’s looking like we’ll be able to achieve unsupervised self-driving with AI4 that is far greater than human safety levels. So — which means it’s not — certainly not immediately needed in the car. At some point, I think it will make sense for us to switch to AI5 in the car, but that’s — but there’s not a pressing issue to do so. So — but at some point, the AI4 hardware is going to get like so old that it’s like, okay, the only reason they’re keeping the factory open is for AI4.

We are planning an AI4 upgrade to use newer generation RAM. So it will go from 16 gigabytes to, I think, 32 gigabytes per SoC. So a total of 64 gigabytes, and probably a 10% increase in compute in sort of into — trillions of operations per second and in memory bandwidth. So that’s AI4.1 or AI4+, probably goes into production middle of next year, I think, depends. It depends on — Samsung is doing the modifications for us. So it sort of depends on when they’re able to finish that — finish those modifications and bring it to production.

Tesla’s management now thinks that Tesla vehicles with Hardware 3 will not be able to run unsupervised FSD; Hardware 3 has much lower memory capacity for Hardware 4, and memory capacity is needed for unsupervised FSD software to run; management is offering a trade-in for Tesla Hardware 3 vehicles to upgrade to Hardware 4; management is also considering setting up small factories to upgrade Hardware 3 on existing vehicles to Hardware 4

Unfortunately, Hardware 3 — I wish it were otherwise, but Hardware 3 simply does not have the capability to achieve unsupervised FSD. We did think at one point, it would have that, but relative to Hardware 4, it has only 1/8 of the memory bandwidth of Hardware 4. And memory bandwidth is one of the key elements needed for unsupervised FSD. And it’s just generally a thing that’s needed for AI. If you’re doing autoregressive transformer memory bandwidth, this is the choke point. So for customers that have bought FSD, what we’re offering is essentially a trade in — like a discounted trade-in for cars that have AI4 hardware. And then we’ll also be offering the ability to upgrade the car, to replace the computer, and you also need to replace the cameras, unfortunately, to go to Hardware 4.

So to do this efficiently, we’re going to have to set up like kind of micro factories or small factories in major metropolitan areas in order to do it efficiently. It’s — because if it’s done just at the service center, it is extremely slow to do so and inefficient. So we basically need like many production lines to make the change. And I do think, over time, it’s going to make sense for us to convert all Hardware 3 cars to Hardware 4 because that’s what enables them to enter the Robotaxi fleet and have unsupervised FSD.

Tesla’s research fab for the TeraFab project will begin construction this year at the company’s Giga Texas campus; management’s still working out details on TeraFab, which is a joint-venture between Tesla and other Elon Musk-related companies (xAI and SpaceX); the construction of the research fab will see Tesla spend around $3 billion, and the research fab is for Tesla to try out new ideas; SpaceX will be in charge of the initial phase of the scaled up TeraFab; Intel will be partnering the TeraFab for some of the core manufacturing technologies; TeraFab will utilise Intel’s 14A process, which is leading-edge but currently not fully mature; the TeraFab will be housing memory, logic, mask, lithography, and advanced packaging all under one roof, whereas the broader fab industry has separate facilities and companies for the different activities; management wants TeraFab to house all the different activities because they think it’s the fastest way to conduct R&D, but they are also aware it’s a long shot; management sees TeraFab as the only way to produce sufficient AI chips for the world, and not to press 3rd-party fabs on pricing; the TeraFab is also a great way for management to test out the radical ideas they have for improving AI chips

We’ve also finalized plans for the chip fab — the research chip fab on the Giga Texas campus, and we’ll start construction of that this year…

…We’re still working out the details of the Terafab deployment. In the near term, Tesla will be building the research fab on our Giga Texas campus. This is something we expect to be probably a $3 billion-ish initiative and capable of maybe a few thousand wafers per month, but it’s really intended to try out ideas, the research fab, both in terms of maybe — we have some ideas for improving the fundamental technology of how chips are made and some of the — there’s some new physics we’d like to test out. But we also want to test out the ability to see if something is working in production. So you need kind of like a few thousand wafer starts a month to make sure that a production process is sound. And then SpaceX is going to take care of like the initial phase of the scaled up Terafab. And that’s what we’ve figured out thus far…

…Intel is excited to partner with us on some of the core manufacturing technologies. So we plan to use Intel’s 14A process, which is state-of-the-art and, in fact, not yet totally complete. So — but given that by the time Terafab scales up, 14A will be probably fairly mature or ready for prime time. 14A seems like the right move…

…I think this will be unique in the world, or at least I’m not aware of any — a place where you have the lithography mask creation, the — and then logic, memory and packaging under one roof in one building. That’s about the fastest I could possibly imagine doing recursive research and development and being able to try out some pretty radical ideas, some of which have — it’s kind of long-shot stuff, but if some of these long shots pan out would be radical improvements in the way chips work…

…Terafab is not some sort of mechanism to generate leverage over our chip suppliers. It’s just literally we don’t see a path to having enough — a sufficient quantity of AI chips down the road as we scale production to high levels. Just the rate at which the industry is growing in logic, but even more so in memory, it just doesn’t — we just anticipate hitting the wall if we don’t make chips ourselves…

…I think that we do have some ideas for how to make maybe radically better AI chips. And these are kind of research ideas there — which means like long shot, but if long shot pays off, it’s maybe a giant improvement. And it’s just easier to do that if we have our own research fab and are developing our own production technologies. So — and if you look sort of long term at, say, having AI satellites, making chips for those. There’s just no way in hell the existing industry can keep up with that. It’s impossible.

Visa (NASDAQ: V)

Visa’s management believes agentic commerce will expand Visa’s market opportunity in 4 ways, namely, (1) accelerating the digitisation of commerce, (2) creation of significantly more transactions by agents, especially in a new category of commerce characterised by micro transactions, (3) accelerating the digitisation of B2B payments, with virtual cards and tokens becoming a preferred way to pay and be paid, and (4) accelerating overall GDP growth by 80-150 basis points

We believe AI and agentic commerce will expand our addressable market in 4 important ways. 

First, like eCommerce and mobile commerce before it, agentic commerce will accelerate the digitization of commerce around the world. And just like the acceleration from eCommerce and mobile commerce, Visa will benefit.

Second, agents will create significantly more transactions. Agents will intelligently split purchases across multiple transactions, optimizing price, timing and value to the buyer. And importantly, in some use cases, we expect agents will pay for their own data and resource consumption transaction by transaction and event by event, which creates an entirely new category of commerce with micro transactions.

Third, we will see accelerated digitization of B2B payments, where there is still enormous friction that AI agents can help remove. They will be able to automate payment initiation directly from invoices and contracts and manage approvals autonomously. In this context, virtual cards and tokenization will become a preferred way to pay and be paid.

And lastly, just like the advent of eCommerce and mobile commerce, agentic commerce will increase economic growth generally. Third parties estimate we are looking at a boost of 80 to 150 basis points of incremental GDP growth from AI and when GDP grows, spending grows and digital payments transactions grow.

Visa’s management believes the company is well positioned to win in agentic commerce for 3 reasons, namely, (1) the massive scale of Visa’s network, which means plenty of proprietary data to work with, (2) the tight security of Visa’s network, and (3) the high level of trust in it; Visa is a proven leader in tokenisation, and management believes tokens will become an essential element in agentic transactions; management thinks people will want their agents to pay with cards, just like how they prefer to use cards for physical and online payments; management recently launched Intelligent Commerce Connect, a network protocol and token vault agnostic on-ramp for agentic commerce; management is seeing early growth in agentic commerce transactions performed with Visa agentic tokens; management thinks the CLI (command line interface), which is effectively a chat box, is becoming a commerce platform, and cards will continue to have strong value in CLI-driven commerce transactions of all sizes; management recently launched Visa CLI as a proof-of-concept for developers to use their Visa credentials to make payments; early feedback for Visa CLI is very positive; management thinks agents will soon realise that no other payment methods, other than Visa cards, offer ease of use, broad acceptance, privacy, easy liquidity management, KYC, user security protection, and rewards; management thinks the limiting factor for agentic commerce is currently trust, which also means users will fall back on payment methods they already trust

Visa is extraordinarily well positioned to win in agentic for 3 important reasons. Our network, security and trust. Our network has enormous scale, more than 175 million seller locations, 5 billion credentials in 200 countries and territories with nearly 14,500 financial institution clients who have opted in to using this network. Payment security is only going to become more difficult and more valued. With our scale comes over 300 billion transactions annually, equating to an average of about 900 million transactions per day, and all of the data that comes with it. Visa has proven it knows how to manage transaction risk, identity risk and fraud, all enabled by this transaction data. And trust. Visa has well-established trust grounded in our standards and brand. We’ve set the standards that enable trusted payments in the digital and emerging agentic ecosystem.

And a big part of our network, security and trust are Visa tokens. Visa is a proven leader in tokenization, which is foundational in eCommerce and is set to become an essential element of trusted transactions in an agentic world.

People overwhelmingly choose to pay with cards face-to-face and online, and they will prefer their agents to pay with cards. And merchants want this, too. We recently launched Intelligent Commerce Connect, which acts as a network protocol and token vault agnostic on-ramp to agentic commerce for agent builders, merchants and enablers. Now while it’s early, we are seeing growth in agentic shopping and the emergence of early agentic commerce, real transactions with Visa agentic tokens.

And AI continues to evolve. With the AI landscape, we are seeing that Claude code and other agentic coding assistants will allow anyone to become a developer. It’s that easy to work in simple command-style tools like the command line interface, or CLI. These agentic coding assistants are a great example of how we see AI and agentic commerce increasing economic growth as they enable anyone to bring their new business ideas to life. We see a world where we will all design, build and launch digital products and experiences ourselves, engage with digital platforms and buy digital services using the CLI, or a slick consumer-friendly version of one as our interface. The CLI itself is becoming a commerce platform, and we believe that the preference and value of cards will be equally strong for all sizes of transactions, including micro transactions. A key to making this happen is enabling safe, simple and easy payments that are widely accepted by all API endpoints. We recently launched Visa CLI as a proof of concept, which shows how easy it is for a developer, soon all of us, to use their Visa credential to pay for digital services like an image, a website builder or more via the CLI. The early feedback we have been receiving from developers is very positive. And as we move forward, we plan to enable CLI commerce at scale, which means scaling the availability of command line tools and card acceptance by promulgating standards, products, rules and pricing…

…In all of these use cases, Visa cards are providing significant value. They’re easy to use, broadly accepted, integrated into the transaction flow, offer privacy, unlike most stablecoins, offer a way to manage liquidity in aggregate rather than funding millions of real-time micro transactions, offer an issuer KYC, user security protections if something goes wrong, and in many cases, cards offer rewards and benefits. We see no other payment method on earth that delivers all of these features. Buyers know this, sellers know this and soon so will agents. We expect more transactions, more value-added services and therefore, more revenue in the years ahead from agentic…

…I think the limiting factor for agentic commerce is trust. I think when we all think about ourselves as buyers and we all think about ourselves having agents go out and transact on our behalf, we are going to fall back on payment methods that we, as users, trust…

…When you think about yourself as a user, when you think about kind of who you’re going to trust your agent to make payments on your behalf, whether those are macro transactions, average transactions or micro transactions, we feel really good about our ability to win those transactions for our users using all of those capabilities.

AI is making Visa’s value-added services better; Visa’s new Large Transaction Model, which has a 5x increase in fraud value capture, is starting to be a foundational model for a variety of AI-powered fraud and risk services at the company; management has been integrating AI features across Visa’s VAS solutions; management thinks AI helps improve the differentiation of Visa’s VAS business even more; there are a variety of AI-driven products within the VAS portfolio that have helped the business perform well

Across Visa, AI is making what we do even better, especially for our value-added services. Our new Visa Large Transaction Model is beginning to act as the foundational model for a variety of AI-powered fraud and risk services at Visa. Early results have shown that it can power up to a 5x increase in fraud value capture. Our team has been integrating new AI-enabled features across our suite of VAS solutions, including the recent release of 6 dispute resolution capabilities. In fact, across all of our services, client adoption has been the fastest among AI embedded services such as Smarter Stand-In Processing and Visa Provisioning Intelligence…

…Our value-added services are highly differentiated and even more so in an AI world…

…We’ve been shipping new, especially AI-driven products in the issuing solutions space. We outperformed in the quarter in our AI-driven stand-in processing platform. We outperformed in our Visa supplier payment services platform. Those are two of the service — issuing solution platforms. In the acceptance side of the business, our Visa account updater platform outperformed. That’s one that allows merchants to automatically upstore credentials when you might have had fraud on your account and it was reissued or something like that. Look at our Risk and Security Solutions area, we saw outsized performance in VCAS, our Visa Consumer Authentication Service, or also in our VAA and VRM platforms, Visa Advanced Authorization and Visa Risk Manager. These are all products that we’ve been deploying in market, largely AI-driven products, and they’ve been driving broad-based out-performance across the value-added services portfolio.

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, Intuitive Surgical, Mastercard, Meta Platforms, Microsoft, Netflix, TSMC, and Visa. Holdings are subject to change at any time.

What We’re Reading (Week Ending 03 May 2026)

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 03 May 2026:

1. Oracle’s Deluge of AI Debt Pushes Wall Street to the Limit – Peter Rudegeair and Berber Jin

Banks including JPMorgan Chase struggled for months to spread the risk of billions of dollars in loans they made to build data centers leased to Oracle in Texas and Wisconsin, people familiar with the matter said. Many financial institutions that would ordinarily buy those loans face restrictions on how much exposure they can have to a single counterparty, and the sheer size of these debt packages pushed them to the limit with Oracle. As a result, bank balance sheets got clogged, constraining the financing prospects of future projects tied to Oracle and OpenAI.

For example, lenders balked at financing the expansion of a data-center complex in Abilene, Texas, if Oracle were the tenant, according to people familiar with the matter. That led the developer, Crusoe, to lease it to Microsoft instead…

…Lenders grew more comfortable with Oracle-related projects after the company said it would raise all the money it needed for 2026 by issuing roughly $50 billion in stock and bonds. Oracle said in a post on X last week that each data center it is developing for OpenAI is moving forward on time.

But even after it raises that amount, Oracle still has additional cash funding needs of $100 billion or more for 2027 and the first half of 2028, according to Morgan Stanley credit analysts. “We’ve pondered how [Oracle’s] considerable funding needs over the next three years may test the depths of different fixed-income markets,” the analysts wrote in February…

…Oracle, though, is in a comparatively weaker financial position than big tech rivals. It has a lower investment-grade credit rating, more debt and is burning cash. Much of its future revenue is tied to a money-losing startup that is facing growing competitive pressure. The cost of protecting Oracle’s bonds against a potential default via credit-default swaps roughly quadrupled between late September and late March, though it has fallen slightly since then…

…Much of the borrowing tied to the OpenAI megacontract was done by projects involving data center developers working with Oracle. The debt was structured as short-term construction loans meant to be syndicated among a group of banks and other institutions. Oracle is the tenant and OpenAI is the subtenant on the deals, but the debt doesn’t sit on Oracle’s balance sheet.

2. OpenAI Misses Key Revenue, User Targets in High-Stakes Sprint Toward IPO – Berber Jin

Chief Financial Officer Sarah Friar has told other company leaders that she is worried the company might not be able to pay for future computing contracts if revenue doesn’t grow fast enough, according to people familiar with the matter. 

Board directors have also more closely examined the company’s data-center deals in recent months and questioned Chief Executive Sam Altman’s efforts to secure even more computing power despite the business slowdown, the people said…

…OpenAI missed an internal goal of reaching one billion weekly active users for ChatGPT by the end of last year, according to people familiar with the goals. The company still hasn’t announced that milestone, unnerving some investors. It also missed its yearly revenue target for ChatGPT as well after Google’s Gemini saw massive growth late last year and ate into OpenAI’s market share, the people said. The company has also struggled with defection rates among subscribers, according to people familiar with those figures.

OpenAI missed multiple monthly revenue targets earlier this year after losing ground to Anthropic in the coding and enterprise markets, people familiar with its finances said.

3. If AI is so great, why isn’t it working? – Vas M.

AI is working for one group of people right now, at scale, because it’s the group of people that rely the least on business logic. It’s software engineers. The biggest winner from 18 months of AI improvement, by miles, has been engineers writing code in Cursor, Claude Code, Codex, etc. Some stats for you if for some reason you still don’t believe in agentic engineering:

  • GitHub’s 2024 study clocked Copilot users at 55% faster on real tasks. 1 hour 11 minutes vs 2 hours 41 minutes on the same work.
  • Anthropic ran an internal study in August 2025 across 132 engineers and 100,000 real Claude conversations. AI cuts developer task completion time by roughly 80%.
  • Sundar Pichai said at the start of 2026 that 75% of new code at Google is AI-generated and engineer-approved. That number was 30% in April 2025.

Yes, the tools still overpromise on the hard stuff: security review, complex distributed systems, novel debugging. Caveat very real and noted. But the bread-and-butter productivity gain on shipping code is the biggest jump engineering has had since the IDE…

…So why does AI work for engineers and not for any of these? What’s different about engineers? As a former software engineer, engineering work has four properties that basically no other enterprise function has. Yes there are nuances but these are directionally correct, please relax in the comments.

  • It’s bounded. A function takes inputs and returns outputs. The scope of “fix this bug” lives inside a file or a module. The dependencies are explicit and importable.
  • It’s checkable. Compilers tell you in milliseconds whether the code parses. Tests tell you whether it works. Type systems catch entire classes of error before runtime. Feedback loop: seconds.
  • The substrate is structured. Code lives in files, in version control, with a deterministic build pipeline underneath. Same input, same output. You can replay any state.
  • The output is verifiable. A pull request is a discrete artifact. A reviewer can look at the diff in 10 minutes and say yes or no.

When you point a capable AI at work that’s bounded, checkable, structured, and verifiable, the leverage is enormous. Cursor and Claude Code are the proof. And if we’re being honest, the biggest reason is that the AI labs (OpenAI, Anthropic, Cursor) poured every single ounce of resources they had into figuring out software engineering. If they can make their own engineers better, they can make the models better, faster, and achieve the ever-elusive “AGI”, which will then make every other task on the planet (Finance, Sales, Operations, Marketing, etc) much easier downstream.

But contrast software engineering with a finance close.

Finance involves AP, AR, intercompany reconciliations, FX, accruals, journal entries, and exception handling that spans NetSuite, Concur, three banks, two ERPs from acquisitions, a custom intake form, and a Slack channel where the controller flags “weird stuff she sees.” The “process” is documented in an SOP that doesn’t match what actually happens. The output is “the close was clean,” which takes two senior accountants two days to verify.

Sales ops involves a CRM, an outbound tool, a calendar, a notes platform, an enrichment vendor, an attribution tool, and a Slack channel where the AE is asking the CRO whether to discount this deal. None of those systems share state cleanly. The process for qualifying a lead is different across reps, even on the same team.

This is what every ops function looks like in every company Varick has ever audited. None of it is bounded, checkable, structured, or verifiable the way code is. And trying to wrangle generic AI to these functions that are incredibly specific to your company and its processes is a fools errand.

Pointing an LLM at this work gives you negative ROI. The operator was doing the work in 30 minutes. Now they’re doing the work in 30 minutes plus another 30 minutes correcting the AI’s mistakes. Most if not every vendors’ “AI for [department]” has the same arc. A nice flashy demo showing how great it works for startups, then a big series A, then quietly killed after it fails to work for enterprise…

…Ok so what does the 5% that ships and stays in production do consistently that makes them so good:

1. They audit before they build. Four weeks (often longer) of mapping the actual workflow before anyone touches a model. The audit produces a digital twin: a live map of how work moves through the org, where the conformance gaps are, what’s pattern-matchable, and what genuinely needs human judgment. The document itself matters less than the alignment it forces between the AI team and the operators. Make sure everyone is aligned on what the bottle-necks are, what the optimal state should be, and what is going to be done to fix it.

2. They decompose the work until most of it is deterministic. LLM goes ONLY where judgment is absolutely required, while plain code goes everywhere else. Most production systems we ship at Varick end up as 5-10 deterministic steps with maybe one or two model calls in specific places. Boring in production is genuinely the goal, and is how we’ve seen the most success.

3. They build a single orchestration layer that sits on top of the existing software stack. At Varick, we call this the single pane of glass. Finance, sales, ops, and engineering agents all live on the same platform, share the same context, and can talk to each other when they need to. Every new use case lands as configuration on top of the platform. In turn, sprawl is dead on arrival.

4. They stay model-agnostic. Abstractions get built at the task level, not at the model level. Each step routes to the best-fit model at any given moment. When OpenAI deprecates a model or Anthropic ships something dramatically better, the routing layer absorbs the change and your workflow keeps running without anyone noticing.

5. They treat the deployment as continuously evolving infrastructure. There is a real team responsible for ongoing tuning, retiring agents that aren’t earning their keep anymore, and shipping improvements every quarter. The deployments that pay off over five years are the ones that get tuned every quarter if not every month, not the ones declared “done” at go-live. You have to get over this fact if you want to succeed with AI. 

4. Software Is Eating the World (But Actually This Time) – Siddharth Ramakrishnan

In 2011, software ate the world. At least that’s what Marc Andreessen told us. But if that’s true, then why does the Bay Area still exist? If software really ate everything, wouldn’t we all have moved to New York or Miami by now?

Well, let’s look at what software actually ate: banks got apps, retail got websites, hospitals got EHR systems, and taxis got dispatched with a few taps instead of a phone call at 2am when you maybe don’t remember exactly where you are.

Software ate the interfaces, but the actual work? That mostly stayed human.

A customer calls about a billing dispute and software routes the call, pulls up the account screen, and then logs the resolution afterward. But here a person is still the one listening, figuring out whether the refund policy applies here, deciding what to do, and actually talking to the customer. A loan officer reviewing an application gets the credit score surfaced by software and the documents pulled up on screen, but they’re the one reading those documents and making the judgment call. For 15 years, software has been really good at the plumbing while humans kept doing the actual work.

Now, AI can actually do the work! A customer service call is becoming an agent loop where the system handles speech recognition, looks up the account via API, pulls the relevant policy, reasons about whether the customer qualifies, triggers the refund, and responds with text-to-speech. An insurance claim is becoming document intake followed by coverage checks, fraud flags, reserve calculations, and settlement workflows, all running as code. A coding task is already 30 rounds of reading files, editing code, running tests, and revising with no human involved at all…

…I think most people dramatically underestimate how much inference these converted workflows actually consume, because they’re picturing one model, one call, one response, and some hallucinations along the way, but the reality is very different.

Take a voice support agent handling something simple but real, like rescheduling a medical appointment. To the customer, it feels like one conversation. Under the hood, it is a small autonomous system running continuously. As the caller speaks, a speech recognition model transcribes audio in real time. An orchestration model then reasons over the transcript, pulls the patient record, checks scheduling constraints, looks up provider availability, decides what to ask next, and calls the relevant tools. Once it has enough information, it synthesizes the result into a response, and a text-to-speech model turns that back into natural audio. In parallel, other models may be monitoring sentiment, checking compliance, or deciding whether the call should be escalated.

The system is doing all the work itself: listening, retrieving, deciding, tool-calling, verifying, and responding in a loop. An 8 minute call might contain only ~3k tokens of raw transcript, but the orchestration layer can easily consume ~40k tokens once you account for repeated reasoning over the growing conversation, retrieved context, and tool outputs, on top of continuous ASR and TTS inference running for the duration of the call. “One AI phone call” is really a multi-model inference stack operating continuously…

…In customer support, a basic FAQ bot in 2023 might have consumed around 3,500 tokens for a ticket, better retrieval pushed that higher, then tool use and reasoning pushed it higher again, and now full voice support stacks are higher still. Coding follows the same pattern, just more violently: what used to be tens of thousands of tokens for a bounded coding task has become hundreds of thousands or even well over a million as agents became capable enough to handle real debugging, refactoring, and multi-file work. Each useful task now justifies much more inference than it did a year or two ago, because the model can actually finish the job.

This is a subtle version of Jevons paradox. The sticker price per token has actually been rising for frontier models, not falling. But the value per million tokens has gone up much faster: a frontier model today can complete a workflow in one coherent session that would have required dozens of brittle attempts a year ago, or simply could not have been done. Effective cost per useful outcome is dropping even as nominal cost per token climbs. And that dynamic is what opens up entirely new categories: complex insurance claims, broad code refactors, long-running research tasks, multi-step back-office processes. These were not meaningfully part of the inference market two years ago because the models could not stay coherent long enough to do them.

The aggregate numbers suggest this is already happening. OpenAI’s API is processing more than 15B tokens per minute as of April 2026, up from 6B half a year earlier. Google went from 9.7T tokens per month to 480T in a year, about 50x growth. OpenAI says reasoning token consumption per enterprise organization grew 320x year over year. Anthropic’s latest reported annualized revenue of $30B (up from $10B to start the year…) speaks for itself, especially given the main driver is Claude Code and their API…

…As models commoditize, the durable application companies will be the ones that see the real work: the tool calls, retries, escalations, corrections, and edge cases that never show up in a benchmark. That is where the system learns how a specific workflow actually runs, and where proprietary context starts to accumulate. Over time, the advantage is not just access to a model. It is knowing how this insurer handles claims, how this hospital works denials, how this codebase breaks, how this finance team closes. The apps that capture that messy operational data will be the ones that improve fastest and defend their position longest.

5. Nike and the Arithmetic of Durability – Andrew Chou

As of April 2026, Nike stock sat below US$45 – a market capitalisation of US$68 billion, its lowest level in over a decade, and a fall of more than 75% from the US$280 billion the company commanded at its 2021 peak.

How does what was once considered one of the widest consumer brand moats in the world, built over half a century, erode over the course of a few short years?

A good starting point is January 2020, when John Donahoe took over as Nike’s new CEO. The board wanted a digital-first operator, and Donahoe had the résumé – ServiceNow, eBay, and Bain – even if he was one of the few leaders in Nike’s history not to have risen through its operating ranks…

…Under Donahoe, Nike began systematically pulling back from these wholesale relationships. The logic was straightforward: move more volume through direct channels, control the brand experience, and capture more margin.

By September 2021, Nike had exited roughly half its retail partners. Big names like Foot Locker, Zappos, Dillard’s, and Big 5 Sporting Goods saw their allocation of the most sought-after models shrink in favour of Nike’s directly owned stores. Gross profit margins expanded immediately.

The vacated shelf space that followed was quickly and eagerly filled by competitors. Adidas, New Balance, Puma, Hoka, On, Brooks, and Salomon—brands that had suddenly found themselves with prime real estate in the stores Nike had walked away from…

…That same model of deep, sport-specific immersion was eventually replicated across basketball, football, tennis, and dozens of other categories. Teams embedded in each discipline accumulated years of insight about athletes, usage patterns, and the fine distinctions that matter in performance products. This kind of expertise accumulates slowly—through proximity to athletes, coaches, biomechanics, and the subtle demands of each sport.

Under Donahoe, Nike restructured around a simpler model: Men’s, Women’s, and Kids. The rationale was familiar—less duplication, cleaner accountability, more consistency across segments—and the resulting redundancies left the org chart looking tidier on paper. Overhead expenses came down immediately.

What it also did was dissolve the sport-by-sport expertise and institutional knowledge accumulated over decades. Product lines that had once been shaped by deep category knowledge were now filtered through broader consumer-demographic lenses…

…Nike has long been famous for marketing that built meaning before it chased sales. The ability to turn a product into a cultural moment was arguably Nike’s most valuable and least replicable asset.

The Banned Air Jordan story is perhaps the purest illustration. In 1984, Michael Jordan wore black-and-red sneakers that violated the NBA’s uniform rules. The league threatened fines. Nike’s response was not to comply—it was to lean in. The company shot a television commercial showing the shoes blacked out by censorship bars, declaring that the league had thrown them out of the game but could not stop you from wearing them. That single ad helped sell 50,000 pairs almost immediately…

…Under the new model, marketing spend shifted from broad, culture-shaping storytelling into programmatic digital advertising designed to drive traffic to Nike’s own e-commerce channels. Performance marketing has direct, measurable KPIs – but by its nature, it harvests existing demand rather than creating it.

Anyone can pay for web traffic, but doing so does not build a competitive advantage. Just ask the direct-to-consumer startups built on performance marketing in the 2010s that failed to sell to a large incumbent with real distribution before the music stopped…

…Nike shares climbed from around $100 when Donahoe took over to an all-time high of $179 in November 2021 – a company valued at roughly $280 billion. The “transformation” was working.

But these gains came from somewhere. They were, in effect, the monetisation of business value painstakingly built over decades: the distribution footprint Knight and his team had cultivated since the 1960s; the product expertise and institutional knowledge that Bowerman’s culture had embedded across dozens of categories; the brand equity that campaigns like the Banned Air Jordan and Just Do It had compounded over generations.

Most business decisions sit on a spectrum between maximising long-term net present value and maximising short-term accounting profit. When the asset being spent is the moat itself, the spending does not show up as a cost. Each of Nike’s three shifts boosted reported profitability immediately and reduced the long-run NPV of the franchise meaningfully. The trajectory of the income statement and the moat moved in opposite directions – but only the income statement was visible quarter to quarter.

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 View On Consumer Spending From The Largest Payments Companies (2026 Q1)

Mastercard and Visa can feel the pulse of consumer spending – what are they seeing now?

Mastercard (NYSE: MA) and Visa (NYSE: V) are two of the largest payments companies in the world. As a result, they have a great view on consumer spending that’s taking place. With both companies reporting their earnings results for the first quarter of 2026 earlier this week, the bottom line is that consumer spending remains strong in the USA and other parts of the world, although there’s some near-term uncertainty because of the current conflict in the Middle East. Here’s what they are seeing.

*What’s shown in italics between the two horizontal lines below are quotes from Mastercard and Visa’s management teams that I picked up from their earnings conference calls.

From Mastercard

1. Mastercard’s management sees consumer and business spending, and the labour market, remaining healthy, although the economic backdrop is uncertain, driven by geopolitical tensions in the Middle East that have affected cross-border travel and global energy supply

Looking at the macro picture, the economic foundation remains generally supportive, with healthy underlying consumer and business spending. However, the backdrop remains uncertain, driven by geopolitical tensions, which has put some pressure on cross-border travel. Overall, labor markets continue to be balanced and wages are still outpacing inflation in most major markets…

…Despite elevated geopolitical risks, the macro economy has remained largely supportive, with healthy, underlying consumer spending and the fundamentals of our business remain strong. With that said, we are operating in a period of heightened uncertainty magnified by the ongoing conflict in the Middle East. Since the outbreak of the conflict at the end of February, we have seen restrictions on travel and a reduction in the world’s energy supply. And as I noted earlier, we are seeing impacts from that in our cross-border travel metrics.

2. Worldwide GDV (gross dollar volume) was up 7% year-on-year in constant-currency basis; cross-border volume was up 13% globally in constant-currency, driven by both travel and non-travel cross-border spending (cross-border volume growth was 14% in 2025 Q4); cross-border volume in 2026 Q1 was affected in March because of impacts on cross-border travel from the conflict in the Middle East; switched transactions was up 9% year-on-year; card growth was 5% in 2026 Q1, with Mastercard ending the quarter with 3.7 billion cards in circulation (there were 3.7 billion cards in 2025 Q4, and year-on-year growth was 6% then); on currency-neutral basis, domestic assessments were up 6%, cross-border assessments were up 18% and transaction processing assessments were up 15%

I’ll speak to the growth rates of our key volume drivers for the first quarter on a local currency basis. Worldwide gross dollar volume, or GDV, increased by 7% year over year. In the US, GDV increased by 4%, with credit growth of 8% and debit growth of 1%. Excluding the impacts from the migration of the Capital One debit portfolio, our US debit GDV growth would have been 7%…

…Outside of the US, volume increased 9% with credit growth of 9% and debit growth of 8%. Overall, cross-border volume increased 13% globally for the quarter, reflecting continued growth in both travel and non-travel related cross-border spending. As one would expect starting in March, we began to see some impact on cross-border travel from the conflict in the Middle East.

…Switched transactions grew 9% year-over-year in Q1. Excluding the impacts from the migration of the Capital One debit portfolio, our switched transaction growth would have been 10%…

…Card growth was 5%. Globally, there are 3.7 billion Mastercard and Maestro branded cards issued…

…All growth rates are described on a currency neutral basis, unless otherwise noted. Looking quickly at each key metric, domestic assessments were up 6% while worldwide GDV grew 7%. The difference is primarily driven by mix, partially offset by pricing. Cross-border assessments increased 18%, while cross-border volumes increased 13%. The five point difference is driven primarily by pricing in international markets. Transaction processing assessments were up 15%, while switch transactions grew 9%. The six ppt difference is primarily due to favorable mix and pricing, slightly offset by lower revenue from FX volatility.

3. In 2026 Q1, Mastercard’s operating metrics had good year-on-year growth and were stable sequentially; in April 2026 so far, Mastercard’s operating metrics continue to be strong with worldwide switched volume growth of 8% (5% in the USA, and 10% outside of the USA), switched transactions growth of 9%, and cross-border volume growth of 9%; cross-border travel volume declined sequentially in April 2026 from 2026 Q1 because of an acceleration in the impact of the Middle East conflict; in all, management continues to see healthy consumer and business spending

Let me comment on the operating metric trends for Q1 and the first 4 weeks of April. As we look across Q1 and April, growth rates of our operating metrics were impacted by timing of holidays, namely Ramadan and Easter. March would have seen the benefits from the timing, while February and April saw a negative impact. Looking at the Q1 operating metrics on a sequential basis, switched metrics were generally in line with Q4 and underlying spend remains stable. Of note, U.S. switched volume was flat sequentially as the strength in consumer and business spend offset the impact from the migration of Capital One’s debit portfolio in the quarter. Excluding Capital One, on a like-for-like basis, U.S. switched volume growth was over 1 ppt higher in Q1 as compared to Q4. Now on to switched transactions; excluding the migration of the Capital One debit growth — sorry, excluding the migration of Capital One debit, growth was generally in line with Q4.

Moving to our cross-border metrics. Our overall cross-border volume remains healthy with growth at 13% in the first quarter. Cross-border card-not-present ex-travel grew at 18% and remained strong. And the sequential decline in cross-border travel was due primarily to the conflict in the Middle East and portfolio shifts.

Now looking specifically at cross-border travel for the first 4 weeks of April, the sequential decline from Q1 is due to an acceleration of the impact of the conflict, the portfolio shifts and the negative impact from the timing I just mentioned. None of these factors relate to any fundamental change, and underlying consumer and business spend remains healthy.

From Visa

1. US payments volume growth was good at 8%, with e-commerce growing faster than physical spend, and it reflected resilience in consumer spending; there was good growth in both US credit and debit volumes; growth across consumer spend bands improved from 2025 Q4 (FY2026 Q1) with the highest spend band continuing to grow the fastest; both discretionary and non-discretionary spend remained strong; management did not see a deterioration in spend in the lower bands; 

U.S. payments volume grew 8% year-over-year, up almost 1.5 points from Q1, reflecting resilience in consumer spending. E-commerce spend outpaced face-to-face spend. Both U.S. credit and debit demonstrated broad-based spend improvement, and we believe both were helped in part by higher tax refunds. Debit grew 7%, up almost 1 point from Q1 and credit grew 10%, up more than 2 points from Q1, with strong travel spend in both consumer and commercial.

Growth across consumer spend band saw incremental improvement from Q1 with the highest spend band continuing to grow the fastest. Across our volume, both discretionary and nondiscretionary spend remains strong. We do not see signs of the lower spend consumer weakening in our volumes.

2. Visa’s cross-border volume growth remained strong in 2026 Q1 (FY2026 Q2) at 11%, and was the same as in 2025 Q4 (FY2026 Q1)

Q2 total cross-border volume was up 11% year-over-year, consistent with Q1. Cross-border eCommerce volume was up 13%, 1 point above Q1. While crypto continued to be a slight drag, the improvement was primarily driven by U.S. inbound volume. Travel-related cross-border volume was up 10%, generally consistent with Q1, led by continued strength in commercial and improved U.S. inbound volume that generally offset the impact in the Middle East that was most pronounced in March.

3. Payments volume on Visa’s network continues to grow in April 2026, with US payments volume up 9%, cross-border volume up over 9%, e-commerce volume up 14%, and processed transactions up 8%; management is seeing near-term uncertainty in cross-border travel spend in the CEMEA (Central Europe, Middle East, and Africa) region because of the Middle East conflict

Now let’s look at drivers through April 21 with volume growth in constant dollars. U.S. payments volume was up 9%, with credit up 10%, and debit up 8% year-over-year. For constant dollar cross-border volume, excluding transactions within Europe, total volume grew 9% year-over-year with eCommerce up 14% and travel up 5%. The step down in travel from March was driven by both the impact from the Middle East conflict and Ramadan timing. When you normalize for Ramadan timing, the total April cross-border volume growth was in line with February levels. Processed transactions grew 8% year-over-year…

…The Middle East conflict has introduced some near-term uncertainty, in particular to cross-border travel spend in the CEMEA region.

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 currently have a vested interest in Mastercard and Visa. Holdings are subject to change at any time.

What We’re Reading (Week Ending 26 April 2026)

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 26 April 2026:

1. Pancreatic cancer mRNA vaccine shows lasting results in an early trial – Kaitlin Sullivan, Marina Kopf and Anne Thompson

Nine days later, Gustafson had surgery to remove the Stage 2 cancer from her pancreas. The day before she was supposed to start chemotherapy, her doctors told her about a clinical trial exploring the use of personalized messenger RNA vaccines for cancer. It was February 2020 — months before mRNA vaccines for Covid would become one of the world’s hottest commodities. Very soon after, Gustafson was the first person to get one for pancreatic cancer.

“It was a no-brainer,” Gustafson said of joining the trial. “I knew that statistically, the odds were against me.”

Less than 13% of people diagnosed with pancreatic cancer live for more than five years, making it one of the deadliest cancers. There is no routine screening for pancreatic cancer, such as colonoscopy or mammogram, and symptoms typically don’t show up until the disease is advanced. Once detected, there are few options for treatment. Only about 20% of cases are operable, which is currently required for someone to be eligible to join a pancreatic cancer vaccine trial…

…The vaccines work as a type of so-called immunotherapy, harnessing a person’s immune system to fight cancer cells. The goal is not to eliminate existing tumors, but instead to stamp out lingering, undetected cancer cells, and later any new cells that form before they can cause a recurrence.

…Pancreatic cancer is the poster child for these difficult-to-treat cancers, Balachandran said, and experts have long believed that people with pancreatic cancer could not generate an immune response against tumors. But after nine doses of the personalized vaccine, Gustafson is one of eight people in the 16-person Phase 1 trial who did just that, producing an army of immune cells called T cells that seek out and destroy tumor cells.

“This is one of the hardest cancers to generate any immune response, let alone such a potent one,” Balachandran said.

Balachandran and his team published the results of the Phase 1 clinical trial last year. At the time, the patients, all of whom had early-stage disease before they joined the trial, had only been tracked for just over three years, and it was unclear whether the immune response would last and lead to the patients living longer, he said. New data collected during the trial’s six-year follow-up period shows that it may.

Six years after treatment, Gustafson and six others who responded to the treatment are still alive, along with two of the eight people who did not respond. Two of the responders, including the one who died, had a cancer recurrence; Gustafson’s cancer has not come back.

“The most important finding here is that the people who mount a response to the vaccine live longer than those who do not,” said Dr. William Freed-Pastor, a physician-scientist at Dana-Farber Cancer Institute, who was not involved with the trial. He cautioned, however, that the results come from a very small group of patients. More research is still needed…

…Earlier research tested mRNA vaccines to treat people with advanced cancer, with disappointing results, “so we thought we didn’t have a vaccine that would work,” said Dr. Robert Vonderheide, the president-elect of the American Association for Cancer Research and director of the Abramson Cancer Center at the University of Pennsylvania.

In reality, newer research like this Phase 1 trial suggests the immunotherapy may work in less advanced cancer.

2. Brad Setser on the War in Iran and the Future of the US Dollar (Transcript Here) – Tracy Alloway, Joe Weisenthal, Brad Setser

Tracy Alloway: Why don’t we start with that historic analogy—the 1970s oil shock. Lots of ink is currently being spilled on whether or not that’s the correct parallel for our current crisis. In your view, how much does this particular oil shock resemble that of 50 years ago?

Brad Setser: There’s the obvious parallel in the sense that the 1970s oil shocks—’73 was a function of the Yom Kippur War and the Arab nations’ reactions to it. The second oil shock in 1979 was a function of the Iranian revolution. Same geographic region, but different in the sense that the US and Israel are the instigators, and different in that, so far at least, the magnitude of the shock is not at all comparable. It’s not at all comparable in price terms. In ’73 and then in ’79, oil doubled or tripled, and by the end of the decade oil had gone up six or seven times in dollar terms, less in real terms. We’ve only gone up maybe 50% max from spot oil for Brent and WTI and next month’s future. It’s a little higher for delivery in Asia, but we are not yet at the magnitude of the shock we saw in the 1970s.

The obvious point is that our economy as a whole—for the US and for the world—is a little less oil-dependent, but I wouldn’t push that too far. The main distinction is that we sort of started it—the US and Israel—and we in theory can end it, although we would only end it if Iran finds its own equilibrium that allows other countries’ oil to pass through the strait. At least so far, the market has not anticipated that this will need the same kind of jump in price to balance supply and demand. That could change. If you look at it in terms of physical interruption of the flow of oil, some of your guests have noted we’re similar, maybe even worse. So we’re in this weird world where the physical interruption is bigger but the price reaction is smaller.

Joe Weisenthal: I’m glad you brought this up. You talk to the commodity guys like we do, and they’re saying, “This is crazy, this is the biggest shock ever.” Guys like me—I’m an efficient-markets guy, I just see what’s on the screen, and it looks like it’s not that big of a deal. You be the third-party arbiter here. How do you make sense of the gap between what we see on our screen versus the shortfall in physical barrels—20 million every single day that aren’t coming to the market?

Brad Setser: It’s not quite 20. You’ve got the East—there’s been some rerouting. It’s somewhere between 10 and 15, which happens to be between 10 and 15% of global supply and between 20 and 30% of global traded oil. It is still a massive, massive shock, and my elasticities would imply a much bigger increase in price if that was a sustained, expected interruption.

You end up dealing with the reality that oil is close to being a perfectly fungible commodity, but it is not a perfectly fungible commodity. A North Atlantic barrel can only get to China or Japan with a long trek around the world, so there’s an extra shipping cost. A lot of the barrels in the North Atlantic are sweet and light—”light” is a measure of the weight of the oil, “sweet” means less sulfur. A lot of the refiners in Asia were set up to refine medium sour. For some things you want heavier grades of oil because you get more diesel out of the heavier grades. Refiners are configured for different grades of oil. When you interrupt the flow—fundamentally the flow from the Gulf countries to Asia—there’s no immediate, instantaneous substitution using barrels from the North Atlantic. That’s the first point.

The second point is that what people think of as traded oil is not actually oil for delivery tomorrow. It is the futures contract for the next month, and the month after that, the world could look completely different. The US has within its ability the capacity to pull back. If the US pulls back—and maybe the Iranians insist on a toll—there is no shortage of oil that could come out. It’ll take a little longer now because of the physical destruction of some of the export facilities in the Gulf, but if you don’t have this particular choke point strangled, the old global oil market was very well supplied. So the futures market has to balance between one possibility—that there is plenty of oil two or three months out and oil is on a trajectory, not immediately because of the damage, back to $60—and another possibility where this persists and oil is at $150 or above. The market’s had trouble figuring that one out…

…Joe Weisenthal: There are obviously differences, but how did the ’70s reshape the world? You had these oil shocks, and people then started talking about “petrodollars”—a word that came into existence. What kind of legacy did those shocks leave on the global financial system?…

…Brad Setser:  Americans are very unhappy—if you remember in the 1970s it was not good for President Carter. When the Iranian revolution came, there were the hostages, but the oil shock did not help his popularity. Americans in general are very unhappy when oil prices are high—it’s one of our national quirks.

In the short run at that time, there was a huge windfall into the Gulf states. The Gulf states piled up dollars, and they were dollars. Most oil—oil was priced in dollars before 1973. It didn’t take a deal to price oil in dollars. The US had been the biggest producer of oil in the 1930s. We were the supplier of oil to the Brits and others during World War II. It was only over the course of the 1950s and ’60s that other parts of the world caught up with US oil production, but the oil industry, in a deep sense, was born in the United States and was always priced in dollars. Saudi Aramco was originally a joint venture with an American company—or maybe even fully owned by an American company, I forget—so it was natural that it was priced in dollars. It wasn’t like in the 1970s you had to do a new deal to price oil in dollars rather than something else. Oil was in dollars.

Those dollars piled up, and it was a period of difficulty in the international monetary system. The US was going off the gold standard; Bretton Woods was breaking down; high inflation was not well contained after the first oil shock. There was an effort to convince the Saudis to keep their large stock of new petrodollars in dollars—not buy euros—and to use them at least in part to buy Treasuries. Even then, the Saudis were a little reluctant to visibly buy Treasuries. Some Bloomberg reporters several years ago went through this history, and the US started masking who was buying Treasuries at the request of the Saudis. The Saudis essentially said, “You guys are supporting Israel, we don’t really want to be seen buying your bonds directly, can you hide it?” And we agreed. Because there was still residual tension between the US and many parts of the Arab world, a lot of the dollars did not flow into the Treasury market. They flowed into bank accounts in London—offshored effectively eurodollars originating from petro-states. Those got recycled and lent in no small part to oil-importing emerging economies, and that is viewed as the start of the buildup of vulnerabilities that led to the Latin American debt crisis in the 1980s.

There’s another part of this whole story that I think people forget, which is sort of irritating me lately. After 1979–80, the Saudis had built up huge stocks of dollars—a great decade for the Saudis in the 1970s. In the ’80s, in order to keep prices high they had to cut production, and eventually that wasn’t enough and the oil price collapsed. By the end of the 1980s, and certainly by the middle of the 1990s, all the dollars that had been built up in the 1970s had been spent. The Saudi cumulative current account balance went back to basically being neutral or in deficit by ’95 and certainly by 2000. So in some sense the petrodollar boom came and it went. By the time of the Asian financial crisis, oil prices were very low—in the $20s—and there were no flows of petrodollars nor a very large stock of petrodollars. There’s sometimes a tendency to think the ’70s just continued and continued, but the reality is that, setting aside the really rich Emirates and Kuwait, the rest of the oil exporters were not in a position to continuously build up and save over most of the period after 1980 until the big run-up in oil from 2003 to 2014…

…Brad Setser: The last point, and this is just to be provocative because I’m tired of people blabbering about the dollar as the global reserve currency and how that’s the foundation of everything: an international large-cap equity portfolio will have a US share of roughly two-thirds—65 to 70%. The Saudi Public Investment Fund—my friend Alex Etra has done some work on it—its international portfolio has a dollar share of 80%, and that’s probably typical, because most private equity funds are going to be pretty dollar-heavy. A typical global reserve portfolio is now at 57% dollars. So the notion that reserves are the source of inflows into dollars is a bit dated. A reserve portfolio will typically have a lower dollar share than a standard return-seeking equities fund, which just because of the outperformance of US large caps will be more overweight dollars…

…Brad Setser:  a quarter of global reserves, to the first approximation, are in China. China still manages its currency against the dollar, but China as a matter of policy brought its formal disclosed dollar reserve share down to 55%, from 79% in 2005. They did not like the optics of financing their strategic rival and holding a lot of Treasuries in visible ways. That’s a bit misleading, because the dollar share of the portfolios of the state banks—which now have a very large share of the total state portfolio—is much higher, around 70%. If you actually net out the offshore liabilities of the state banks and just look at the net, the euro offshore portfolio is matched by euro offshore liabilities. The dollar offshore portfolio is matched by dollars onshore. In a sense, the BOP flow through the state banks was, setting aside some of the CNY lending which has gone up, almost 100% dollars…

…Brad Setser: Now, we are in a world where an enormous share of the world’s financial wealth—both people looking for safety in reserve assets, people looking for a bit more yield than you can get out of a safe G10 government bond, the private credit/CLO world, and people wanting the equity home runs—all those investors globally are now quite overweight US assets. As a result, the dollar is quite strong. To me, the core question is not really whether geopolitics will change things, assuming we don’t get into a full-on blow-up with Europe, which would accelerate some shifts. The real question is: is this intense overweight in the dollar sustainable when we have fairly reckless policies? The answer so far has been yes.

3. Token Cost Conundrums – Abdullah Al-Rezwan

Each model has its own tokenizer that decides how many tokens your prompt becomes. Feed the exact same prompt to GPT-5.4 and Claude Opus 4.7, and Claude might slice it into 2–3x as many pieces. So even if the headline price were exactly the same, you’d pay 2–3x more for identical content…

…”We sent identical inputs through each provider’s official token counting API and normalized against OpenAI’s…

…”The differences are dramatic. On tool-heavy workloads, claude-opus-4-7 costs 5.3x more than gpt-5.4 even though their list prices are only 2x apart. The rankings also flip depending on what you’re sending: Gemini is the cheapest option for text and structured data, but becomes 46% more expensive than OpenAI on tool definitions.

The only way to know what you’re actually paying is to measure it.”…

…Similarly, after understanding these nuances, I think any enterprise would be really imprudent to standardize on just one model developer. This is because the customer loses bargaining power, a benchmark, and the ability to distinguish real quality differences from billing artifacts. If the seller controls both the meter and the service, and the buyer has no parallel benchmark, the buyer is highly likely to end up paying more over the long term. Even if the model developer isn’t sneakily charging you higher price, without any benchmark, how will the customer press the model developer to lower their price or even understand that they’re paying too high a price?…

…Nonetheless, the smart move does seem to be multi-model capability (even if 95% of volume goes to one vendor) plus internal benchmarks run on your actual prompts. That gives you the optionality to switch and more importantly, the negotiating leverage to push back at contract renewal. Given this context, I believe it will be exceptionally unlikely that enterprise AI will ever be dominated by one model developer. Anthropic may be dominating enterprise AI today, but OpenAI and Google will also likely have plenty of opportunities to gain further ground.

4. Elite law firm Sullivan & Cromwell admits to AI ‘hallucinations’ – Sujeet Indap and Kaye Wiggins

Sullivan & Cromwell told a US federal bankruptcy court that a major filing it made in a high-profile case contained multiple “hallucinations” made by AI software…

…The case in question revolves around S&C’s representation of liquidators appointed by legal authorities in the British Virgin Islands who are pursuing actions against Prince Group and its owner Chen Zhi.

US federal prosecutors last year charged Zhi with wire fraud and money laundering, accusing him of “directing Prince Group’s operation of forced-labour scam compounds across Cambodia . . . that stole billions of dollars from victims in the United States and around the world”.

In a separate action, US prosecutors also filed a civil forfeiture complaint seeking to seize nearly $9bn worth of bitcoin that the US authorities said represented the proceeds of the Prince Group crimes. Zhi was arrested earlier this year in Cambodia and extradited to China after a request from Beijing.

Prince Group is incorporated in the British Virgin Islands and the Chapter 15 proceeding in the US court system is designed to get the US government to formally recognise the powers of the BVI liquidators to represent creditors and victims in the US legal proceedings, liquidators told the court.

In multiple instances, S&C in the April 9 filing erroneously summarised the conclusions made in other cases, according to a list of strike-through corrections the firm submitted to the judge.

S&C has an enterprise licence for ChatGPT according to multiple people familiar with the firm’s operations. According to S&C’s website, at least five high-level partners have been assigned to the Prince Group bankruptcy case.

5. Anthropic’s Mythos Model Is Being Accessed by Unauthorized Users – Rachel Metz

A handful of users in a private online forum gained access to Mythos on the same day that Anthropic first announced a plan to release the model to a limited number of companies for testing purposes, said the person, who asked not to be named for fear of reprisal. The group has been using Mythos regularly since then, though not for cybersecurity purposes, said the person, who corroborated the account with screenshots and a live demonstration of the model.

Anthropic has said Mythos is capable of identifying and exploiting vulnerabilities “in every major operating system and every major web browser when directed by a user to do so.” As a result, the company has taken pains to ensure that the technology is only available to a select batch of software providers through an initiative called Project Glasswing, with the goal of allowing those firms to test and safeguard their own systems from potential cyberattacks…

…The users relied on a mix of tactics to get into Mythos. These included using access the person had as a worker at a third-party contractor for Anthropic and trying commonly used internet sleuthing tools often employed by cybersecurity researchers, the person said. The users are part of a private Discord channel that focuses on hunting for information about unreleased models, including by using bots to scour for details that Anthropic and others have posted on unsecured websites such as GitHub…

…The group is interested in playing around with new models, not wreaking havoc with them, the person said. The group has not run cybersecurity-related prompts on the Mythos model, the person said, preferring instead to try tasks like building simple websites in an attempt to avoid detection by Anthropic.

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.

Some Signs of AI Froth

Companies are seeing their stock prices surge mani-fold just by adding “AI” to their name.

The late French writer Jean-Baptiste Alphonse Karr has a phrase, “Plus ça change, plus c’est la même chose” which translates into “the more things change, the more they stay the same.” This aptly describes the financial markets.

During the heady days of the Dotcom Bubble in the late 1990s, companies saw their stock prices surge simply by changing their name to include a reference to the internet. In a 2002 academic finance paper, A Rose.com by Any Other Name, Michael Cooper, Orlin Dimitrov, and Raghavendra Rau wrote (emphasis mine):

“We document a striking positive stock price reaction to the announcement of corporate name changes to Internet-related dotcom names. This “dotcom” effect produces cumulative abnormal returns on the order of 74 percent for the 10 days surrounding the announcement day.”

There have been recent rhymes in the stock market, but of the AI (artificial intelligence) variety.

On 15 April 2026, Allbirds announced a financing agreement along with changes in its business direction (laughably, from consumer footwear to providing compute for AI) and name (to NewBird AI). In response, its stock price surged 582% to US$17 on the day of the changes. NewBird AI’s stock price has since retreated to US$8, but it is still significantly higher than the pre-name-change price of less than US$3.

Later in the same day saw Myseum add “AI” to its name to highlight “the Company’s core technology platform that will integrate proprietary privacy-first artificial intelligence (AI) into its secure messaging and social media platforms.” The company’s stock price jumped by 129% the next day to close at US$3.30; at the day’s peak of US$5.77, Myseum AI’s stock price was up by 300%. The stock price is currently hovering near US$3.

The acclaimed investor Howard Marks has a great investing quote: “We may never know where we’re going, but we’d better have a good idea where we are.” And where we are right now, from what I see, is a bubbly place in AI-land.

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 don’t have a vested interest in any company mentioned. Holdings are subject to change at any time.

What We’re Reading (Week Ending 19 April 2026)

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 19 April 2026:

1. A Bakery, a Fortress, and Three Fired Central Bankers – Thomas Chua

Between 1991 and 1995, Croatia fought for independence as Yugoslavia dissolved. At its core, it was a war between a Croatian state seeking independence and Serbia wanting all territories where Serbs lived to be under Serbian control. Serbs were roughly 12% of Croatia’s population, but backed by the Yugoslav army, they pushed for roughly one third of the land.

An estimated 250,000 to 300,000 Croats were expelled from their homes, their houses looted or destroyed…

…But of all the stories I heard across Croatia, the most impactful came from our guide in Trogir.

Her grandmother believed one of her sons (the tour guide’s uncle) had been killed in the war. Heartbroken, this woman, living in a rural village, took her entire life savings and set out to find her son’s body so she could bring him home for a proper burial.

She couldn’t find him.

Eventually, she walked into a bakery and asked if anyone had seen her son’s body. They said no. She placed all her life savings on the table and told them: this is yours if you can find my son’s body. Please let me know.

The people at the bakery refused the money and said they would help, but not for the money. The grandmother left it on the table regardless.

Months later, her son came home. Alive. With her life savings in his hand. The bakery had found him and passed the money back.

In the middle of a war where Croats and Serbs were killing each other, where homes were being bombed and families torn apart, the people at that bakery who helped this grieving mother find her son were Serbs.

Not everyone supports the war. There can still be kindness across enemy lines…

…The tour guides all shared something similar. The pain never fully goes away, even if their rational minds tell them to let bygones be bygones. But they all said the same thing about the next generation: the children don’t carry the same weight. And that gives them hope that pain from the war will heal…

…I sat down at a casual spot and ordered a kebab. Nothing fancy. The bill came to 300 lira. I checked the Google reviews for the same place, and photos from a few years back showed kebab prices around 25 to 35 lira. That’s not a typo. Prices here change so fast that some of the menus had white stickers plastered over the old prices, one layer on top of another. Some restaurants had just given up on the lira entirely and started quoting in euros instead.

Our tour guide shared how prices had spiralled out of control over the past few years, and how the government is almost certainly underreporting the real inflation rate. The official numbers are bad enough.

Turkey’s official annual inflation rate was around 20% in 2021. By October 2022, it had hit 85%. It’s come down since, to around 31% as of March 2026, but independent analysts believe the real numbers are significantly higher.

Meanwhile, the Turkish lira went from about 8 per US dollar in early 2021 to around 44 per dollar today. That’s over 80% of its value gone in five years…

…How did this happen? President Erdogan holds an unconventional economic belief: that high interest rates cause inflation, not the other way around. This is the opposite of mainstream economics, where central banks raise rates to cool an overheating economy. Erdogan has called himself an “enemy of interest rates” and has also cited Islamic beliefs against usury as part of his reasoning…

…Between 2019 and 2021, Erdogan fired three central bank governors in roughly two years. The most dramatic was in March 2021, when he sacked Governor Naci Agbal just two days after the bank hiked interest rates to 19% to curb inflation. Agbal had been on the job less than five months and had been winning investor confidence. His replacement did exactly what Erdogan wanted: slashed rates from 19% down to 14%. The lira lost 44% of its value in 2021 alone.

And they kept cutting. By late 2022, the central bank had pushed rates down to 9%, even as inflation was running above 80%. The lira went into freefall. Ordinary Turks watched their purchasing power evaporate.

After winning re-election in 2023, Erdogan quietly reversed course. A new economic team was brought in and interest rates were hiked aggressively, eventually reaching 50% by March 2024. It was an implicit admission that the previous policy had failed, though Erdogan has never said so publicly.

The lesson is straightforward: when the central bank loses its independence, the consequences are severe and they fall hardest on ordinary people. A president who fires central bankers for doing their job, who replaces them with loyalists willing to cut rates into the teeth of 80% inflation, isn’t just making a policy error. He’s destroying the institutional credibility that takes decades to build and years to repair.

2. The coming El Niño of 2026 – Michael Fritzell

But first, let me explain what El Niño is. It’s essentially a climate pattern that drives global temperatures to rise, leading to droughts across Asia and Africa.

In normal years, winds blow from the eastern Pacific Ocean near South America to the western Pacific Ocean near Asia. These winds push warm water towards Asia. In normal years, this warm water causes clouds to form and rain to fall in Asia.

And since the warm water moves away from South America, the remaining water close to South America tends to be cool.

The so-called El Niño weather cycle disrupts this pattern. Instead of winds moving west, the warm water stays in the middle of the Pacific, or even moves east.

This causes:

  • Less rainfall in Asia, leading to droughts in Australia, Southeast Asia and even parts of Africa
  • More rainfall in the Southern United States and South America, leading to flooding in those regions…
  • …The US National Oceanic and Atmospheric Administration gives a 61% chance of El Niño emerging by July 2026.
  • Roughly half of the team at the European Centre for Medium-Range Weather Forecasts expect temperatures in the main El Niño region in the Pacific Ocean to exceed 2.5 degrees Celsius above the seasonal average by October 2026. Making it one of the most intense El Niños of the past century..

…First, droughts will negatively impact palm oil yields for Malaysian and Indonesian plantation companies, perhaps by as much as 10-20%. That’s how much output was impacted by the unusually strong El Niño of 1997…

…Droughts in Asia tend to reduce hydroelectric output, boosting the demand for coal in India and Indonesia. So coal prices could be heading higher, all else equal. And Indonesian coal miners stand to benefit…

…There have been a few instances, such as 2017, when key weather agencies forecasted an El Niño, yet none materialised.

However, I think there’s an asymmetry here, given that investors are not yet prepared for the potential of a super-El Niño, which could rival the one we saw in 1997.

3. China shock 2.0: the flood of high-tech goods that will change the world – Ryan McMorrow, Sam Fleming, Peter Foster, and Joe Leahy

Twenty years ago the global economy was shaken by a first “China shock” as a wave of low-cost goods destroyed the business models of manufacturers in advanced economies, displacing millions of workers and feeding discontent that fuelled populist politicians including US President Donald Trump.

Now a second shock is under way — one that is even more threatening to China’s trading partners: an assault on high-end manufacturing.

Vicious domestic competition, coupled with vast industrial scale, ample pools of engineering talent and some of the highest subsidies in the world, has generated world-beating Chinese champions in EVs, solar panels, batteries, wind turbines and a lengthening list of advanced manufacturing sectors…

…After racking up a record trade surplus in goods that surpassed $1tn in 2025, China boosted exports by nearly 15 per cent year on year in the first three months of 2026…

…BYD, the world’s largest EV maker, saw its average selling price per car fall from Rmb143,100 in 2021 to Rmb119,223 last year. Nio, one of China’s premium EV brands, has lowered the price of its flagship ES8 SUV by about 20 per cent since its 2018 debut, despite packing much more technology into the car.

Chief executive William Li says cutting costs has been a focus as they have redesigned the car. “For the first-generation ES8, the vehicle structure used 97.4 per cent aluminium, which was very expensive,” he says. “Today, we can achieve the same strength with less aluminium.”

Li adds that the group has brought the manufacture of components such as semiconductors in-house and localised the sourcing of parts such as the air suspension, which was once imported from Germany…

…“There is an ideological hardwiring at the top of the Chinese hierarchy to favour production over consumption,” says Daleep Singh, a former White House adviser under Joe Biden who is now chief global economist at PGIM, the asset management group.

“China will continue to rely on the rest of the world to absorb their excess production because the domestic political cost of empowering their own consumers is too high.”…

…The surge in Chinese exports in the first three months of 2026 was driven by shipments to the EU, up 21.1 per cent, and to south-east Asia, up 20.5 per cent year on year — even as exports to the US fell…

…A further, critical factor is the Chinese currency. Lower inflation relative to Chinese trading partners has led to a real exchange rate devaluation in the past three years, helping boost net exports and the current account surplus, which stood at 3.7 per cent of GDP last year.

The IMF estimates the country’s real effective exchange rate — which measures the real value of the currency against a basket of competitors — is undervalued by around 16 per cent, fuelling the competitive advantage enjoyed by Chinese exporters.

China has kept exports competitive by buying dollars and depreciating the currency, accumulating “shadow reserves” through a complex web of state-owned banks.

Then, crucially, there is Beijing’s industrial policy.

China has a ream of policies to help companies get off the ground, with local governments in particular battling with each other to offer the best subsidies, cheap land, financing and tax breaks to lure in manufacturers and seed new industries on their turf.

The competition between localities can be so great that some businesses move from one place to the next as they chase subsidies and investment. They have become known as “migratory bird enterprises”…

…The way the Chinese system works, local officials have every incentive to protect their companies.

Value added tax generates nearly 40 per cent of China’s tax revenue, and the central government splits the receipts with the localities where products are made, giving them a direct stake in keeping factories running.

Adding local production capacity also creates the growth that officials are largely judged on, and any large-scale lay-off could threaten social stability, Beijing’s overriding priority.

“Officials are scared of missing their GDP targets. Nobody is scared of overcapacity,” says another founder, who asks to remain unnamed. “As long as you’re manufacturing, there’s VAT revenue. Whether you sell [a product] or make a profit, that doesn’t really affect them.”…

…Recent OECD analysis underscores the role of subsidies. Company-level analysis of Chinese industry by the 38-member organisation estimates that Chinese businesses are subsidised at between three and nine times the rate of their rich-world counterparts.

As well as grants and tax breaks, the OECD data finds that the biggest subsidies come in the form of loans from Chinese state banks offering below-market rates to Chinese companies that undercut international competition.

While such dynamics have helped Chinese groups dominate globally, profits are vanishing. In the solar industry, overcapacity has led to vast losses, which China’s top six publicly traded solar groups indicated would cumulatively total Rmb43bn for 2025.

Yet the subsidies continue. One of those six companies, Jinko Solar, received Rmb1.3bn in subsidies in the first half of 2025 but still lost Rmb3bn in the period…

…As Chinese factories rushed into solar, production capacity skyrocketed. The country has the ability to manufacture 1,200GW of solar panels annually, roughly double the 647GW installed worldwide last year, according to the China Photovoltaic Industry Association and energy think-tank Ember.

“Why was it possible to build capacity exceeding global demand by double in such a short time?” asked Li Dongsheng, the chair of television and solar conglomerate TCL. “The key reason is the distortion of resource allocation and inappropriate local government participation,” he said in an interview with local media last month.

4. Corporate dark arts gone awry: how executive incentives can destroy shareholder value $NNBR $GME $HAIN – Andrew Walker

A comp scheme that could encourage management to destroy value to maximize their own payout.

Gamestop (GME) serves as a perfect example here. In January, they gave their CEO a huge option package: the CEO got >171m options struck at $20.66/share (the stock’s closing price). The options don’t expire for 10 years, and they only vest if the company hits certain market cap and EBITDA targets…

…You can certainly see the logic behind the award: GME’s market cap is <$10B, and their 2026 EBITDA was ~$345m. This comp package is encouraging massive market cap and EBITDA growth in order to even begin vesting.

Corporate governance ninjas can probably already see the issue with this package: it encourages any growth in market cap and EBITDA, not per share numbers. That incentive carries a host of issues. To take it to the most extreme: the CEO could easily hit all of his targets by issuing stock like a wild man in order to boost the company’s market cap. He could then take all of that cash and go on an acquisition spree in order to drive the company’s EBITDA up. It doesn’t matter whether the acquisitions create value for shareholders; if they boost EBITDA, they help from a vesting perspective…

…A comp package could actually disincentivize management from maximizing shareholder value.

Why does this one scare me? Because I’m so focused on incentives, and I’m always worried I’ll be lured into a situation where the incentives look positive but are actually insidious.

A live example will show this best: consider NNBR. In 2023, the stock was trading for just over $1/share, and they recruited a new CEO with a contract that would give him up to 2.5m shares if the stock price could hold $11/share…

…Fast forward to today, and things haven’t gone that well. The stock is back down to $1.50/share (though some early strength in the stock resulted in the $2 and $3 tranches vesting), and the company is reviewing strategic alternatives. Imagine you’re the CEO and had two choices right now: sell the whole company for $3/share, or max out the company’s credit line, head to Vegas, plop down at a roulette table, and bet it all on lucky #13.

If we ignored the fact that option #2 would result in some jail time, the CEO is actually incentivized to pursue that “lever up and risk it all” option. Why? Selling the company doesn’t help him vest more units, so he’s not super incentivized to pursue a sale (particularly because it puts him out of a job). In contrast, if he got lucky with the “lever up and risk it all” strategy all those PSUs would go in the money and he’d grab a multi-million dollar windfall…

…Someone highlighted COOK’s pay to me recently, and I’d be remiss if I didn’t mention it. COOK’s financial performance for 2025 missed all of their executive team’s performance goals, resulting in their stock declining >50% during the year and “no payments under the program to the Company’s named executive officers”…. but “the Board decided to award Jeremy Andrus, the Company’s Chief Executive Officer, and Michael Joseph (Joey) Hord, the Company’s Chief Financial Officer, discretionary cash bonuses equal to $956,250 and $270,938, respectively, due to their significant contributions to the Company in 2025 and to promote retention.” Well done guys; if I was a shareholder I know I’d be thrilled with that decision!

5. Letter to the 20-year-old investor – Chin Hui Leong

If you are closer to 20, you have an edge that no amount of money can buy. More on that later…

…I actually started investing much earlier, in 2002. Back then, there weren’t many choices. I bought the only unit trust available that tracked the US-based S&P 500…

…But when I bought my first individual stock in 2005, things changed. I actually felt more comfortable holding individual stocks than I did when holding index funds…

…Since 1928, the S&P 500 has fallen 10 per cent (or more) roughly every 1.8 years and 20 per cent every five years or so. When that happens, if you’re watching the index too closely, you’ll be upset.

You’ll start looking for reasons why it declined; my advice is don’t.

The S&P 500 is made up of 500 stocks…

…Trying to figure out why all 500 – or even 30 – stocks fell at once is too much work…

…When I held individual stocks, whenever a stock price fell, I could look at how much cash the company had. I could check whether its products were still selling. I could see whether it was generating profits and free cash flow…

…Between 2005 and 2010, the S&P 500 peaked in 2007, only to fall spectacularly during the global financial crisis. While the US market recovered starting from 2009, the index ended 2010 roughly where it started five years earlier…

…During what was rated as one of the deepest recessions in 70 years, I started noticing that certain companies were thriving.

The companies included Apple, Amazon, Booking Holdings, and Netlifx. They were among the 25 stocks I bought and held for a decade or more…

…Through it all, there were benefits I did not expect. I had a window into the future. I knew that online streaming was coming before it happened. I knew that same-day delivery was possible back in 2009…

…Amazon is up 39 times from when I bought it in 2010. Netflix has grown over 313 times. Booking Holdings is up 21 times. And Apple, which people thought had saturated its market a decade ago, is up 26 times…

…If you started investing at 20, or even earlier, time is on your side.

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, Apple, and Netflix. Holdings are subject to change at any time.