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

We’re thick in the action of the latest earnings season for the US stock market – for the first quarter of 2025 – 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:

• 2023 Q1 – here and here
• 2023 Q2 – here and here
• 2023 Q3 – here and here
• 2023 Q4 – here and here
• 2024 Q1 – here and here
• 2024 Q2 – here and here
• 2024 Q3 – here and here
• 2024 Q4 – here, here, and here

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

Airbnb (NASDAQ: ABNB)

Airbnb’s management thinks that designing end-to-end travel is very difficult and travelers often find planning travel to be very complicated, so travelers do it very infrequently; management thinks that a great user interface is the key to designing a great end-to-end travel experience for Airbnb users, and AI will be an important way to do it

I think a lot of companies have tried to like design an end-to-end travel. I think designing end-to-end travel is very, very hard. It’s funny — there’s this funny thing. One of the most common start-up ideas for entrepreneurs is to do a travel planning app. And yet travel planning apps almost always fail. So it’s almost like a riddle, why do travel planning apps fail, and everyone really tries to do it? And the reason why is because to plan travel is very complicated. In fact, it’s so complicated many people have assistants and a big part of their job is to plan travel for them. And yet you use it infrequently. So it’s a very difficult thing to do and you do it infrequently. And so therefore, a lot of companies have failed to design like a so-called connected trip. So I think to do this, a lot of it is to design a really good user experience. And I think that’s one of the things that we’re going to try to do to really design a great end-to-end experience, to able to book your entire trip, and much more. I think the user interface will be important. I think AI will be an important way to do this as well…

…We’re focused on making everything instant book and easy to use. We’re trying to make sure that the end-to-end travel experience is really, really wonderful with great Airbnb design, and we’re going to bring more AI into the application so that Airbnb, you can really solve your own problems with great self-solve through AI customer service agents.

Airbnb’s management recently rolled out an AI customer service agent; 50% of Airbnb’s US users are already using the customer service agent and it will soon be rolled out to 100% of Airbnb’s US users; management thinks Airbnb’s AI customer service agent is the best of its kind in travel, having already led to a 15% reduction in users needing to contact human agents; the AI customer service agent will be more personalised and agentic in the years ahead

We just rolled out our AI customer service agent this past month. 50% U.S. users are now using the agent, and we’ll roll it out to 100% of U.S. users this month. We believe this is the best AI-supported customers travel agent in travel. It’s already led to a 15% reduction in people needing to contact live human agents and it’s going to get significantly more personalized and agentic over the years to come.

Alphabet (NASDAQ: GOOG)

AI Overviews in Search now has more than 1.5 billion monthly users; AI Mode has received early positive reaction; usage growth of AI Overviews continues to increase nearly a year after its launch; management is leaning heavily into AI Overviews; management released the AI Mode in March as an experiment; AI Mode searches are twice as long as traditional search queries; AI Mode is getting really positive feedback from early users; the volume of commercial queries on Google Search has increased with the launch of AI Overviews; AI Overviews is now available in 15 languages and 140 countries; AI Overviews continues to monetise at a similar rate to traditional Search; reminder that ads within AI Overviews was launched in mobile in the USA in late-2024; an example of longer search queries in AI Mode is product comparisons; management thinks AI Overviews in Search and Gemini as 2 distinct consumer experiences; management thinks of AI Mode as a way to discover how the most advanced users are using AI-powered search

AI Overviews is going very well with over 1.5 billion users per month, and we are excited by the early positive reaction to AI Mode…

…Nearly a year after we launched AI Overviews in the U.S., we continue to see that usage growth is increasing as people learn that Search is more useful for more of their queries. So we are leaning in heavily here, continuing to roll the feature out in new countries to more users and to more queries. Building on the positive feedback for AI Overviews, in March, we released AI Mode, an experiment in labs. It expands what AI Overviews can do with more advanced reasoning, thinking and multimodal capabilities to help with questions that need further exploration and comparisons. On average, AI Mode queries are twice as long as traditional search queries. We’re getting really positive feedback from early users about its design, fast response time and ability to understand complex, nuanced questions…

…As we’ve mentioned before, with the launch of AI Overviews, the volume of commercial queries has increased. Q1 marked our largest expansion to date for AI Overviews, both in terms of launching to new users and providing responses for more questions. The feature is now available in more than 15 languages across 140 countries. For AI Overviews, overall, we continue to see monetization at approximately the same rate, which gives us a strong base in which we can innovate even more…

…On the ads of — in AI Overviews, last — late last year, actually, we launched them within the AI Overviews on mobile in the U.S. And this builds on our previous rollout of ads above and below. So this was a change that we have…

…I mentioned people typing in longer queries. There’s a lot more complex, nuanced questions. People are following through more. People are appreciating the clean design, the fast response time and the fact that they can kind of be much more open-ended, can undertake more complicated tasks. Product comparisons, for example, has been a positive one, exploring how tos, planning a trip…

…On AI-powered search and how do we see our consumer experience. Look, I do think Search and Gemini, obviously, will be 2 distinct efforts, right? I think there are obviously some areas of overlap, but they’re also — like expose very, very different use cases. And so for example, in Gemini, we see people iteratively coding and going much deeper on a coding workflow, as an example. So I think both will be around…

…AI Mode is the tip of the tree for us pushing forward on an AI-forward experience. There will be things which we discover there, which will make sense in the context of AI Overviews, so I think will flow through to our user base. But you almost want to think of what are the most advanced 1 million people using Search for, the most advanced 10 million people and then how do 1.5 billion people use Search for.

Alphabet’s management rolled out Alphabet’s latest foundation model, Gemini 2.5, in 2025 Q1; Gemini 2.5 is widely recognised as the best model in the industry; Gemini 2.5 Pro debuted at No.1 on the Chatbot Arena in 2025 Q1 by a significant margin; activer users in AI Studio and Gemini API is up 200% since the start of 2025; Alphabet introduced Gemini 2.5 Flash in April 2025; Gemini models are now found in all of Alphabet’s 15 products with at least 0.5 billion users each; Alphabet is upgrading Google Assistant on mobile devices to Gemini, and will also upgrade tablets, cars, and devices that connect to phones later this year; the Pixel 9a phone with Gemini integration was launched to strong reviews; the Gemini Live camera feature, among others, will soon be rolled out to all Android devices

This quarter was super exciting as we rolled out Gemini 2.5, our most intelligent AI model, which is achieving breakthroughs in performance, and it’s widely recognized as the best model in the industry…

…We released Gemini 2.5 Pro last month, receiving extremely positive feedback from both developers and consumers. 2.5 Pro is state-of-the-art on a wide range of benchmarks and debuted at #1 on the Chatbot Arena by a significant margin. 2.5 Pro achieved big leaps in reasoning, coding, science and math capabilities, opening up new possibilities for developers and customers. Active users in AI Studio and Gemini API have grown over 200% since the beginning of the year…

…Last week, we introduced 2.5 Flash, which enables developers to optimize quality and cost…

…All 15 of our products with 0.5 billion users now use Gemini models…

…We are upgrading Google Assistant on mobile devices to Gemini. And later this year, we’ll upgrade tablets, cars and devices that connect to your phones such as headphones and watches. The Pixel 9a launched very strong reviews, providing the best of Google’s AI offerings like Gemini Live and AI-powered camera features. And Gemini Live camera and screen sharing is now rolling out to all Android devices, including Pixel and Samsung S25.

Google Cloud is offering the industry’s widest range of TPUs and GPUs; Alphabet’s 7th generation TPU, Ironwood, has 10x better compute power and 2x better power efficiency than the previous generation TPU; Google Cloud is the first cloud provider to offer NVIDIA’s Blackwell family of GPUs; Google Cloud will be offering NVIDIA’s upcoming Rubin family of GPUs

Complementing this, we offer the industry’s widest range of TPUs and GPUs and continue to invest in next-generation capabilities. Ironwood, our seventh-generation TPU and most powerful to date, is the first designed specifically for inference at scale. It delivers more than 10x improvement in compute power or a recent high-performance TPU while being nearly twice as power efficient. Our strong relationship with NVIDIA continues to be a key advantage for us and our customers. We were the first cloud provider to offer NVIDIA’s groundbreaking B200 and GB200 Blackwell GPUs, and we’ll be offering their next-generation Vera Rubin GPUs.

Alphabet’s management is rolling out the company’s latest image and video generation models; Alphabet has launched its open-sourced Gemma 3 model in March 2025; Gemma models have been downloaded more than 140 million times; Alphabet is developing robotics AI models; Alphabet has launched a multi-agent AI research system called AI Co-Scientist; the AlphaFold model has been used by more than 2.5 million researchers

Our latest image and video generation models, Imagen 3 and Veo 2, are rolling out broadly and are powering incredible creativity. Turning to open models. We launched Gemma 3 last month, delivering state-of-the-art performance for its size. Gemma models have been downloaded more than 140 million times. Lastly, we are developing AI models in new areas where there’s enormous opportunity, for example, our new Gemini Robotics models. And in health, we launched AI Co-Scientist, a multi-agent AI research system, while AlphaFold has now been used by over 2.5 million researchers.

Google Cloud’s AI developer platform, Vertex AI, now has more than 200 foundation models available, including Alphabet’s in-house models and third-party models

Our Vertex AI platform makes over 200 foundation models available, helping customers like Lowe’s integrate AI. We offer industry-leading models, including Gemini 2.5 Pro, 2.5 Flash, Imagen 3, Veo 2, Chirp and Lyria, plus open-source and third-party models like Llama 4 and Anthropic.

Google Cloud is the leading cloud platform for building AI agents; Google Cloud has an open source framework for building AI agents and multi-agent systems called Agent Development Kit; Google Cloud has a low-code agent-building tool called Agent Designer; KPMG is using Google Cloud to deploy AI agents to employees; Google Cloud has the Google Agentspace product that helps employees in organisations use AI agents widely; Google Cloud offers pre-packaged AI agents across various functions including coding and customer engagement; Alphabet is working on agentic experiences internally and deploying it across the company; Alphabet’s customer service teams have deployed AI agents to dramatically enhance the user experience and is teaching Google Cloud customers how to do so

We are the leading cloud solution for companies looking to the new era of AI agents, a big opportunity. Our Agent Development Kit is a new open-source framework to simplify the process of building sophisticated AI agents and multi-agent systems. And Agent Designer is a low-code tool to build AI agents and automate tasks in over 100 enterprise applications and systems.

We are putting AI agents in the hands of employees at major global companies like KPMG. With Google Agentspace, employees can find and synthesize information from within their organization, converse with AI agents and take action with their enterprise applications. It combines enterprise search, conversational AI or chat and access to Gemini and third-party agents. We also offer pre-packaged agents across customer engagement, coding, creativity and more that are helping to provide conversational customer experiences, accelerate software development, and improve decision-making…

…Particularly with the newer models, I think we are working on early agentic workflows and how we can get those coding experiences to be much deeper. We are deploying it across all parts of the company. Our customer service teams are deeply leading the way there. We’ve both dramatically enhanced our user experience as well as made it much more efficient to do so. And we are actually bringing all our learnings and expertise in our solutions through cloud to our other customers. But beyond that, all the way from the finance team preparing for this earnings call to everything, it’s deeply embedded in everything we do.

Waymo is now serving 250,000 trips per week (was 150,000 in 2024 Q4), up 5x from a year ago; Waymo launched its paid service in Silicon Valley in 2025 Q1; Waymo has expanded in Austin, Texas, and will launch in Atlanta later this year; Waymo will launch in Washington DC and Miami in 2026; Waymo continues to make progress in airport access and freeway driving; management thinks Alphabet will not be able to scale Waymo by themselves, so partners are needed

Waymo is now safely serving over 0.25 million paid passenger trips each week. That’s up 5x from a year ago. This past quarter, Waymo opened up paid service in Silicon Valley. Through our partnership with Uber, we expanded in Austin and are preparing for our public launch in Atlanta later this summer. We recently announced Washington, D.C. as a future ride-hailing city, going live in 2026 alongside Miami. Waymo continues progressing on 2 important capabilities for riders, airport access and freeway driving…

More businesses are adopting Alphabet’s AI-powered campaigns; Alphabet’s recent work with AI is helping advertisers reach customers and searches where advertising would previously not be showed; Alphabet is infusing AI at every step of the marketing process for advertisers, for example, (1) advertisers can now generate a broader variety of lifestyle imagery customized to their business, (2) in PMax, advertisers can automatically source images from their landing pages and crop them, (3) on media buying, AI-powered campaigns continue to help advertisers find new customers, (4) in Demand Gen, advertisers can more precisely manage ad placements and understand which assets work best at a channel level; users of Demand Gen now see an average 26% year-on-year increase in conversions per dollar spend; when Demand Gen is paired with Product Feed, advertisers see double the conversion per dollar spend year-over-year on average; Royal Canin used Demand Gen and PMax campaigns and achieved a 2.7x higher conversion rate, a 70% lower cost per acquisition for purchases, a 8% higher value per user

More businesses, big and small, are adopting AI-powered campaigns, and the deployment of AI across our Ads business is driving results for our customers and for our business. Throughout 2024, we launched several features that leverage LLMs to enhance advertiser value, and we’re seeing this work pay off. The combination of these launches now allows us to match ads to more relevant search queries. And this helps advertisers reach customers and searches where we would not previously have shown their ads.

Focusing on our customers, we continue to solve advertisers’ pain points and find opportunities to help them create, distribute and measure more performant ads, infusing AI at every step of the marketing process. On Audience Insights, we released new asset audience recommendations, which tell businesses the themes that resonate most with their top audiences. On creatives, advertisers can now generate a broader variety of lifestyle imagery customized to their business to better engage their customers and use them across PMax, demand gen, display and app campaigns. Additionally, in PMax, advertisers can automatically source images from their landing pages and crop them, increasing the variety of their assets. On media buying, advertisers continue to see how AI-powered campaigns help them find new customers. In Demand Gen, advertisers can more precisely manage ad placements across YouTube, Gmail, Discover and Google Display Network globally and understand which assets work best at a channel level. Thanks to dozens of AI-powered improvements launched in 2024, businesses using Demand Gen now see an average 26% year-on-year increase in conversions per dollar spend for goals like purchases and leads. And when using Demand Gen with Product Feed, on average, they see more than double the conversion per dollar spend year-over-year…

…Royal Canin combined Demand Gen and PMax campaigns to find more customers for its cat and dog food products. The integration resulted in a 2.7x higher conversion rate, a 70% lower cost per acquisition for purchases and increased the value per user by 8%.

Google Cloud still has more AI demand than capacity in 2025 Q1 (as it did in 2024 Q4) 

Recall I’ve stated on the Q4 call that we exited the year in Cloud specifically with more customer demand than we had capacity. And that was the case this quarter as well.

30% of new code at Alphabet is now generated by AI (it was 25% in 2024 Q3)

We’re continuing to make a lot of progress there in terms of people using coding suggestions. I think the last time I had said, the number was like 25% of code that’s checked in. It involves people accepting AI-suggested solutions. That number is well over 30% now. But more importantly, we have deployed more deeper flows.

Amazon (NASDAQ: AMZN)

AWS grew 17% year-on-year in 2025 Q1, and is now at a US$117 billion annualised revenue run rate (was US$115 billion in 2024 Q4); management used to think AWS could be a multi-hundred billion dollar revenue run rate business without AI and now that there’s AI, they think AWS could be even bigger; AWS’s AI business is now at a multi-billion annual revenue run rate and is growing triple-digits year-on-year; the shifting from on-premise to the cloud is still a huge tailwind for AWS, and now even more so as companies that want realize the full potential of AI will need to shift to the cloud; AWS is currently still supply constrained and it will be on a lot more new chips in the coming months; management thinks that the supply chain issues with chips will get better as the year progresses

AWS grew 17% year-over-year in Q1 and now sits at a $117 billion annualized revenue run rate…

…Before this generation of AI, we thought AWS had the chance to ultimately be a multi-hundred billion dollar revenue run rate business. We now think it could be even larger…

…Our AI business has a multibillion-dollar annual revenue run rate, continues to grow triple-digit year-over-year percentages and is still in its very early days…

…Infrastructure modernization is much less sexy to talk about than AI, but fundamental to any company’s technology and invention capabilities, developer productivity, speed and cost structure. And for companies to realize the full potential of AI, they’re going to need their infrastructure and data in the cloud…

…During the first quarter, we continued to see growth in both generative AI business and non-generative AI offerings as companies turn their attention to newer initiatives, bring more workloads to the cloud, restart or accelerate existing migrations from on-premises to the cloud and tap into the power of Generative AI…

…We — as fast as we actually put the capacity in, it’s being consumed. So I think we could be driving — we could be helping more customers driving more revenue for the business if we had more capacity. We have a lot more Trainium2 instances and the next generation of NVIDIA’s instances landing in the coming months…

…I do believe that the supply chain issues and the capacity issues will continue to get better as the year proceeds.

Management is directing Amazon to invest aggressively in AI; Amazon is building 1000-plus AI applications across the company; the next generation of Alexa is Alexa+; Amazon is using AI in its fulfilment network, robotics, shopping, and more

If you believe your mission is to make customers’ lives easier and better every day, and you believe that every customer experience will be reinvented with AI, you’re going to invest very aggressively in AI, and that’s what we’re doing. You can see that in the 1,000-plus AI applications we’re building across Amazon. You can see that with our next generation of Alexa, named Alexa+. You can see that in how we’re using AI in our fulfillment network, robotics, shopping, Prime Video and advertising experiences. And you can see that in the building blocks AWS is constructing for external and internal builders to build their own AI solutions.

AWS’s in-house AI chip, Trainium 2, is starting to lay in capacity in larger quantities with significant appeal and demand; AWS will always be offering AI chips from multiple providers, but Trainium 2 offers a compelling option with 30%-40% better price performance; management believes that the price of inference needs to be much lower for AI to be successful, and they think the price of inference will go down; Anthropic is still building its next few models with Trainium 2

Our new custom AI chip Trainium2 is starting to lay in capacity in larger quantities with significant appeal and demand. While we offer customers the ability to do AI in multiple chip providers and will for as long as I can foresee, customers doing AI at any significant scale realize that it can get expensive quickly. So the 30% to 40% better price performance that Trainium2 offers versus other GPU-based instances is compelling. For AI to be as successful as we believe it can be, the price of inference needs to come down significantly…

…I would say that we’ve been bringing on a lot of P5, which is a form of NVIDIA chip instances, as well as landing more and more Trainium2 instances as fast as we can…

…Anthropic is running — building the next few training models on top of our Trainium2 chip on AWS…

…As they’re waiting to see the cost of inference continue to go down, which it will.

The latest premier Amazon Nova model was launched yesterday and it delivers frontier intelligence and industry-leading price performance; thousands of customers are already using Amazon Nova models; Amazon Nova Sonic, a speech-to-speech foundation model, was recently released and it enables developers to build voice-based AI applications; Amazon Nova Sonic has lower word error rates and higher win rates over other comparable models; AWS recently released a research preview of Amazon Nova Act, a new AI model that can perform actions within a web browser; Amazon Nova Act aims to move the current state-of-the-art accuracy of multi-step agentic actions from 30%-60% to 90%-plus

We offer our own Amazon Nova state-of-the-art foundation models in Bedrock with the latest premier model launching yesterday. They deliver frontier intelligence and industry-leading price performance, and we have thousands of customers already using them, including Slack, Siemens, Sumo Logic, Coinbase, FanDuel, Glean and Blue Origin. A few weeks ago, we released Amazon Nova Sonic, a new speech-to-speech foundation model that enables developers to build voice-based AI applications that are highly accurate, expressive and human-like. Nova Sonic has lower word error rates and higher win rates over other comparable models for speech interactions…

…We’ve just released a research preview of Amazon Nova Act, a new AI model trained to perform actions within a web browser. It enables developers to break down complex workflows into reliable atomic commands like search or checkout or answer questions about the screen. It also enables them to add more detailed instructions to these commands where needed, like don’t accept the insurance upsell. Nova Act aims to move the current state-of-the-art accuracy of multistep agentic actions from 30% to 60% to 90-plus percent with the right set of building blocks to build these action-oriented agents.

Amazon’s management sees question-and-answer being the only current use-case for AI agents, but they want AI agents to be be capable of performing a wide variety of complex tasks and they have built Alexa+ to be such an agent; management launched a new lightning fast AI agent coding experience in Amazon Q in 2025 Q1 and customers are loving it; management has made generally available GitLab Duo with Amazon Q, which enables AI agents to assist multi-step tasks; Alexa+ is meaningfully smarter and more capable than the previous Alexa; Alexa+ is free with Prime and available for non-Prime customers at $19.99 per month; Alexa+ is just starting to be rolled out in the USA and will be introduced to other countries later in 2025; users really like Alexa+ thus far; Alexa+ is now with more than 100,000 users; Amazon already has 0.5 billion devices in people’s homes and cars that can easily distribute Alexa+; management thinks users will have to relearn a little on how to communicate with Alexa+, but the communication experience is now much better; management asked Alexa+  about good Italian restaurants in New York and Alexa+ helped to make a reservation

To date, virtually all of the agentic use cases have been of the question-answer variety. Our intention is for agents to perform wide-ranging complex multistep tasks by organizing a trip or setting the lighting, temperature and music ambience in your house for dinner guests or handling complex IT tasks to increase business productivity. There haven’t been action-oriented agents like this until Alexa+…

…This past quarter, Amazon Q, the most capable generative AI-powered assistant for accelerating software development and leveraging your own data, launched a lightning fast new agent coating experience within the command line interface that can execute complex workflows autonomously. Customers are loving this. We also made generally available GitLab Duo with Amazon Q, enabling AI agents to assist multi-step tasks such as new feature development, code-based upgrades for Java 8 and 11, while also offering code review and unit testing, all within the same familiar GitLab platform…

…We introduced Alexa+, our next-generation Alexa personal assistant, who is meaningfully smarter and more capable than our prior self can both answer virtually any question and take actions and is free with Prime or available to non-Prime customers for $19.99 a month. We’re just starting to roll this out in the U.S., and we’ll be expanding to additional countries later this year. People are really liking Alexa+ this far…

…So we’ve worked hard on that in Alexa+. We’ve been — we started rolling out over the last several weeks. It’s with now over 100,000 users with more rolling out in the coming months. And so far, the response from our customers has been very, very positive…

…We’re very fortunate in that we have over 0.5 billion devices out there in people’s homes and offices and cars. So we have a lot of distribution already…

…To some degree, there will be a little bit of rewiring for people on what they can do because you get used to patterns. I mean even the simple thing of not having to speak, Alexa speak anymore, we we’re all used to saying, Alexa, before we want every action to happen. And what you find is you really don’t have to do it the first time, and then really the conversation is ongoing where you don’t have to say Alexa anymore. And I’ve been lucky enough to have the alpha and the beta that I’ve been playing with for several months, and it took me a little bit of time to realize they didn’t have to keep saying Alexa, it’s very freeing when you don’t have to do that…

…When I was in New York, when we were announcing, I asked her, what were the — we did the event way downtown. I asked her what was great Italian restaurants or pizza restaurants, she gave me a list and she asked me if she wanted me to make a reservation. I said yes. And she made the reservation and confirmed the time, like that. When you get into those types of routines and you have those types of experience, they’re very, very useful.

The majority of Amazon’s capital expenditure (capex) in 2025 Q1 was for AWS’s technology infrastructure, including the Trainium chips

Turning to our cash CapEx, which was $24.3 billion in Q1. The majority of this spend is to support the growing need for technology infrastructure. It primarily relates to AWS as we invest to support demand for our AI services and increasingly in custom silicon like Trainium as well as tech infrastructure to support our North America and International segments. We’re also investing in our fulfillment and transportation network to support future growth and improve delivery speeds and our cost structure. This investment will support growth for many years to come.

The vast majority of successful startups are built on AWS; high-profile startups building AI coding agents are on AWS

If you look at the start-up space, the vast majority of successful start-ups over the last 10 to 15 years have run on top of AWS…

…If you just look at the growth of these coding agents in the last few months, these are companies like Cursor or Vercel, both of them run significantly on AWS.

Amazon’s management thinks that current AI apps have yet to really tackle customer experiences that are going to be reinvented and many other agents that are going to be built

What’s interesting in AI is that we still haven’t gotten to all the other customer experiences that are going to be reinvented and all the other agents that are going to be built. They’re going to take the role of a lot of different functions today. And those are — they’re — even though we have a lot of combined inference in those areas, I would say we’re not even at the second strike of the first batter in the first inning. It is so early right now.

AWS operating margin improved from 37.6% in 2024 Q1 to 39.5% in 2025 Q1, but margins will fluctuate from time to time; AWS’s margin strength is from the business’s strong growth, the impact of some continued investments, and AWS’s custom chips; the investments include software optimisations for server capacity, low-cost custom networking equipment, and power usage in data centers

AWS operating income was $11.5 billion and reflects our continued growth coupled with our focus on driving efficiencies across the business. As we said before, we expect AWS operating margins to fluctuate over time, driven in part by the level of investments we’re making at any point in time…

…We had a strong quarter in AWS, as you mentioned, the margin performance. I would attribute it to the strong growth that we’re seeing, coupled with the impact of some continued investment we’re making in innovation and technology. I’ll give you some examples. So we invest in software and process improvements and ends up optimizing our server capacity, which helps our infrastructure cost. We’ve been developing more efficient network using our low-cost custom networking gear. We’re working to maximize the power usage in our existing data centers, which both lowers our costs and also reclaims power for other newer workloads. And we’re also seeing the impact of advancing custom silicon like Graviton. It provides lower cost not only for us, but also for our customers, better price performance for them.

Apple (NASDAQ: AAPL)

Apple is currently shipping an LLM (large language model) on the iPhone 16 where some of the queries are being handled on the device itself

As you know, we’re shipping an LLM on the iPhone 16 today. And there are — some of the queries that are being used by our customers are on-device, and then others go to the private cloud where we’ve essentially mimicked the security and privacy of the device into the cloud. nd then others, for world knowledge, are with the integration with ChatGPT.

The new Mac Studio has Apple’s M4 Max and M3 Ultra chips, and it can run large language models with over 600 billion parameters entirely in memory

The new Mac Studio is the most powerful Mac we’ve ever shipped, equipped with M4 Max and our new M3 Ultra chip. It’s a true AI powerhouse capable of running large language models with over 600 billion parameters entirely in memory.

Apple has released VisionOS 2.4 which unlocks the first set of Apple Intelligence features for Vision Pro users

VisionOS 2.4 unlocks the first set of Apple Intelligence features for Vision Pro users while inviting them to explore a curated and regularly updated collection of spatial experiences with the Spatial Gallery app.

Apple’s management has released iOS 18.4, which brings Apple Intelligence to more languages (including Singlish); Apple has built its own foundation models for everyday tasks; new Apple Intelligence features in iOS 18 include, Writing Tools, Genmoji, Image Playground, Image Wand, Clean Up, Visual Intelligence, and a seamless connection to ChatGPT

Turning to software. We just released iOS 18.4, which brought Apple Intelligence to more languages, including French, German, Italian, Portuguese, Spanish, Japanese, Korean and simplified Chinese as well as localized English to Singapore and India…

…At WWDC24, we announced Apple Intelligence and shared our vision for integrating generative AI across our ecosystem into the apps and features our users rely on every day. To achieve this goal, we built our own highly capable foundation models that are specialized for everyday tasks. We designed helpful features that are right where our users need them and are easy to use. And we went to great lengths to build a system that protects user privacy whether requests are processed on-device or in the cloud with Private Cloud Compute, an extraordinary step forward for privacy and AI.

Since we launched iOS 18, we’ve released a number of Apple Intelligence features from helpful Writing Tools to Genmoji, Image Playground, Image Wand, Clean Up, visual intelligence and a seamless connection to ChatGPT. We made it possible for users to create movies of their memories with a simple prompt and added AI-powered photo search, smart replies, priority notifications, summaries for mail, messages and more. We’ve also expanded these capabilities to more languages and regions.

Apple’s in-house chips are designed with a neural engine that powers AI features across Apple’s products and 3rd-party apps; management thinks the neural engine makes Apple products the best devices for generative AI

AI and machine learning are core to so many profound features we’ve rolled out over the years to help our users live a better day. It’s why we designed Apple silicon with a neural engine that powers so many AI features across our products and third-party apps. It’s also what makes Apple products the best devices for generative AI.

Apple still needs more time to work on the more personalised Siri that was unveiled by management recently

With regard to the more personal Siri features we announced, we need more time to complete our work on these features so they meet our high-quality bar. We are making progress and we look forward to getting these features into customers’ hands.

Apple has low capital expenditures for AI relative to other US technology giants because it uses 3rd-party data centers so they are mostly operating expenses; Apple’s new $500 billion investment in the USA could signal more capital expenditures and data center investments

On the data center side, we have a hybrid strategy. And so we utilize third parties in addition to the data center investments that we’re making. And as I’ve mentioned in the $500 billion, there’s a number of states that we’re expanding in. Some of those are data center investments. And so we do plan on making investments in that area

Arista Networks (NYSE: ANET)

Arista Networks’ management remains confident of reaching $750 million in back-end AI revenue in 2025 even with the uncertainty surrounding US tariffs; the 1:1 ratio between front-end and back-end AI spending for Arista Networks’ products still remains, but management thinks it’s increasingly hard to parse between front-end and back-end

Our cloud and AI momentum continues as we remain confident of our $750 million front-end AI goal in 2025…

…Just a quick clarification before we go into Q&A. Jayshree meant we were reiterating our back-end goal of $750 million, not front-end AI…

…[Question] Is that 1:1 ratio for the front-end back is still intact in your perspective?

[Answer] On the front-end ratio, yes, we’ve said it’s generally 1:1. It’s getting harder and harder to measure front end and back end. Maybe we’ll look at the full AI cluster differently next year. But I think 1:1 is still a good ratio. It varies. Some of them just build a cluster and don’t worry about the front end and others worry about it entirely holistically. So it does vary, but I think the 1:1 is still a good ratio…

…[Question] You reiterated the $750 million back-end target, but you’ve kind of had this $1.5 billion kind of AI target for 2025. And just wondering, is the capability of that more dependent on kind of the tariffs given kind of some of the front-end spend?

[Answer] Regarding tariffs, I don’t think it will have a material difference on the $750 million number or the $1.5 billion. We got the demand. So unless we have some real trouble shipping it or customers change their mind, I think we’re good with both those targets for the year.

Arista Networks is progressing well with its 4 major AI customers; 1 of the 4 customers have been in NVIDIA’s Infiniband solution for a long time, so they’ll be small for Arista Networks; 2 of the 4 are heading towards 50,000 GPU deployments by end-2025, maybe even 100,000 GPUs; 3 of the 4 customers are already in production with the 4th progressing well towards production; management has a lot of visibility from the 4 major AI customers for 2025 and 2026 and it’s looking good; the 4 major AI customers are mostly deploying Arista Networks’ 800-gig switches

We are progressing well in all 4 customers and continue to add smaller ones as well…

…Let me start with the 4 customers. All of them are progressing well. One of them is still new to us. They’ve been in Infiniband for a long time, so they’ll be small. I would say 2 of them are heading towards 50,000 GPU deployments by end of the year, maybe they’ll be at 100 but I can be most certainly sure of 50,000, heading to 100,000. And then the other one is also in production. So I had talked about all 4 going into production. Three are already in production, the fourth one is well underway…

…[Question] If I can go back to the 4 Tier 1s that you’re working with on the AI back end and the progress that you updated on that front. Are these customers now giving you more visibility just given the tariff landscape and that you would need to sort of build inventory for some of the finished codes? And can you just update us how they’re handling the situation on that front? And particularly then, as you think about — I think the investor focus is a lot about sort of 2026 and potential sort of changes in the CapEx landscape from these customers at that point. Are you getting any forward visibility from them? Any sort of early signs for 2026 on these customers?

[Answer] We definitely have all the visibility in the world for this year, and we’re feeling good. We’re getting unofficial visibility because they all know our lead times are tied to some fairly long lead times from our partners and suppliers. So I would say 2026 is looking good. And based on our execution of 2025 and plans we’re putting together, we should have a great year in 2026 as well for AI sector specifically…

…[Question] Do you see the general cadence of hyperscalers deploying 800-gig switch ports this year? I ask because I believe your Etherlink family of switches became generally available in late 2024.

[Answer] I alluded to this earlier in 2024, the majority of our AI trials were on 400 gig at that time. So you’re right to observe that with our Etherlink portfolio really getting introduced in the second half of ’24 that a lot of our 800-gig activity has picked up in 2025, some of which will be reflected in shipments and some of it which will be part of our deferred. So it’s a good observation and an accurate one that this is the year of 800, like last year it was the year of 400.

Arista Networks’ management plans for the company to be the premier and preferred network for NVIDIA’s next-generation GPUs; Arista Networks’ Etherlink portfolio makes it easy to identify and localise performance issues in accelerated compute AI clusters

At the GTC event in March of 2025, we heard all about NVIDIA’s planned GPU road map every 12 to 18 months, and Arista intends to be the premier and preferred scale-out network for all of those GPUs and AI accelerators. Traditional GPUs have a collective communication libraries or CCL, as they’re known, that try to discover the underlying network topology using localization techniques. With this accelerated compute approach, the discrepancies between the discovered topology and the one that actually happens can impact AI job completion times. Arista’s ethylene portfolio highlights the accelerated networking approach, bringing that single point of network control and visibility as a differentiation. This makes it extremely crisp to identify and localize performance issues especially as the size of the AI cluster grows to 50,000 and 100,000 XPUs with the Arista AI Spine and leaf network designs.

Arista Networks’ campus portfolio provides cost-effective access points for agentic AI applications

Arista’s cognitive campus portfolio features our advanced spine with power or Ethernet-wired lease capabilities, along with a wide range of cost-effective wireless or 7 indoor and outdoor access points for the newer IoT and agentic applications.

The data center ecosystem is still somewhat new to AI and the suppliers are figuring things out together

But everybody is new to AI, they’ve never really put together a network design for 4-rail or 8-rail or how does it connect into the GPUs and what is the NIC [network interface card] attachment? What is the accessories in terms of cables or optics that connect? So this movement from trials to production causes us to bring a whole ecosystem together for the first time.

Arista Networks’ management thinks that when it comes to AI use-cases, Arista Networks’ products will play a far bigger role than whitebox networking manufacturers, even though whiteboxes will always be around and management is even happy to help customers build networking solutions that encompass both Arista Networks’ products and whiteboxes; Arista Networks was able to help a small AI customer build a network for a cluster of a few hundred GPUs very quickly after the customer struggled to do so with whiteboxes

I’ve always said, that white box is not new. It’s been with us since the beginning of time. In fact, when Arista got started, a couple of our customers had already implemented internally various implementations of white box. So there is a class of customers who will make the investments in engineering and operations to build their own network and manage it. And it’s a very different business model. It operates typically at 10% gross margins. I don’t think you want Arista to go there. And it’s very hardware-centric and doesn’t require the rich software foundation and investments that we’ve made. So first, I’ll start by saying we will always and will continue to coexist with white box. There are times that you’ve noticed this, too, that because Arista builds some very superior hardware, that even if they don’t use our EOS, they like to have our blue box, as I often call it, the Arista hardware that’s engineered much better than any others with a more open OS like Sonic or FBOSS or at least the attributes of running both EOS and an open-source networking system. So I think we view this as a natural part of selection in a customer base where if it’s a simple use case, they’re going to use something cost effective. But if it’s a really complex use case, like the AI spine or roles that require and demand more mission-critical features, Arista always plays a far bigger role in premium, highly scalable, highly valued software and hardware combinations than we do in a stand-alone white box. So we’ll remain coexistent peacefully, and we’re not in any way threatened by it. In fact, I would say we work with our customers to make sure as they’re building permutations and combinations of the white box, that we can work with that and build the right complement to that with our Etherlink portfolio…

…We had a customer, again, not material. We said, “I can’t get these boxes. I can’t make them run. I cannot get an AI network.” And one of my most technical sales leaders said, hey, we got a chance to build an AI cluster here for a few hundred GPUs. We jumped on it. Obviously, that customer is small and have been largely using white boxes and is now about to install an AI leaf and an AI spine, and we had to get it to him before the tariff deadline. So as an example of not material, but how quickly these decisions get made when you have the right product, right performance, right quality, right mission-critical nature and you can deal with that traffic pattern better than anyone else can. So it happens. It’s not big because we’ve got so much commitment in a given quarter from a customer, but when it is, we ask with great deal of nimbleness and agility to do that.

Arista Networks’ management is happy to support any kind of advanced packaging technologies – such as co-packaged optics or co-packaged copper – for back-end AI networks in the company’s products; management has yet to see any major adoption of co-packaged optics for back-end AI networks

[Question] I’d love to get your latest views around co-packaged optics. NVIDIA introduced its first CPO switches, GCC, for scale-out. And I was wondering whether that had any impact on your views regarding CPO adoption in back-end AI networks in coming years.

[Answer] It’s had no impact. It’s very early days. I think you’ve seen — Arista doesn’t build optics, but Arista enables optics and we’ve always been at the forefront, especially with Andy Bechtolsheim and his team of talented tech individuals that whether it is pluggable optics with LPO or how we define the OSFP connector for MSAs or 100 gig, 400 gig, it’s something we take seriously. And our views on CPOs, it’s not a new idea. It’s been demonstrated in prototype for, I don’t know, 10 to 20 years. The fundamental lack of adoption to date on CPO, it’s relatively high failure rates and it’s mostly been in the labs. So what are some of the advantages of CPO? Well, it has a linear interface. It has lower power than DSP for long-haul optics. It has a higher channel count. And I think if pluggable optics can achieve some of that in the best of both worlds, then you can overcome that with pluggable optics or even co-packaged copper. So Arista has no religion. We will do co-package copper. We’ll do co-package optics. We will do pluggable optics, but it’s too early to call this a real production-ready product that’s still in very early experiments and trials.

Arista Networks’ management is not seeing any material pull-forward in demand for its products because of US tariffs

[Question] We know tariffs are coming later in the year. Whether the strength you’re seeing is the result of early purchases of customers ahead of tariffs in order to save some dollars?

[Answer] Even if our customers try to pull it in and get it all by July, we would be unable to supply it. So that would be the first thing. So I’m not seeing the pull-ins that are really material in any fashion. I am seeing a few customers trying to save $1 here, $1 there to try and ship it before the tariff date but nothing material. Regarding pull-ins for 4 to 6 quarters, again, our best visibility is near term. And if we saw that kind of behavior, we would see a lot of inventory sitting in our customers, which we don’t. In fact, that’s long enough to ship faster and ship more.

2 years ago, Arista Networks’ management saw all its Cloud Titan customers pivot to AI and slow down their cloud spending; management is seeing more balanced spending now, with a more surgical focus on AI

2 years ago, I was very nervous because the entire cloud titans pivoted to AI and slowed down their cloud. Now we see a more balanced spend. And while we can’t measure how much of this cloud and how much of it is AI, if they’re kind of cobbled together, we are seeing less of a pivot, more of a surgical focus on AI and then a continued upgrade of the cloud networks as well. So compared to ’23, I would say the environment is much more balanced between AI and cloud.

Arista Networks’ management sees competitive advantages in the company’s hardware design, development, and operation that are hard to replicate even for its Cloud Titan customers

[Question] What functionality about the blue box actually makes it defensible versus what hyperscalers can kind of self-develop?

[Answer] Let me give you a few attributes of what I call the blue box, and I’m not saying others don’t have it, but Arista has built this as a mission, although we’re known for our software. We’re just as well known for our hardware. When you look at everything from a form factor of a one RU that we build to a chassis, we’ve got a tremendous focus on signal integrity, for example, all the way from layer 1, multilayer PCB boards, a focus on quality, a focus on driving distances, a focus on integrating optics for longer distances, a focus on driving MACsec, et cetera. So that’s a big focus. The second is hardware diagnostics. Internal to the company, we call it Arista boot. We’ve got a dedicated team focused on not just the hardware but the firmware to make it all possible in terms of troubleshooting because when these boards get super complex, you know where the failure is and you’re running at high-speed 200 [indiscernible] 30s. So things are very complex. So the ability to pinpoint and troubleshoot is a big part of what we do. And then there’s additional focus on the mechanical, the power supplies, the cooling, all of which translate to better power characteristics. Along with our partners and chip vendors, there’s a maniacal focus on not just high performance but low power. So some of the best attributes come from our blue boxes, not only for 48 ports, but all the way up to 576 ports of an AI spine or double that if you’re looking for dual capabilities. So well-designed, high-quality hardware is a thing of beauty, but also think of complexity that not everyone can do.

With neo AI cloud customers, Arista Networks’ management is observing that they are very willing to forsake NVIDIA’s GPUs and networking solutions and try other AI accelerators and Ethernet; management thinks that the establishment of the Ultra Ethernet Consortium in 2024 has a role to play in the increasing adoption of Ethernet for AI networking; with the Cloud Titans, management is also observing that they are shifting towards Ethernet; management thinks that the shift from Infiniband to Ethernet is faster than the the shift from NVIDIA’s GPUs to other companies’ GPUs

[Question] There’s a general perception that most of them are buying NVIDIA-defined clusters and networking. So I wonder if you could comment on those trends, their interest in moving past InfiniBand? And also are there opportunities developing with some of these folks to kind of multi-source their AI connectivity to different providers?

[Answer] We’re seeing more adventurous spirit in the neo-cloud customers because they want to try alternatives. So some of them are absolutely trying other AI accelerators like Lisa and AMD and my friends there. Some of them are absolutely looking at Ethernet, not InfiniBand as a scale-out. And that momentum has really shifted in the last year with the Ultra Ethernet Consortium and the spec coming out in May. I just want to give a shout-out to that team and what we have done. So I think Ethernet is a given that there’s an awful lot of legacy of InfiniBand that will obviously sort itself out. And a new class of AI accelerators we are seeing more niche players, more internal developments from the cloud titans, all of which is mandating more Ethernet. So I think between your 2 questions, I would say the progress from InfiniBand to Ethernet is faster, the progress from the ones they know and the high-performance GPU from NVIDIA versus the others is still taking time.

ASML (NASDAQ: ASML)

ASML’s management still sees AI (artificial intelligence) as the key growth driver; ASML will hit upper range of guidance for 2025 if AI demand continues to be strong, while ASML will hit the lower range of guidance if there is uncertainty among its customers

Consistent with our view from last quarter, the growth in artificial intelligence remains the key driver for growth in our industry. If AI demand continues to be strong and customers are successful in bringing on additional capacity to support the demand, there is a potential opportunity towards the upper end of our range. On the other hand, there is still quite some uncertainty for a number of our customers that can lead to the lower end of our range. 

ASML’s management is still positive on the long-term outlook for ASML, with AI being a driver for growth

Looking longer term, the semiconductor market remains strong with artificial intelligence, creating growth in recent quarters, and we see some of the future demand for AI solidifying, which is encouraging. 

ASML’s management thinks inference will become a larger part of AI demand going forward

I think there has been a lot of emphasis in the past quarters on the training side of life. I think more and more, which I think is logical, that you also see more and more emphasis being put on the inferencing side of the equation. So I think you will see the inferencing part becoming a larger component of AI demand on a go-forward basis.

ASML’s management is unable to tell what 2027 will look like for AI demand, but the commitment to AI chips in the next 2 years is very strong

You are looking at major investment, investment has been committed, investment that a lot of company believe they have to make in order to basically enter this AI race, I think the threshold to change this behavior is pretty high. And this is why — this is what our customers are telling us. And that’s also why we mentioned that, based on those conversations, we still see ’25, ’26 as growth years. That’s largely driven by AI and by that dynamic. Now ’27 start to be a bit further away, so you’re asking us too much, I think, to be able to answer basically what AI may look like in ’27. But if you look at the next couple of year, so far, the commitment to the AI investment and, therefore, the commitment also to deliver the chips for AI has been very solid.

Coupang (NYSE: CPNG)

Coupang’s management is investing in automation (such as automated picking, packing and sorting) and machine learning to deploy inventory more precisely to improve the customer experience and reduce costs

This quarter, we saw benefits from advances in our automated picking, packing and sorting systems and machine learning utilization that deploys inventory with more precise prediction of demand. This, coupled with our focus on operational excellence, enables us to continually improve the customer experience while also lowering their cost of service.

Datadog (NASDAQ: DDOG)

Existing customer usage growth in 2025 Q1 was in line with management’s expectations; management is seeing high growth in Datadog’s AI cohort, and stable growth in the other cohorts

Overall, we saw trends for usage growth from existing customers in Q1 that were in line with our expectations. We are seeing high growth in our AI cohort as well as consistent and stable growth in the rest of the business.

Datadog’s ,anagement continues to see increase in interest in next-gen AI capabilities and analysis; 4,000 Datadog customers at the end of 2025 Q1 used 1 or more Datadog AI integrations (was 3,500 in 2024 Q4), up 100% year-on-year; companies using end-to-end data observability to manage model performance, security, and quality, has more than doubled in the past 6 months; management has observed that data observability has become a big enabler of building AI workloads; the acquisition of Metaplane helps Datadog build towards a comprehensive data observability suite; management thinks data observability will be a big opportunity for Datadog

We continue to see rising customer for next-gen AI capabilities and analysis. At the end of Q1, more than 4,000 customers used one or more Datadog AI integrations, and this number has doubled year-over-year. With end-to-end data observability, we are seeing continued growth in customers and usage as they seek to manage end-to-end model performance, security and quality. I’ll call out the fact that the number of companies using end-to-end data observability has more than doubled in the past 6 months…

…[Question] What the vision is about moving into data observability and how consequential an opportunity it could be for Datadog?

[Answer] The field is evolving into a big enabler or it can be positive enabler, if you don’t do it right, for building enterprise workloads — for AI workloads, sorry. So in other words, making sure the data is being extracted from the the right place, transformed the right way and is being fed into the right AI models on the other hand…

…We only had some building blocks for data observability. We built data streams monitoring product for streaming data that comes out of few, such as Kafka, for example. We built their job monitoring product that monitors back jobs and large transformation jobs. We have a database monitoring product that looks at the way you optimize queries and optimize base performance and cost. And by adding data quality and data pipelines, with Metaplane, we have a full suite basically that allows our customers to manage everything from getting the data from their core data storage into all of the products and AI workloads and reports they need to go populate that data. And so we think it’s a big opportunity for us.

Datadog’s management has improved Bits AI, and is using next-gen AI to help solve customer issues quickly and move towards auto remediation

We are adding to Bits AI, with capabilities for customers to take action with workflow automation and App Builder, using next GenAI to help our customers immediate issues more quickly and move towards auto remediation in the future.

Datadog has made 2 recent acquisitions; Eppo is a feature management and experimentation platform; management sees automated experimentation as an important part of modern application development because of the use of AI in coding; Metaplane is a data observability platform that works well for new enterprise AI workloads; management is seeing more AI-written code in both its customers and the company itself; management thinks that as AI writes more code, more value will come from being able to observe and understand the AI-written code in production environments, which is Datadog’s expertise; the acquisitions of Eppo and Metaplane are to position Datadog for the transition towards a world of AI-written code

We recently announced a couple of acquisitions.

First, we acquired Eppo, a next-generation feature management and experimentation platform. The Eppo platform helps increase the velocity of releases, while also lowering risk by helping customers to release and validate features in a controlled manner. Eppo augments our efforts in product analytics, helping customers improve the variance and tie feature performance to business outcomes. More broadly, we see automated experimentation as a key part of modern application development, with the rapid adoption of the agent generative code, as well as more and more of the application logic itself being implemented with nondeterministic AI models. 

Second, we also acquired Metaplane, the data observability platform built for modern data teams. Metaplane helps prevent, detect and resolve their availability and quality issues across the company’s data warehouses and data pipelines. We’ve seen for several years now that better freshness and quality were critical for applications and business analytic. And we believe that they are becoming key enablers of the creation of new enterprise AI workloads, which is why we intend to integrate the Metaplane capabilities into our end-to-end dataset offerings…

…There is definitely a big transition that is happening right now, like we see the rise of AI written code. We see it across our customers. We also see it inside of Datadog, where we’ve had very rapid adoption of this technology as well…

…The way we see it is that it means that there’s a lot less value in writing the code itself, like everybody can do it pretty quickly, can do a lot of it. You can have the machine to do a lot of it, and you complement it with a little bit of your own work. But the real difficulty is in validating that code, making sure that it’s safe, making sure it runs well, that it’s performing and that it does what it’s supposed to do for the business. Also making sure that when 15 different people are changing the code at the same time, all of these different changes come together and work the right way, and you understand the way these different pieces interact in the way. So the way we see it is this move out a lot of their value from writing the code to observing it and understanding it in production environments, which is what we do. So a lot of the investments we’re making right now, including some of the acquisitions we’ve announced, build towards that, and making sure that we’re in the right spot.

Datadog signed a 7-figure expansion deal with a leading generative AI company; the generative AI company needs to reduce tool fragmentation; the generative AI company is replacing commercial tools for APM (application performance monitoring) and log management with Datadog, and is expanding to 5 Datadog products

We signed a 7-figure expansion as an annualized contract with a leading next GenAI company. This customer needs to reduce tool fragmentation to keep on top of its hyper growth in usage and employee headcount. With this expansion, the customer will use 5 Datadog products and will replace commercial tool for APM and log management.

AI-native customers accounted for 8.5% of Datadog’s ARR in 2024 Q4 (was 6% in 2024 Q4); AI-native customers contributed 6 percentage points to Datadog’s year-on-year growth in 2025 Q1, compared to 2 percentage points in 2024 Q1; management thinks AI-native customers will continue to optimise cloud and observability usage in the future; AI-native contracts that come up for renewal are healthy; Datadog has huge customer concentration with the AI-native cohort; Datadog has more than 10 AI-native customers that are spending $1 million or more with Datadog; the strong performance of the AI-native cohort in 2025 Q1 is fairly broad-based; Datadog is helping the AI-native customers mostly with inference, and not training; when Datadog sees growth among AI-native customers, that’s growth of AI adoption because the AI-native customers’ workloads are mostly customer-facing

We saw a continued rise in contribution from AI-native customers who represented about 8.5% of Q1 ARR, up from about 6% of ARR last quarter and up from about 3.5% of ARR in the year ago quarter. AI-native customers contributed about 6 points of year-over-year revenue growth in Q1 versus about 5 points last quarter and about 2 points in the year ago quarter. We continue to believe that adoption of AI will benefit Datadog in the long term, but we remain mindful that we may see volatility in our revenue growth on the backdrop of long-term volume growth from this cohort as customers renew with us on different terms and as they may choose to optimize cloud and observability usage…

…[Question] Could you talk about what you’re seeing from some of those AI-native contracts that have already come up for renewal and just how those conversations have been trending?

[Answer] All the contracts that come up for renewal, they are healthy. The trick with the cohort is that it’s growing fast. There’s also a revenue concentration there. We now have our largest customer in the cohort, and they’re growing very fast. And on the flip side of that, we also have a larger number of large customers that are also growing. So we — I think we mentioned more than 10 customers now that are spending $1 million or more with us in that AI-native cohort and that are also growing fast…

…On the AI side, we do have, as I mentioned, one customer large and the others there, they’re contributing more of the new revenue than the others. But we see growth in the rest of the cohort as well. So again, it’s fairly typical…

…For the AI natives, actually, what we help them with mostly is not training. It’s running their applications and their inference workloads as customer-facing. Because what’s training for the AI natives tends to be largely homegrown one-off and different from — between each and every one of them. We expect that as and if most other companies and enterprises do significant training, that this will not be the case. This will not be one-off and homegrown. But right now, it is still the AI natives that do most of the training, and they still do it in a way that’s largely homegrown. So when we see growth on the AI-native cohorts, that’s growth of AI adoption because that’s growth of customer-facing workloads by and large.

Datadog’s management sees the trend of cloud migration as being steady; management sees cloud migration being partly driven by customers’ desires to adopt AI, because migrating to the cloud is a prerequisite for AI

[Question] What are the trend lines on the cloud migration side?

[Answer] It’s consistent with what we’ve seen before. It’s also consistent with what you’ve heard from the hyperscalers over the past couple of weeks. So I would say it’s steady, unremarkable. It’s not really trending up nor trending down right now. But we see the same desire from customers to move more into the cloud and to lay the groundwork so they can also add up AI, because digital transformation and cloud migrations are prerequisites for that.

Datadog’s management thinks there will be more products for Datadog to build as AI workloads shift towards inferencing; management is seeing its LLM Observability product getting increasing usage as customers move AI workloads into production; management wants to build more products across the stack, from closer to the GPU to AI agents; 

On the workloads turning more towards inference, so there’s definitely more product to build there. So we have a — so we built an LLM Observability product that is being — that is getting increasing usage from customers as they move into production. And we think there’s more that we need to build both down the stack closer to the GPUs and up the stack closer to the agents that are being built on top of these models.

Datadog’s management is already seeing returns on Datadog’s internal investments in AI in terms of employee productivity; in the long-term, there’s the possibility that Datadog may need lesser headcount because of AI

[Question] Internally, how do you think about AI from an efficiency perspective?

[Answer] For right now, I think we’re seeing the returns in productivity, whether that be salespeople getting more information or R&D. We’re essentially trying to create an environment where we’re encouraging the various departments to use it and learning from it. Long term, there might well be efficiency gains — there may be efficiency gains that can be manifested in headcount.

Mastercard (NYSE: MA)

Mastercard’s management sees contactless payments and tokenised transactions as important parts of agentic AI digital commerce; Mastercard has announced Mastercard Agent Pay, which will facilitate safe, frictionless and programmable transactions across AI platforms; Mastercard is working with important AI companies such as Microsoft and OpenAI to deliver agentic payments

Today, 73% of all in-person switched transactions are contactless and approximately 35% of all our switch transactions are tokenized. These technologies will continue to play an important role as we move forward into the next phase of digital commerce, such as Agentic AI. We announced Mastercard Agent Pay to leverage our Agentic tokens as well as franchise rules, fraud and cybersecurity solutions. Combined, these will help partners like Microsoft to facilitate safe, frictionless and programmable transactions across AI platforms. We will also work with companies like OpenAI to deliver smarter, more secure and more personalized agentic payments. The launch of Agent Pay is an important step in redefining commerce in the AI era.

Mastercard closed the Recorded Future acquisition in 2024 Q4 (Recorded Future provides AI-powered solutions for real-time visibility into potential threats related to fraud); Recorded Future just unveiled the AI-powered Malware Intelligence; Malware Intelligence enables proactive threat prevention

On the cybersecurity front, Recorded Future just unveiled malware intelligence. It’s a new capability enabling proactive threat prevention for any business using real-time AI-powered intelligence insights.

Mastercard’s management sees AI as being deeply ingrained in Mastercard’s business; Mastercard’s access to an enormous amount of data is an advantage for Mastercard in deploying AI; in 2024, a third of Mastercard’s products in its value-added services and solutions segment was powered by AI

AI is deeply ingrained in our business. We have access to an enormous amount of data, and this uniquely positions us to enhance our AI’s performance, resulting in greater accuracy and reliability. And we’re deploying AI to enable many solutions in market today. In fact, in 2024, AI enabled approximately 1 in 3 of our products within value-added services and solutions.

Meta Platforms (NASDAQ: META)

Meta’s management is focused on 5 opportunities within AI namely, improved advertising, more engaging experiences, business messaging, Meta AI and AI devices; the 5 opportunities are downstream of management’s attempt to build artificial general intelligence and leading AI models and infrastructure in an efficient manner; management thinks the ROI of Meta’s investment in AI will be good even if Meta does not succeed in all the 5 opportunities;

As we continue to increase our investments and focus more of our resources on AI, I thought it would be useful today to lay out the 5 major opportunities that we are focused on. Those are improved advertising, more engaging experiences, business messaging, Meta AI and AI devices. And these are each long-term investments that are downstream from us building general intelligence and leading AI models and infrastructure. Even with our significant investments, we don’t need to succeed in all of these areas to have a good ROI. But if we do, then I think that we will be wildly happy with the investments that we are making…

…We are focused on building full general intelligence. All of the opportunities that I’ve discussed today are downstream of delivering general intelligence and doing so efficiently.

Meta’s management’s goal with the company’s advertising business is for businesses to simply tell Meta their objectives and budget, and for Meta to do all the rest with AI; management thinks that Meta can redefine advertising into an AI agent that delivers measurable business results at scale

Our goal is to make it so that any business can basically tell us what objective they’re trying to achieve like selling something or getting a new customer and how much they’re willing to pay for each result and then we just do the rest. Businesses used to have to generate their own ad creative and define what audiences they wanted to reach, but AI has already made us better at targeting and finding the audiences that will be interested in their products than many businesses are themselves, and that keeps improving. And now AI is generating better creative options for many businesses as well. I think that this is really redefining what advertising is into an AI agent that delivers measurable business results at scale.

Meta tested a new advertising recommendation model for Reels in 2025 Q1 called Generative Ads Recommendation Model, or GEM, that has improved conversion rates by 5%; 30% more advertisers are using Meta’s AI creative tools in 2025 Q1; GEM is twice as efficient at improving ad performance for a given amount of data and compute; GEM’s better efficiency helped Meta significantly scale up the amount of compute used for model training; GEM is now being rolled out to additional surfaces across Meta’s apps; the initial test of Advantage+’s streamlined campaign creation flow for sales, app and lead campaigns is encouraging and will be rolled out globally later in 2025; Advantage+ Creative is seeing strong adoption; all eligible advertisers can now automatically adjust the aspect ratio of their existing videos and generate images; management is testing a feature that uses gen AI to place clothing on virtual models; management has seen a 46% lift in incremental conversions in the testing of the incremental attribution feature and will roll out the feature to all advertisers in the coming weeks; improvements in Meta’s advertising ranking and modeling drove conversion growth that outpaced advertising impressions growth in 2025 Q1

In just the last quarter, we are testing a new ads recommendation model for Reels, which has already increased conversion rates by 5%. We’re seeing 30% more advertisers are using AI creative tools in the last quarter as well…

…In Q1, we introduced our new Generative Ads Recommendation Model, or GEM, for ads ranking. This model uses a new architecture we developed that is twice as efficient at improving ad performance for a given amount of data and compute. This efficiency gain enabled us to significantly scale up the amount of compute we use for model training with GEM trained on thousands of GPUs, our largest cluster for ads training to date. We began testing the new model for ads recommendations on Facebook Reels earlier this year and have seen up to a 5% increase in ad conversions. We’re now rolling it out to additional surfaces across our apps…

…We’re seeing continued momentum with our Advantage+ suite of AI-powered solutions. We’ve been encouraged by the initial test of our streamlined campaign creation flow for sales, app and lead campaigns, which starts with Advantage+ turned on from the beginning for advertisers. In April, we rolled this out to more advertisers and expect to complete the global rollout later this year. We’re also seeing strong adoption of Advantage+ Creative. This week, we are broadening access of video expansion to Facebook Reels for all eligible advertisers, enabling them to automatically adjust the aspect ratio of their existing videos by generating new pixels in each frame to optimize their ads for full screen surfaces. We also rolled out image generation to all eligible advertisers. And this quarter, we plan to continue testing a new virtual try-on feature that uses gen AI to place clothing on virtual models, helping customers visualize how an item may look and fit…

…We continue to evolve our ads platform to drive results that are optimized for each business’ objectives and the way they measure value. One example of this is our incremental attribution feature, which enables advertisers to optimize for driving incremental conversions or conversions we believe would not have occurred without an ad being shown. We’re seeing strong results in testing so far with advertisers using incremental attribution in tests seeing an average 46% lift in incremental conversions compared to their business-as-usual approach. We expect to make this available to all advertisers in the coming weeks…

…Year-over-year conversion growth remains strong. And in fact, we continue to see conversions grow at a faster rate than ad impressions in Q1, so reflecting increased conversion rates. And ads ranking and modeling improvements are a big driver of overall performance gains.

Improvements in the past 6 months to Meta’s content recommendation systems have driven increases of 7% in time spent on Facebook, 6% on Instagram, and 35% on Threads; video consumption in Facebook and Instagram grew strongly in 2025 Q1 because of improvements to Meta’s content recommendation systems; management sees opportunities for further gains in improving the content recommendation systems in 2025; Meta is making progress on longer-term efforts to improve its content recommendation systems in two areas, (1) develop increasingly efficient recommendation systems by incorporating innovations from LLM model architectures, and (2) integrating LLMs into content recommendation systems to better identify what is interesting to a user; management’s testing of Llama in Threads’ recommendation systems has led to a 4% increase in time spent from launch; management is exploring how Llama can be deployed in recommendation systems for photo and video content, which management expects can improve Meta AI’s personalisation by better understanding users’ interests and preferences through their use of Meta’s apps; management launched a new feed in Instagram in the US in 2025 Q1 of content a user’s friends have left a note on or liked and the new feed is producing good results; management has launched the Blend experience that blends a user’s Reels algorithm in direct messages with friends; the increases of 7% in time spent on Facebook and 6% on Instagram seen in the last 6 months is on top of uplift in time spent on Facebook and Instagram that management had already produced in the first 9 months of 2024

In the last 6 months, improvements to our recommendation systems have led to a 7% increase in time spent on Facebook, 6% increase on Instagram and 35% on Threads…

…In the first quarter, we saw strong growth in video consumption across both Facebook and Instagram, particularly in the U.S., where video time spent grew double digits year-over-year. This growth continues to be driven primarily by ongoing enhancements to our recommendation systems, and we see opportunities to deliver further gains this year.

We’re also progressing on longer-term efforts to develop innovative new approaches to recommendations. A big focus of this work will be on developing increasingly efficient recommendation systems so that we can continue scaling up the complexity and compute used to train our models while avoiding diminishing returns. There are promising techniques we’re working on that will incorporate the innovations from LLM model architectures to achieve this. Another area that is showing early promise is integrating LLM technology into our content recommendation systems. For example, we’re finding that LLM’s ability to understand a piece of content more deeply than traditional recommendation systems can help better identify what is interesting to someone about a piece of content, leading to better recommendations.

We began testing using Llama in Threads recommendation systems at the end of last year given the app’s text-based content and have already seen a 4% lift in time spent from the first launch. It remains early here, but a big focus this year will be on exploring how we can deploy this for other content types, including photos and videos. We also expect this to be complementary to Meta AI as it can provide more relevant responses to people’s queries by better understanding their interests and preferences through their interactions across Facebook, Instagram and Threads…

…In Q1, we launched a new experience on Instagram in the U.S. that consists of a feed of content your friends have left a note on or liked, and we’re seeing good results. We also just launched Blend, which is an opt-in experience in direct messages that enables you to blend your Reels algorithm with your friends to spark conversations over each other’s interest…

…We shared on the Q3 2024 call that improvements to our AI-driven feed and video recommendations drove a roughly 8% lift in time spent on Facebook and a 6% lift on Instagram over the first 9 months of last year. Since then, we’ve been able to deliver similar gains in just 6 months’ time with improvements to our AI recommendations delivering 7% and 6% time spent gains on Facebook and Instagram, respectively.

AI is enabling the creation of better content on Meta’s apps; the better content includes AI generating content directly for users and AI helping users produce better content; management thinks that the content created on Meta’s apps will be increasingly interactive over time; management recently launched the stand-alone Edits app that contains an ultra-high resolution, short-form video camera, and generative AI tools to remove backgrounds of video or animate still images; more features on Edits are coming soon; 

AI is also enabling the creation of better content as well. Some of this will be helping people produce better content to share themselves. Some of this will be AI generating content directly for people that is personalized for them. Some of this will be in existing formats like photos and videos, and some of it will be increasingly interactive…

…Our feeds started mostly with text and then became mostly photos when we all got mobile phones with cameras and then became mostly video when mobile networks became fast enough to handle that well. We are now in the video era, but I don’t think that this is the end of the line. In the near future, I think that we’re going to have content in our feeds that you can interact with and that it will interact back with you rather than you just watching it…

…Last week, we launched our stand-alone Edits app, which supports the full creative process for video creators from inspiration and creation to performance insights. Edits has an ultra-high resolution, short-form video camera and includes generative AI tools that enable people to remove the background of any video or animate still images with more features coming soon.

Countries like Thailand and Vietnam with low-cost labour actually conduct a lot of business through Meta’s messaging apps but management thinks this phenomena is absent in developed economies because of the high cost of labour; management thinks that AI will allow businesses in developed economies to conduct business through Meta’s messaging apps; management thinks that every business in the future will have AI business agents that are easy to set up and can perform customer support and sales; Meta is currently testing AI business agents with small businesses in the USA and a few countries across Meta’s apps; management has launched a new agent management experience to make it easier for businesses to train their AI; management’s vision is for that to be one agent that’s interacting with a consumer regardless of where he/she is engaging with the business AI; feedback from the tests are that the AI business agents are saving businesses a lot of time and helping them determine which conversations to spend more time on

In countries like Thailand and Vietnam, where there is a low cost of labor, we see many businesses conduct commerce through our messaging apps. There’s actually so much business through messaging that those countries are both in our top 10 or 11 by revenue even though they’re ranked in the 30s in global GDP. This phenomenon hasn’t yet spread to developed countries because the cost of labor is too high to make this a profitable model before AI, but AI should solve this. So in the next few years, I expect that just like every business today has an e-mail address, social media account and website, they’ll also have an AI business agent that can do customer support and sales. And they should be able to set that up very easily given all the context that they’ve already put into our business platforms…

…We are currently testing business AIs with a limited set of businesses in the U.S. and a few additional countries on WhatsApp, Messenger and on ads on Facebook and Instagram. We’ve been starting with small business and focusing first on helping them sell their goods and services with business AIs…

…We’ve launched a new agent management experience and dashboard that makes it easier for businesses to train their AI based on existing information on their website or WhatsApp profile or their Instagram and Facebook pages. And we’re starting with the ability for businesses to activate AI in their chats with customers. We are also testing business AIs on Facebook and Instagram ads that you can ask about product and return policies or assist you in making a purchase within our in-app browser…

…No matter where you engage with the business AI, it should be one agent that recalls your history and your preferences. And we’re hearing encouraging feedback, particularly that adopting these AIs are saving the business that we’re testing with a lot of time and helping to determine which conversations make sense for them to spend more time on.

Meta AI now has nearly 1 billion monthly actives; management’s focus for Meta AI in 2025 is to establish Meta AI as the leading personal AI for personalization, voice conversations, and entertainment; management thinks people will eventually have an AI to talk to throughout the day on smart-glasses and this AI will be one of the most important and valuable services that has ever been created; management recently released the first Meta AI stand-alone app; the Meta AI stand-alone app is personalised to the user’s behaviour on other Meta apps, and it also has a social feed for discovery on how others are using Meta AI; initial feedback on the Meta AI stand-alone app is good; management expects to focus on scaling and deepening engagement on Meta AI for at least the next year before attempting to monetise; management saw engagement on Meta AI improve when testing Meta AI’s ability to personalize responses by remembering people’s prior queries and their usage of Meta’s apps; management has built personalisation into Meta AI across all of Meta’s apps; the top use cases for Meta AI currently include information gathering, writing assistance, interacting with visual content, and seeking help; WhatsApp has the strongest usage of Meta AI, followed by Facebook; a standalone Meta AI app is important for Meta AI to become the leading personal AI assistant because WhatsApp is currently not the primary messaging app used in the USA; management thinks that people are going to use different AI agents for different things; management thinks having memory of a user will be a differentiator for AI agents

Across our apps, there are now almost 1 billion monthly actives using Meta AI. Our focus for this year is deepening the experience and making Meta AI the leading personal AI with an emphasis on personalization, voice conversations and entertainment. I think that we’re all going to have an AI that we talk to throughout the day, while we’re browsing content on our phones, and eventually, as we’re going through our days with glasses. And I think that this is going to be one of the most important and valuable services that has ever been created.

In addition to building Meta AI into our apps, we just released our first Meta AI stand-alone app. It is personalized. So you can talk to it about interests that you’ve shown while browsing Reels or different content across our apps. And we built a social feed into it. So you can discover entertaining ways that others are using Meta AI. And initial feedback on the app has been good so far.

Over time, I expect the business opportunity for Meta AI to follow our normal product development playbook. First, we build and scale the product. And then once it is at scale, then we focus on revenue. In this case, I think that there will be a large opportunity to show product recommendations or ads as well as a premium service for people who want to unlock more compute for additional functionality or intelligence. But I expect that we’re going to be largely focused on scaling and deepening engagement for at least the next year before we’ll really be ready to start building out the business here…

…Earlier this year, we began testing the ability for Meta AI to better personalize its responses by remembering certain details from people’s prior queries and considering what that person engages with on our apps. We are already seeing this lead to deeper engagement with people we’ve rolled it out to, and it is now built into Meta AI across Facebook, Instagram, Messenger and our new stand-alone Meta AI app in the U.S. and Canada…

…The top use case right now for Meta AI from a query perspective is really around information gathering as people are using it to search for and understand and analyze information followed by social interactions from — ranging from casual chatting to more in-depth discussion or debate. We also see people use it for writing assistance, interacting with visual content, seeking help…

…WhatsApp continues to see the strongest Meta AI usage across our Family of Apps. Most of that WhatsApp engagement is in one-on-one Threads, followed by Facebook, which is the second largest driver of Meta AI engagement, where we’re seeing strong engagement from our feed deep dives integration that lets people ask Meta AI questions about the content that’s recommended to them…

…I also think that the stand-alone app is going to be particularly important in the United States because WhatsApp, as Susan said, is the largest surface that people use Meta AI and which makes sense. If you want to text an AI, having that be closely integrated and a good experience in the messaging app that you use makes a lot of sense. But we’re — while we have more than 100 million people use WhatsApp in the United States, we’re clearly not the primary messaging app in the United States at this point. iMessage is. We hope to become the leader over time. But we’re in a different position there than we are in most of the rest of the world on WhatsApp. So I think that the Meta AI app as a stand-alone is going to be particularly important in the United States to establishing leadership in — as the main personal AI that people use…

…I think that there are going to be a number of different agents that people use, just like people use different apps for different things. I’m not sure that people are going to use multiple agents for the same exact things, but I’d imagine that something that is more focused on kind of enterprise productivity might be different from something that is somewhat more optimized for personal productivity. And that might be somewhat different from something that is optimized for entertainment and social connectivity. So I think there will be different experiences…

…Once an AI starts getting to know you and what you care about in context and can build up memory from the conversations that you’ve had with it over time, I think that will start to become somewhat more of a differentiator.

Meta’s management continues to think of glasses as the ideal form factor for an AI device; management thinks that the 1 billion people in the world today who wear glasses will likely all be wearing smart glasses in the next 5-10 years; management thinks that building the devices people use for Meta’s apps lets the company deliver the best AI and social experiences; sales of the Ray-Ban Meta AI glasses have tripled in the last year and usage of the glasses is high; Meta has new launches of smart glasses lined up for later this year; monthly actives of Ray-Ban Meta AI glasses is up 4x from a year ago, with the number of people using voice commands growing even faster; management has rolled out live translations on Ray-Ban Meta AI glasses to all markets for English, French, Italian and Spanish; management continues to want to scale the Ray-Ban Meta AI glasses to 10 million units or more for its 3rd generation; management intends to run the same monetisation playbook with the Ray-Ban Meta AI glasses as Meta’s other products

Glasses are the ideal form factor for both AI and the metaverse. They enable you to let an AI see what you see, hear what you hear and talk to you throughout the day. And they let you blend the physical and digital worlds together with holograms. More than 1 billion people worldwide wear glasses today, and it seems highly likely that these will become AI glasses over the next 5 to 10 years. Building the devices that people use to experience our services lets us deliver the highest-quality AI and social experiences…

…Ray-Ban Meta AI glasses have tripled in sales in the last year. The people who have them are using them a lot. We’ve got some exciting new launches with our partner, EssilorLuxottica, later this year as well that should expand that category and add some new technological capabilities to the glasses…

…We’re seeing very strong traction with Ray-Ban Meta AI glasses with over 4x as many monthly actives as a year ago. And the number of people using voice commands is growing even faster as people use it to answer questions and control their glasses. This month, we fully rolled out live translations on Ray-Ban Meta AI glasses to all markets for English, French, Italian and Spanish. Now when you are speaking to someone in one of these languages, you’ll hear what they say in your preferred language through the glasses in real time…

…If you look at some of the leading consumer electronics products of other categories, by the time they get to their third generation, they’re often selling 10 million units and scaling from there. And I’m not sure if we’re going to do exactly that, but I think that that’s like the ballpark of the opportunity that we have…

…As a bunch of the products start to hit and start to grow even bigger than the number that I just said is just sort of like the sort of a near-term milestone, then I think we’ll continue scaling in terms of distribution. And then at some point, just like the other products that we build out, we will feel like we’re at a sufficient scale that we’re going to primarily focus on making sure that we’re monetizing and building an efficient business around it.

Meta released the first few Llama 4 models in April 2025 and more Llama 4 models are on the way, including the massive Llama 4 Behemoth model; management thinks leading-edge AI models are critical for Meta’s business, so they want the company to control its own destiny; by developing its own models, Meta is also able to optimise the model to its infrastructure and use-cases; an example of the optimisation is the Llama 4 17-billion model that comes with low latency to suit voice interactions; another example of the optimisation is the models’ industry-leading context window length which helps Meta AI’s personalisation efforts; Llama 4 Behemoth is important for Meta because all the models the company is using internally, and some of the models the company will develop in the future, are distilled from Behemoth

We released the first Llama 4 models earlier this month. They are some of the most intelligent, best multimodal, lowest latency and most efficient models that anyone has built. We have more models on the way, including the massive Llama 4 Behemoth model…

…On the LLM, yes, there’s a lot of progress being made in a lot of different dimensions. And the reason why we want to build this out is — one is that we think it’s important that for kind of how critical this is for our business that we sort of have control of our own destiny and are not depending on another company for something so critical. But two, we want to make sure that we can shape the development to be optimized for our infrastructure and the use cases that we want.

So to that end, Llama 4, the shape of the model with 17 billion parameters per expert was designed specifically for the infrastructure that we have in order to provide the low latency experience to be voice optimized. One of the key things, if you’re having a voice conversation with AI, is it needs to be low latency. So that way, when you’re having a conversation with it, there’s isn’t a large gap between when you stop speaking and it starts. So everything from the shape of the model to the research that we’re doing to techniques that go into it are kind of fit into that.

Similarly, another thing that we focused on was context window length. And in some of our models, we have really — we’re industry-leading on context window length. And part of the reason why we think that that’s important is because we’re very focused on providing a personalized experience. And there are different ways that you can put personalization context into an LLM, but one of the ways to do it is to include some of that context in the context window. And having a long context window that can incorporate a lot of the background that the person has shared across our apps is one way to do that…

…I think it’s also very important to deliver big models like Behemoth, not because we’re going to end up serving them in production, but because of the technique of distilling from larger models, right? The Llama 4 models that we’ve published so far and the ones that we’re using internally and some of the ones that we’ll build in the future are basically distilled from the Behemoth model in order to get the 90%, 95% of the intelligence of the large model in a form factor that is much lower latency and much more efficient.

Meta’s management is accelerating the buildout of Meta’s AI capacity, leading to higher planned investment for 2025; Meta’s capex growth in 2025 is for both generative AI and core business needs with the majority of overall capex supporting Meta’s core business; management continues to build infrastructure in a flexible way where the company can react to how the AI ecosystem develops in the coming years; management is increasing the efficiency of Meta’s workloads and this has helped the company to achieve strong returns from its core AI initiatives

We are accelerating some of our efforts to bring capacity online more quickly this year as well as some longer-term projects that will give us the flexibility to add capacity in the coming years as well. And that has increased our planned investment for this year…

…Our primary focus remains investing capital back into the business with infrastructure and talent being our top priorities…

…Our CapEx growth this year is going toward both generative AI and core business needs with the majority of overall CapEx supporting the core. We expect the significant infrastructure footprint we are building will not only help us meet the demands of our business in the near term but also provide us an advantage in the quality and scale of AI services we can deliver. We continue to build this capacity in a way that grants us maximum flexibility in how and when we deploy it to ensure we have the agility to react to how the technology and industry develop in the coming years…

…The second way we’re meeting our compute needs is by increasing the efficiency of our workloads. In fact, many of the innovations coming out of our ranking work are focused on increasing the efficiency of our systems. This emphasis on efficiency is helping us deliver consistently strong returns from our core AI initiatives.

Meta’s management sees a number of long-term tailwinds that AI can provide for Meta’s business, including making advertising a larger share of global GDP, and freeing up more time for people to engage in entertainment

Over the coming years, I think that the increased productivity from AI will make advertising a meaningfully larger share of global GDP than it is today…

…Over the long term, as AI unlocks more productivity in the economy, I also expect that people will spend more of their time on entertainment and culture, which will create an even larger opportunity to create more engaging experiences across all of these apps.

Meta’s management still expects to develop an AI coding agent sometime in 2025 that can operate as a mid-level engineer; management expects this AI coding agent to be do a substantial part of Meta’s AI research and development in 2026 H2; management is focused on building AI that can run experiments to improve Meta’s recommendation systems

I’d say it’s basically still on track for something around a mid-level engineer kind of starting to become possible sometime this year, scaling into next year. So I’d expect that by the middle to end of next year, AI coding agents are going to be doing a substantial part of AI research and development. So we’re focused on that. Internally, we’re also very focused on building AI agents or systems that can help run different experiments to increase recommendations across our other AI products like the ones that do recommendations across our feeds and things like that.

Microsoft (NASDAQ: MSFT)

Microsoft’s management is seeing accelerating demand across industries for cloud migrations; there are 4 things happening to drive cloud migrations, (1) classic migration, (2) data growth, (3) growth in cloud-native companies’ consumption, and (4) growth in AI consumption, which also requires non-AI consumption 

When it comes to cloud migrations, we saw accelerating demand with customers in every industry, from Abercrombie in French, to Coca-Cola and ServiceNow expanding their footprints on Azure…

…[Question] On your comment about accelerating demand for cloud migrations. I’m curious if you could dig in and extrapolate a little more what you’re seeing there.

[Answer] One is, I’ll just say, the classic migration of whether it’s SQL, Windows Server. And so that sort of again got good steady-state progress because the reality is, I think everyone is now, perhaps there’s another sort of kick in the data center migrations just because of the efficiency the cloud provides. So that’s sort of one part.

The second piece is good data growth. You saw some — like Postgres on Azure — I mean, forgetting even SQL server, Postgres on Azure is growing. Cosmos is growing. The analytics stuff I talked about with Fabric. It’s even the others, whether it is Databricks or even Snowflake on Azure are growing. So we feel very good about Fabric growth and our data growth.

Then the cloud-native growth. So this is again before we even get to AI, some of the core compute consumption of cloud-native players is also pretty very healthy. It was healthy throughout the quarter. We projected to go moving forward as well.

Then the thing to notice is the ratio, and I think we mentioned this multiple times before, if you look underneath even ChatGPT, in fact, that team does a fantastic job of thinking about not only their growth in terms of AI accelerators they need, they use Cosmos DB, they use Postgres. They use core compute and storage. And so there’s even a ratio between any AI workload in terms of AI accelerator to others.

So those are the 4 pockets, I’d say, or 4 different trend lines, which all have a relationship with each other.

Foundry is now used by developers in over 70,000 companies, from enterprises to startups, to design, customize and manage their AI apps and agents; Foundry processed  more than 100 trillion tokens in 2025 Q1, up 5x from a year ago; Foundry now has industry-leading model fine tuning tools; the latest models from AI heavyweights including OpenAI and Meta are available on Foundry;  Microsoft’s Phi family of SLMs (small language model) now has over 38 million downloads (20 million downloads in 2024 Q4); Foundry will soon introduce an LLM (large language model) with 1 billion parameters that can run on just CPUs

Foundry is the agent in AI app factory. It’s now used by developers at over 70,000 enterprises and digital natives from Atomicwork to Epic, Fujitsu and Gainsight to H&R Block and LG Electronics to design, customize and manage their AI apps and agents. We processed over 100 trillion tokens this quarter, up 5x year-over-year, including a record 50 trillion tokens last month alone. And 4 months in, over 10,000 organizations have used our new agent service to build, deploy and scale their agents.

This quarter, we also made a new suite of fine-tuning tools available to customers with industry-leading reliability, and we brought the latest models from OpenAI along with new models from Cohere, DeepSeek, Meta, Mistral, Stability to Foundry. And we’ve expanded our Phi family of SLMs with new multimodal and mini models. All-up, Phi has been downloaded 38 million times. And our research teams are taking it one step further with BitNet b1.58, a billion parameter, large language model that can run on just CPUs coming to the Foundry.

With agent mode in VS Code, Github Copilot can now iterate on code, recognize errors, and fix them automatically; there are other Github agent modes that provide coding support to developers; Microsoft is previewing a first-of-its-kind SWE (software engineering) agent that can execute developer tasks; GitHub Copilot now has 15 million users, up 4x from a year ago; GitHub Copilot is used by a wide range of companies; VS Code has more than 50 million monthly active users

We’re evolving GitHub Copilot from paired to peer programmer with agent mode in VS Code, Copilot can now iterate on code, recognize errors and fix them automatically. This adds to other Copilot agents like Autofix, which helps developers remediate vulnerabilities as well as code review agent, which has already reviewed over 8 million pull requests. And we are previewing a first-of-its-kind SWE-agent capable of asynchronously executing developer tasks. All-up, we now have over 15 million GitHub Copilot users, up over 4x year-over-year. And both digital natives like Twilio and enterprises like Cisco, HPE, Skyscanner and Target continue to choose GitHub Copilot to their developers with AI throughout the entire dev life cycle. With Visual Studio and VS Code, we have the world’s most popular editor with over 50 million monthly active users.

Microsoft 365 Copilot is now used hundreds of thousands of customers, up 3x from a year ago; deal sizes for Microsoft 365 Copilot continue to grow; a record number of customers in 2025 Q1 returned to buy more seats for Microsoft 365 Copilot; new researcher and analyst deep reasoning agents can analyze vast amounts of web and enterprise data on-demand directly within Microsoft 365 Copilot; Microsoft is introducing agents for every role and business process; customers can build their own AI agents with no/low code with Copilot Studio and these agents can handle complex tasks, including taking action across desktop and web apps; 230,000 organisations, including 90% of the Fortune 500, have already used Copilot Studio; customers created more than 1 million custom agents across SharePoint and Copilot Studio, up 130% sequentially

Microsoft 365 Copilot is built to facilitate human agent collaboration, hundreds of thousands of customers across geographies and industries now use Copilot, up 3x year-over-year. Our overall deal size continues to grow. In this quarter, we saw a record number of customers returning to buy more seats. And we’re going further. Just last week, we announced a major update, bringing together agents, notebooks, search and create into a new scaffolding for work. Our new researcher and analyst deep reasoning agents analyze vast amounts of web and enterprise data to deliver highly skilled expertise on demand directly within Copilot…

…We are introducing agents for every role and business process. Our sales agent turns contacts into qualified leads and with sales chat reps can quickly get up to speed on new accounts. And our customer service agent is deflecting customer inquiries and helping service reps resolve issues faster.

With Copilot Studio, customers can extend Copilot and build their own agents with no code, low code. More than 230,000 organizations, including 90% of the Fortune 500 have already used Copilot Studio. With deep reasoning and agent flows in Copilot Studio, customers can build agents that perform more complex tasks and also handle deterministic scenarios like document processing and financial approvals. And they can now build Computer Use Agents that take action on the UI across desktop and web apps. And with just a click, they can turn any SharePoint site into an agent, too. This quarter alone, customers created over 1 million custom agents across SharePoint and Copilot Studio, up 130% quarter-over-quarter.

Azure grew revenue by 33% in 2025 Q1 (was 31% in 2024 Q4), with 16 points of growth from AI services (was 13 points in 2024 Q4); management brought capacity online for Azure AI services faster than expected;  Azure’s non-AI business saw accelerated growth in its Enterprise customer segment as well as some improvement in its scale motions; management thinks the real outperfomer within Azure in 2025 Q1 is the non-AI business; the strength in the AI business in 2025 Q1 came because Microsoft was able to match supply and demand somewhat, and also deliver supply early to some customers; management thinks it’s getting harder to separate an AI workload from a non-AI workload

In Azure and other cloud services, revenue grew 33% and 35% in constant currency, including 16 points from AI services. Focused execution drove non-AI services results, where we saw accelerated growth in our Enterprise customer segment as well as some improvement in our scale motions. And in Azure AI services, we brought capacity online faster than expected…

…The real outperformance in Azure this quarter was in our non-AI business. So then to talk about the AI business, really, what was better was precisely what we said. We talked about this. We knew Q3 that we had and hadn’t really match supply and demand pretty carefully and so didn’t expect to do much better than we had guided to on the AI side. We’ve been quite consistent on that. So the only real upside we saw on the AI side of the business was that we were able to deliver supply early to a number of customers…

…[Question] You mentioned that the upside on Azure came from the non-AI services this time around. I was wondering if you could just talk a little bit more about that.

[Answer] In general, we saw better-than-expected performance across our segments, but we saw acceleration in our largest customers. We call that the Enterprise segment in general. And then in what we talked about of our scale motions, where we had some challenges in Q2, things were a little better. And we still have some work to do in our scale motions, and we’re encouraged by our progress. We’re excited to stay focused on that as, of course, we work through the final quarter of our fiscal year…

…It’s getting harder and harder to separate what an AI workload is from a non-AI workload.

Around half of Microsoft’s cloud and AI-related capex in 2025 Q1 (FY2025 Q3) are for long-lived assets that will support monetisation over the next 15 years and more, while the other half are for CPUs and GPUs; management expects Microsoft’s capex in 2025 Q2 (FY2025 Q4) to increase sequentially, but the guidance for total capex for FY2025 H2 is unchanged from previous guidance (previously, expectation was for capex for 2025 Q1 and 2025 Q2 to be at similar levels as 2024 Q4 (FY2025 Q2); FY2026’s capex is still expected to grow at a lower rate than in FY2025; the mix of spend in FY2026 will shift to short-lived assets in FY2026; demand for Azure’s AI services is growing faster than capacity is being brought online and management expects to have some AI capacity constraints beyond June 2025 (or FY2025 Q4); management’s goal with Microsoft’s data center investments is to be positioned for the workload growth of the future; management thinks pretraining plus test-time compute is a big change in terms of model-training workloads; Microsoft is short of power in fulfilling its data center growth plans; Microsoft’s data center builds have very long lead-times; in Microsoft’s 2024 Q4 (FY 2025 Q1) earnings call, management expected Azure to no longer be capacity-constrained by the end of 2025 Q2 (FY2025 Q4) but demand was stronger than expected in 2025 Q1 (FY2025 Q3); management still thinks they can get better and better capital efficiency from the cloud and AI capex; Azure’s margin on the AI business now is far better than what the margin was when the cloud transition was at a similar stage

Roughly half of our cloud and AI-related spend was on long-lived assets that will support monetization over the next 15 years and beyond. The remaining cloud and AI spend was primarily for servers, both CPUs and GPUs, to serve customers based on demand signals, including our customer contracted backlog of $315 billion…

…We expect Q4 capital expenditures to increase on a sequential basis. H2 CapEx in total remains unchanged from our January H2 guidance. As a reminder, there can be quarterly spend variability from cloud infrastructure build-outs and the timing of delivery of finance leases…

…Our earlier comments on FY ’26 capital expenditures remain unchanged. We expect CapEx to grow. It will grow at a lower rate than FY ’25 and will include a greater mix of short-lived assets, which are more directly correlated to revenue than long-lived assets…

… In our AI services, while we continue to bring data center capacity online as planned, demand is growing a bit faster. Therefore, we now expect to have some AI capacity constraints beyond June…

…the key thing for us is to have our builds and lease be positioned for what is the workload growth of the future, right? So that’s what you have to [ goal ] seek to. So there’s a demand part to it, there is the shape of the workload part to it, and there is a location part to it. So you don’t want to be upside down on having one big data center in one region when you have a global demand footprint. You don’t want to be upside down when the shape of demand changes because, after all, with essentially pretraining plus test-time compute, that’s a big change in terms of how you think about even what is training, right, forget inferencing…

…We will be short power. And so therefore — but it’s not a blanket statement. I need power in specific places so that we can either lease or build at the pace at which we want…

…From land to build to build-outs can be lead times of 5 to 7 years, 2 to 3 years. So we’re constantly in a balancing position as we watch demand curves…

…I did talk about in my comments, we had hoped to be in balance by the end of Q4. We did see some increased demand as you saw through the quarter. So we are going to be a little short still, say, a little tight as we exit the year…

…[Question] You’ve said in the past that you can attain better and better capital efficiency with the cloud business and probably cloud and AI business. Where do you stand today?

[Answer] The way, of course, you’ve seen that historically is right when we went through the prior cloud transitions, you see CapEx accelerate, you build out data center footprint.,, You slowly filled GPU capacity. And over time, you see software efficiencies and hardware efficiencies build on themselves. And you saw that process for us for goodness now quite a long time. And what Satya’s talking about is how quickly that’s happening on the AI side of the business and you add to that model diversity. So think about the same levers plus model efficiency, those compounds. Now the one thing that’s a little different this time is just the pace. And so when you’re seeing that happen, pace in terms of efficiency side, but also pace in terms of the build-out. So it can mask some of the progress… Our margins on the AI side of the business are better than they were at this point by far than when we went through the same transition in the server to cloud transition…

…I think the way to think about this is you can ask the question, what’s the difference between a hosting business and a hyperscale business? It’s software. That’s, I think, the gist of it. Yes, for sure, it’s a capital-intensive business, but capital efficiency comes from that system-wide software optimization. And that’s what makes the hyperscale business attractive and that’s what we want to just keep executing super well on.

Microsoft’s management sees Azure as Microsoft’s largest business; management thinks that the next platform shift in technology, which is AI, is built on the last major platform, which was for cloud computing, so this benefits Microsoft

There’s nothing certain for sure in the future, except for one thing, which is our largest business is our infrastructure business. And the good news here is the next big platform shift builds on that. So it’s not a complete rebuild, having gone through all these platform shifts where you have to come out on the other side with a full rebuild. If there is good news here is that we have a good business in Azure that continues to grow and the new platform depends on that.

It’s possible that software optimizations with AI model development and deployment could lead to even longer useful lives for GPUs, but management wants to observe this for longer

[Question] Could we start to consider the possibility that software enhancements might extend the useful life assumption that you’re using for GPUs?

[Answer] In terms of thinking about the depreciable life of an asset, we like to have a long history before we make any of those changes. So we’re focused on getting every bit of useful life we can, of course, out of assets. But to Satya’s point, that tends to be a software question more than a hardware one.

Netflix (NASDAQ: NFLX)

Netflix’s content talent are already using AI tools to improve the content production process; management thinks AI tools can enable lower-budget projects to access top-grade VFX; Rodrigo Prieto is directing his first feature film with Netflix in 2025, Pedro Paramo, and he’s able to use AI tools for de-aging VFX at a much lower cost than The Irishman film that Prieto worked on 5 years ago; the entire budget for Pedro Paramo is similar to the cost of VFX alone for The Irishman; management’s focus with AI is to find ways for AI to improve the member and creator experience

So our talent today is using AI tools to do set references or previs, VFX sequence prep, shop planning, all kinds of things today that kind of make the process better. Traditionally, only big budget projects would have access to things like advanced visual effects such as de-aging. So today, you can use these AI-powered tools so to enable smaller budget projects to have access to big VFX on screen.

A recent example, I think, is really exciting. Rodrigo Prieto was the DP on The Irishman just 5 years ago. And if you remember that movie, we were using very cutting edge, very expensive de-aging technology that still had massive limitations, still creating a bunch of complexity on set for the actors. It was a giant leap forward for sure, but nowhere near what we needed for that film. So this year, just 5 years later, Rodrigo is directing his first feature film for us, Pedro Páramo in Mexico. Using AI-powered tools he was able to deliver this de-aging VFX to the screen for a fraction of what it cost on The Irishman. In fact, the entire budget of the film was about the VFX cost on The Irishman…

…So our focus is simple, find ways for AI to improve the member and the creator experience.

Netflix’s management is building interactive search into Netflix which is based on generative AI

We’re also building out like new capabilities, an example would be interactive search. That’s based on generative technologies. We expect that will improve that aspect of discovery for members.

Paycom Software (NYSE: PAYC)

Paycom’s GONE is the industry’s first fully automated time-off solution, utilising AI, that automates all time off requests; prior to GONE, 10% of an organisation’s labour cost was unmanaged; GONE can generate ROI of up to 800%, according to Forrester; GONE helped Paycom be named by Fast Company as one of the world’s most innovative companies

Our award-winning solution, GONE, is a perfect example of how Paycom simplifies tests through automation and AI. GONE is the industry’s first fully automated time-off solution that decisions all time-off requests based on customizable guidelines set by the company’s time-off rules. Before GONE, 10% of an organization’s labor cost went substantially unmanaged, creating scheduling errors, increased cost from overpayments, staffing shortages and employee uncertainty over pending time-off requests. According to a Forrester study, GONE’s automation delivers an ROI of up to 800% for clients. GONE continues to receive recognition. Most recently, Fast Company magazine named Paycom, one of the world’s most innovative companies for a second time. This honor specifically recognized GONE and is a testament to how Paycom is shaping our industry by setting new standards for automation across the globe.

PayPal (NASDAQ: PYPL)

PayPal’s management is leaning into agentic commerce; PayPal recently launched the payments industry’s first remote MCP (Model Context Protocol) server to enable AI agent frameworks to integrate with PayPal APIs; the introduction of the MCP allows any business to create an agentic commerce experience; all major AI players are involved with PayPal’s annual Developer Days to engage PayPal’s developer community

At Investor Day, I told you we were leaning into agentic commerce…

…Just a few weeks ago, we launched the industry’s first remote MCP server and enabled the leading AI agent frameworks to seamlessly integrate with PayPal APIs. Now any business can create agentic experience that allow customers to pay, track shipments, manage invoices and more, all powered by PayPal and all within an AI client. As we speak, developers are gathering in our San Jose headquarters for our annual Developer Days. Every major player in AI is represented, providing demos and engaging with our developer community.

Shopify (NASDAQ: SHOP)

Shopify’s management recently launched TariffGuide.ai, an AI-powered tool that provides duty rates based on just a product description and the country of origin, helping merchants source the right products in minutes

And just this past week, we launched TariffGuide.ai. This AI driven tool provides duty rates based on just a product description and the country of origin. Sourcing the right products from the right country can mean the difference between a 0% and a 15% duty rate or higher, And TariffGuide.ai allows merchants to do this in minutes, not days.

Shopify CEO Tobi Lutke penned a memo recently on his vision on how Shopify should be workin with AI; AI is becoming 2nd nature to how Shopify’s employees work, where employees use AI reflexively; before any team requests for additional headcount, they need to first assess if AI can meet their goals; Shopify has built a dozen MCP (model context protocol) servers in the last few weeks to enable anyone in Shopify to ask questions and find resources more easily; management sees AI being a cornerstone of how Shopify delivers value; management is investing more in AI, but the increased investment is not a driver for the lower gross margin in Shopify’s Subscription Solutions segment in 2025 Q1; management does not expect the Subscription Solutions segment’s gross margin to change much in the near term; Shopify has shown strong operating leverage partly because of its growing internal use of AI

AI is at the core of how we operate and is transforming our work processes. For those who have not seen it, I encourage you to check out Toby’s recent company wide email on AI that has now been shared publicly. At Shopify, we take AI seriously. In fact, it’s becoming second nature to how we work. By fostering a culture of reflexive AI usage, our teams default to using AI first, reflexive being the key term here. This also means that before requesting additional headcount or resources, teams are required to start with assessing how they can meet their goals using AI first. This approach is sparking some really fascinating explorations and discussions around the company, challenging the way we think, the way we operate, and pushing us to look ahead as we redefine our decision making processes. In the past couple of weeks, we built a dozen MCP servers that make Shopify’s work legible and accessible. And now anyone within Shopify can ask questions, find resources, and leverage those tools for greater efficiency. This reflexive use of AI goes well beyond internal improvements. It supercharges our team’s capabilities and drives operational efficiencies, keeping us agile. And as we continue to innovate, AI will remain a cornerstone of how we deliver value across the board…

…Gross profit for Subscription Solutions grew 19%, slightly less than the 21% revenue growth for Subscription Solutions. The lower rate was driven primarily by higher cloud and infrastructure hosting costs needed to support higher volumes and geographic expansion. Although we are investing more in AI, it is not a significant factor in this increase. Over the past 5 years, the gross margin for Subscription Solutions has centered around 80%, plus or minus a couple of hundred basis points in any given quarter, and we do not anticipate that trend changing in the near term…

…Our continued discipline on head count across all 3 of R&D, sales and marketing and G&A continued to yield strong operating leverage, all while helping us move even faster on product development aided by our increasing use of AI.

Shopify’s management rearchitected the AI engine of Sidekick, Shopify’s AI merchant assistant, in 2025 Q1; monthly average users of Sidekick has more than doubled since the start of 2025; early results of Sidekick are really strong for both large and small merchants

In Q1, key developments for Sidekick included a complete rearchitecture of the AI engine for deeper reasoning capabilities, enhancing processing of larger business datasets and accessibility in all supported languages, allowing every Shopify merchant to use Sidekick in their preferred language. And these changes, well, they’re working. In fact, our monthly average users of Sidekick continue to climb more than doubling since the start of 2025. Now this is still really early days, but the progress we are making is already yielding some really strong results for merchants, both large and small. 

Shopify acquired Vantage Discovery in 2025 Q1; Vantage Discovery works on AI-powered, multi-vector search; management thinks the acquisition will improve the overall consumer search experience delivered by Shopify’s merchants

In March, we closed the acquisition of Vantage Discovery, which helps accelerate the development of AI-powered, multi-vector search across our search, APIs, shop and storefront search offerings. This acquisition is one piece of a broader strategy to ensure that our merchants are able to continue meeting buyers regardless of where they’re shopping or discovering great products…

…The Vantage team coming in who are rock stars in AI are going to help take our search abilities to the next level.

Shopify’s management is seeing more and more commerce searches starting away from a search engine; Shopify is already working with AI chatbot providers on AI shopping; management thinks that AI shopping is a huge opportunity; management thinks AI agents will be a great opportunity for Shopify too

One of the things we think about is that wherever commerce is taking place, Shopify will be there. And obviously, one of the things we are seeing is that more and more searches are starting on places beyond just somebody’s search engine. That’s a huge opportunity whereby more consumers are going to be searching for great products…

…We’ve talked about some of the partnerships in the past. You’ve seen what we’ve done with Perplexity and OpenAI. We will continue doing that. We’re not going to front run our product road map when it comes to anything, frankly. But we do think though that AI shopping, in particular, is a huge opportunity…

…[Question] How does Shopify view the emergence of AI agents in terms of do you guys see this as an opportunity or more of a threat because, on one hand, they could facilitate direct checkout with their own platforms. On the other hand, this may also unlock some new sales channel for Shopify merchants, very similar to sort of what happened with social media commerce

[Answer] We think it’s a great opportunity. Look, the more channels that exist in the world, the more complexity it is for merchants and brands, that’s where the value of Shopify really shines. So if there’s a new surface area, whether it’s through AI agents or through just simply LLMs and AI wrappers, that consumer goes to, to look for a new pair of sneakers or a new cosmetic or a piece of furniture, they want to have access to the most interesting products for the most important brands, and those are all on Shopify. So for us, we think that all of these new areas where commerce is happening is a great thing. It allows Shopify to increase its value.

Taiwan Semiconductor Manufacturing Company (NYSE: TSM)

TSMC’s management continues to expect AI accelerators revenue to double in 2025; management has factored China-bans on US chips into TSMC’s 2025 outlook; AI-related demand outside of China appears to have become even stronger over the last 3 months

We reaffirm our revenue from AI accelerated to double in 2025. The AI accelerators we define as AI GPU, AI ASIC and HPM controllers for AI training and inference in the data center. Based on our customers’ strong demand, we are also working hard to double our CoWoS capacity in 2025 to support their needs…

…[Question] The geopolitical risk, micro concerns is one of the major uncertainty nowadays. Last 2 days, we have like H20 being banned in China, blah, blah, blah. So how does that impact to TSMC’s focus and production planning, right? Do we have enough other customers and demand to keep our advanced node capacity fully utilized? Or how does that change our long-term production planning moving forward?

[Answer] Of course, we do not comment on specific customers or product, but let me assure you that we have taken this into consideration when providing our full year’s growth outlook. Did I answer the question?…

…[Question] AI is still expected to double this year despite the U.S. ban on AI GPUs into China. And I guess, China was a meaningful portion of accelerated shipments well over 10% of volumes. So factoring this in, it would imply your AI outlook this year, still doubling would mean that the AI orders have improved meaningfully outside of China in the last sort of 3 months. Is that how we should interpret your comment about you still expect the business to double?

[Answer] 3 months ago, we are — we just cannot supply enough wafer to our customer. And now it’s a little bit balanced, but still, the demand is very strong. And you are right, other than China, the demand is still very strong, especially in U.S.

TSMC’s management has a disciplined approach when building capacity and management recognises how important the discipline is given the high forecasted demand for AI-related chips

At TSMC, higher level of capital expenditures is always correlated with higher growth opportunities in the following years. We reiterate our 2025 capital budget is expected to be between USD 38 billion and USD 42 billion as we continue to invest to support customers’ growth. About 70% of the capital budget will be allocated for advanced process technologies. About 10% to 20% will be spent for specialty technologies and about 10% to 20% will be spent for advanced packaging, testing, mass-making and others. Our 2025 CapEx also includes a small amount related to our recently announced additional $100 billion investment plan to expand our capacity in Arizona…

…To address the structural increase in the long-term market demand profile, TSMC employed a disciplined and robust capacity planning system. This is especially important when we have such high forecasted demand from AI-related business. Externally, we work closely with our customers and our customers’ customers to plan our capacity. Internally, our planning system involves multiple teams across several functions to assess and evaluate the market demand from both a top-down and bottom-up approach to determine the appropriate capacity build.

TSMC’s management expects the Foundry 2.0 industry to grow 10% year-on-year in 2025, driven by AI-related demand and mild recovery in other end markets; management expects TSMC to outperform the Foundry 2.0 industry in 2025

Looking at the full year of 2025, we expect Foundry 2.0 industry growth to be supported by robust AI-related demand and a mild recovery in other end market segment. In January, we had forecasted a Foundry 2.0 industry to grow 10 points year-over-year in 2025, which is consistent with IDC’s forecast of 11% year-over-year growth for Foundry 2.0…

…We are confident TSMC can continue to outperform the Foundry 2.0 industry growth in 2025.

TSMC’s management thinks impact from recent AI models, including DeepSeek, will lower the barrier to future long-term AI development; TSMC’s management continues to expect mid-40% revenue CAGR from AI accelerators in the 5-years starting from 2024

Recent developments are also positive to AI’s long-term demand outlook. In our assessment, the impact from AI recent models, including DeepSeek, will drive greater efficiency and help lower the barrier to future AI development. This will lead to wider usage and greater adoption of AI models, which all require use of leading-edge silicon. These developments only serve to strengthen our conviction in the long-term growth opportunities from the industry megatrend of 5G, AI and HPC…

…Based on our planning framework, we are confident that our revenue growth from AI accelerators will approach a mid-40s percentage CAGR for the next 5 years period starting from 2024.

TSMC’s 2nd fab in Arizona will utilise N3 process technology and is already complete and management wants to speed up volume production schedule to meet AI-related demand

Our first fab in Arizona has already successfully entered high-volume production in 4Q ’24, utilizing N4 process technology with a yield comparable to our fab in Taiwan. The construction of our second fab, which will utilize the 3-nanometer process technology, is already complete and we are working on speeding up the volume production schedule based on the strong AI-related demand from our customers. Our third and fourth fab will utilize N2 and A16 process technologies and with the expectation of receiving all the necessary permits are scheduled to begin construction later this year. Our fifth and sixth fab will use even more advanced technologies. The construction and ramp schedule for this fab will be based on our customers’ demand.

TSMC’s management believes its A16 technology has a best-in-class backside power delivery solution that is also the first in the industry; A16 is best suited for specific HPC (high-performance computing) products, which means it is best suited for AI-related workloads; A16 is scheduled for volume production in 2026 H2

We also introduced a 16 feature in super power rail or SPR as a separate offering. Compared with the N2P, A16 provides a further 8% to 10% speed improvement at the same power or 15% to 20% power improvement at the same speed and additional 7% to 10% chip density gain. A16 is best suited for specific HPC products with complex signal route and dense power delivery network. Volume production is scheduled for second half 2026.

Tesla (NASDAQ: TSLA)

Tesla’s management continues to expect fully autonomous Tesla rides in Austin, Texas in June 2025; management will sell full autonomy software for Model Y in Austin; management now demarcates CyberCab as a separate product, and all of the other models (S, 3, X, Y) that is compatible with autonomous software as being robotaxis; management reiterates that once Tesla can solve for autonomy in 1 city, it can very quickly scale because Tesla’s autonomous solution is a general solution, not a city-specific solution; Tesla’s autonomous solution involves AI and a specific Tesla-designed AI chip, as opposed to expensive sensors and high-precision maps; the fully autonomous Teslas in June 2025 in Austin will be Model Ys; management expects full autonomy in Tesla’s fleet to ramp up very quickly; management is confident that Tesla will have large-scale autonomy by 2026 H2, meaning, millions of fully autonomous Tesla vehicles by 2026 H2; even with the introduction of full autonomy, management thinks there will be some localised parameters – effectively a mixture of experts model – set for safety; management thinks Tesla’s autonomous solution can scale well because when the FSD (Full Self Driving) software was deployed in China, it used very minimal China-specific data and yet could work well in China; validation of Tesla’s autonomous solution will be important in determining its rate of acceptance; there are now convoys of Teslas in Austin running autonomously in testing in order to compress Tesla’s AI’s learning curve; a consumer in China used FSD on a narrow mountain dirt road; management expects FSD unsupervised to be available for personal use by end of 2025; Musk thinks the first Model Y ro drive itself from factory to customer will happen later in 2025; newly-manufactured Model Ys are already driving themselves around in Tesla factories

We expect to have — be selling fully autonomous rides in June in Austin as we’ve been saying for now several months. So that’s continued…

…Unsupervised autonomy will first be sold for the Model Y in Austin, and then actually, should parse out the term for robotic taxi or robotaxi and just generally like what’s the Cybercab because we’ve got a product called the Cybercab. And then any Tesla, which could be an S, 3, X or Y that is autonomous is a robotic taxi or a robotaxi. It’s very confusing. So the vast majority of the Tesla fleet that we’ve made is capable of being a robotaxi or a robotic taxi…

…Once we can make the system work where you can have paid rides, fully autonomously with no one in the car in 1 city, that is a very scalable thing for us to go broadly within whatever jurisdiction allows us to operate. So because what we’re solving for is a general solution to autonomy, not a city-specific solution for autonomy, once we make it work in a few cities, we can basically make it work in all cities in that legal jurisdiction. So if it’s — once we can make the pace to work in a few cities in America, we can make it work anywhere in America. Once we can make it work in a few cities in China, we can make it work anywhere in China, likewise in Europe, limited only by regulatory approvals. So this is the advantage of having a generalized solution using artificial intelligence and an AI chip that Tesla designed specifically for this purpose, as opposed to very expensive sensors and high-precision maps on a particular neighborhood where that neighborhood may change or often changes and then the car stops working. So we have a general solution instead of a specific solution…

…The Teslas that will be fully autonomous in June in Austin are fully Model Ys. So that is — it’s currently on track to be able to do paid rides fully autonomously in Austin in June, and then to be in many other cities in the U.S. by the end of this year.

It’s difficult to predict the exact ramp sort of week by week and month by month, except that it will ramp up very quickly. So it’s going to be like some — basically an S-curve where it’s very difficult to predict the intermediate slope of the S-curve, but you kind of know where the S-curve is going to end up, which is the vast majority of the Tesla fleet being autonomous. So that’s why I feel confident in predicting large-scale autonomy around the middle of next year, certainly the second half next year, meaning I bet that there will be millions of Teslas operating autonomously, fully autonomously in the second half of next year, yes…

…It does seem increasingly likely that there will be a localized parameter set sort of — especially for places that have, say, a very snowy weather, like I say, if you’re in the Northeast or something like this — you can think of — it’s kind of like a human. Like you can be a very good driver in California but are you going to be also a good driver in a blizzard in Manhattan? You’re not going to be as good. So there is actually some value in — you can still drive but your probability of an accident is higher. So the — it’s increasingly obvious that there’s some value to having a localized set of parameters for different regions and localities…

…You can see that from our deployment of FSD supervised in China with this very minimal data that’s China-specific, the model is generalized quite well to completely different driving styles. That just like shows that the AI-based solution that we have is the right one because if you had gone down the previous rule-based solutions, sort of like more hard-coded HD map-based solutions, it would have taken like many, many years to get China to work. You can see those in the videos that people post online themselves. So the generalized solution that we are pursuing is the right one that’s going to scale well…

…You can think of this like location-specific parameters that Elon alluded to as a mixture of experts. And if you are sort of familiar with the AI models, Grok and others, they all use this mixture of experts to sort of specialize the parameters to specific tasks while still being general…

…What are the critical things that need to get right, one thing I would like to note is validation. Self-driving is a long-tail problem where there can be a lot of edge cases that only happen very, very rarely. Currently, we are driving around in Austin using our QA fleet, but then super [ rare ] to get interventions that are critical for robotaxi operation. And so you can go many days without getting a single intervention. So you can’t easily know whether you are improving or regressing in your capacity. And we need to build out sophisticated simulations, including neural network-based video generation…

…There’s just always a convoy of Teslas going — just going all over to Austin in circles. But yes, I just can’t emphasize this enough. In order to get a figure on the long-tail things, it’s 1 in 10,000, that says 1 in 20,000 miles or 1 in 30,000. The average person drives 10,000 miles in a year. So not trying to compress that test cycle into a matter of a few months. It means you need a lot of cars doing a lot of driving in order to compress that to do in a matter of a month what would normally take someone a year…

…I saw one guy take a Tesla on — autonomously on a narrow dirt road across like a mountain. And I’m like, still a very brave person. And I said this driving along the road with no barriers where he makes a mistake, he’s going to plunge to his doom. But it worked…

…[Question] when will FSD unsupervised be available for personal use on personally-owned cars?

[Answer] Before the end of this year… the acid test being you should — can you go to sleep in your car and wait until your destination? And I’m confident that will be available in many cities in the U.S. by the end of this year…

…I’m confident also that later this year, the first Model Y will drive itself all the way to the customer. So from our — probably from a factory in Austin and our one in here in Fremont, California, I’m confident that from both factories, we’ll be able to drive directly to a customer from the factory…

…We have — it has been put to use — it’s doing useful work fully autonomously at the factories, as Ashok was mentioning, the cars drive themselves from end of line to where they supposed to be picked up by a truck to be taken to the customer… It’s important to note in the factories, we don’t have dedicated lengths or anything. People are coming out every day, trucks delivering supplies, parts, construction.

Tesla’s management expects thousands of Optimus robots to be working in Tesla factories by end-2025; management expects Optimus to be the fastest product to get to millions of units per year; management thinks Tesla can get to 1 million units annually in 4-5 years; management expects to make thousands of Optimus robots at the end of this year; there’s no existing supply chain for all of Optimus’s components, so Tesla has to build a supply chain from scratch; the speed of manufacturing of a product is governed by the speed of the slowest item in the supply chain, but in Optimus’s case, there are many, many such items since it’s so new; Optimus production is currently rate-limited by restrictions on rare-earth magnets from China but management is working on it; management still has no idea how Optimus’s supply chain will look like at maturity

Making good progress in Optimus. We expect to have thousands of Optimus robots working in Tesla factories by the end of this year beginning this fall. And we expect to see Optimus faster than any product, I think, in history to get to millions of units per year as soon as possible. I think we feel confident in getting to 1 million units per year in less than 5 years, maybe 4 years. So by 2030, I feel confident in predicting 1 million Optimus units per year. It might be 2029…

…This year, we’ll make a few — we do expect to make thousands of Optimus robots, but most of that production is going to be at the end of the year…

…Almost everything in Optimus is new. There’s not like an existing supply chain for the motors, gearboxes, electronics, actuators, really anything in the Optimus apart from the AI for Tesla, the Tesla AI computer, which is the same as the one in the car. So when you have a new complex manufactured product, it will move as fast as the slowest and the least lucky component in the entire thing. And as a first order approximation, there’s like 10,000 unique things. So that’s why anyone who tells you they can predict with precision, the production ramp of the truly new product is — doesn’t know what they’re talking about. It is literally impossible…

…Now Optimus was affected by the magnet issue from China because the Optimus actuators in the arm to use permanent magnet. Now Tesla, as a whole, does not need to use permanent magnets. But when something is volume constrained like an arm of the robot, then you want to try to make the motor as small as possible. And then — so we did design in permanent magnets for those motors and those were affected by the supply chain by basically China requiring an export license to send out any rare earth magnets. So we’re working through that with China. Hopefully, we’ll get a license to use the rare earth magnets. China wants some assurances that these are not used for military purposes, which obviously they’re not. They’re just going into a humanoid robot. So — and it’s a nonweapon system…

…[Question] Wanted to ask about the Optimus supply chain going forward. You mentioned a very fast ramp-up. What do you envision that supply chain looking like? Is it going to require many more suppliers to be in the U.S. now because of the tariffs?

[Answer] We’ll have to see how things settle out. I don’t know yet. I mean some things we’re doing, as we’ve already talked about, which is that we’ve already taken tremendous steps to localize our supply chain. We’re more localized than any other manufacturer. And we have a lot of things kind of underway that to increase the localization to reduce supply chain risk associated with geopolitical uncertainty.

Tesla’s supervised FSD (full-self driving) software is safer than a human driver; management has been using social media (X, or Twitter) to encourage people to try out Tesla’s FSD software; management did not directly answer a question on FSD pricing once the vehicle can be totally unsupervised

Not only is FSD supervised safer than a human driver, but it is also improving the lifestyle of individuals who experience it. And again, this is something you have to experience and anybody who has experienced just knows it. And we’ve been doing a lot lately to try and get those stories out, at least on X, so that people can see how other people have benefited from this…

…[Question] Can we envision when you launch unsupervised FSD that there could be sort of a multitiered pricing approach to unsupervised versus supervised similar to what you did with autopilot versus FSD in the past?

[Answer] I mean this is something which we’ve been thinking about. I mean just so now for people who have been trying FSD and who’ve been using FSD, they think given the current pricing is too cheap because for $99, basically getting a personal shop… I mean we do need to give people more time to — if they want to look at — like a key breakpoint is, can you read your text messages or not? Can you write a text message or not? Because obviously, people are doing this, by the way, with unautonomous cars all the time. And if you just go over and drive down the highway and you’ll see people texting while driving doing 80-mile an hour… So that value — it will really be profound when you can basically do whatever you want, including sleep. And then that $99 is going to seem like the best $99 you ever spent in your life.

Tesla’s management thinks Waymo vehicles are too expensive compared to Teslas; Waymo has expensive sensor suites; management thinks Tesla will have lion’s share of the robotaxi market; a big difference between Tesla and Waymo is that Tesla is also manufacturing the cars whereas Waymo is retrofitting cars from other parties; management thinks Tesla’s vision-only approach will not have issues with cameras becoming blinded by glare and stuff because the system uses direct photon counting and bypasses image signal processing

The issue with Waymo’s cars is it costs way more money, but that is the issue. The car is very expensive, made in low-volume. Teslas are probably cost 1/4, 20% of what a Waymo costs and made in very high volume. Ironically, like we’re the ones who made the bet that a pure AI solution with cameras and [ already ] what the car actually will listen for sirens and that kind of thing. It’s the right move. And Waymo decided that an expensive sensor suite is the way to go, even though Google is very good at AI. So I’m wondering…

….As far as I’m aware, Tesla will have, I don’t know, 99% market share or something ridiculous…

…The other thing which people forget is that we’re not just developing the software solution, we are also manufacturing the cars. And like you know what like Waymo has, they’re taking cars and then trying to…

…[Question] You’re still sticking with the vision-only approach. A lot of autonomous people still have a lot of concerns about sun glare, fog and dust. Any color on how you anticipate on getting around those issues? Because my understanding, it kind of blinds the camera when you get glare and stuff.

[Answer] Actually, it does not blind the camera. We use an approach which is a direct photon count. So when you see a processed image, so the image that goes from the — with sort of photon counter, the silicon photon counter, that they get — goes through a digital signal processor or image signal processor. That’s normally what happens. And then the image that you see looks all washed out because if it’s — you pointed a camera at the sun, the post-processing of the photon counting washes things out. It actually adds noise. So quite a big breakthrough that we made some time ago was to go with direct photon counting and bypass the image signal processor. And then you can drive pretty much straight at the sun, and you can also see in what appears to be the blackest of night. And then here in fog, we can see as well as people can, probably better, but in fact probably slightly better than people than the average person anyway.

Tesla’s AI software team and chip-design team was built from scratch with no acquisitions; management thinks Tesla’s team is the best

It is worth noting that Tesla has built an incredible AI software team and AI hardware chip design team from scratch, didn’t acquire anyone. We just built it. So yes, it’s really — I mean I don’t see anyone being able to compete with Tesla at present.

Tesla’s management thinks China is ahead of the USA in physical AI with respect to autonomous drones because China has the ability to manufacture autonomous drones, but the USA does not;  management thinks Tesla is ahead of any company in the world, even Chinese companies, in terms of humanoid autonomous robots 

[Question] Between China and United States, who, in your opinion, is further ahead on the development of physical AI, specifically on humanoid and also drones?

[Answer] A friend of mine posted on X, I reposted it. I think of a prophetic statement, which is any country that cannot manufacture its own drones is going to be the vassal state of any country that can. And we can’t — America cannot currently manufacture its own drones. Let that sink in, unfortunately. So China, I believe manufactures about 70% of all drones. And if you look at the total supply chain, China is almost 100% of drones are — have a supply chain dependency on China. So China is in a very strong position. And here in America, we need to tip more of our people and resources to manufacturing because this is — and I have a lot of respect for China because I think China is amazing, actually. But the United States does have such a severe dependency on China for drones and be unable to make them unless China gives us the parts, which is currently the situation.

With respect to humanoid robots, I don’t think there’s any company and any country that can match as well. Tesla and SpaceX are #1. And then I’m a little concerned that on the leaderboard, ranks 2 through 10 will be Chinese companies. I’m confident that rank 1 will be Tesla.

The Trade Desk (NASDAQ: TTD)

Trade Desk’s industry-leading Koa AI tools are embedded across Kokai; adoption of Kokai is now ahead of schedule, with 2/3 of clients using it; the bulk of spending on Trade Desk now takes place on Kokai; management continues to expect all Trade Desk clients to be using Kokai by end-2025; management is confident that Kokai will be seen as the most powerful buying platform by the industry by end-2025

The injection of our industry-leading Koa AI tools across every aspect of our platform has been a game changer, and we are just getting started…

…The core of Kokai has been delivered and adoption is now ahead of schedule. Around 2/3 of our clients are now using it and the bulk of the spend in our platform is now running through Kokai. We expect all clients to be using it by the end of the year…

…I’m confident that by the end of this year, we will reflect on Kokai as the most powerful buying platform the industry has ever seen, precisely because it combines client needs with the strong point of view on where value is shifting and how to deliver the most efficient return on ad spend.

…Kokai adoption now represents the majority of our spend, almost 2/3, a significant acceleration from where we ended 2024.

Deutsche Telekom used Kokai’s AI tools and saw an 11x improvement in post-click conversions and an 18x improvement in the cost of conversions; Deutsche Telekom is now planning to use Kokai across more campaigns and transition from Trade Desk’s previous platform, Solimar, like many other Trade Desk clients

Deutsche Telekom. They’re running the streaming TV service called Magenta TV, and they use our platform to try to grow their subscriber base…

…Using seed data from their existing customers, Deutsche Telekom was able to use the advanced AI tools in our Kokai platform to find new customers and define the right ad impressions across display and CTV, most relevant to retain those new customers successfully, and the results were very impressive. They saw an 11x improvement in post-click conversions attributed to advertising and an 18x improvement in the cost of those conversions. Deutsche Telekom is now planning to use Kokai across more campaigns, a transition that is fairly typical as clients move from our previous platform, Solimar to our newer, more advanced AI fuel platform, Kokai.

Visa (NASDAQ: V)

Visa recently announced the Authorize.net product that features AI capabilities, including an AI agent; Authorize.net enables all different types of payments;

In Acceptance Solutions, we recently announced 2 new product offerings. The first is a completely new version of Authorize.net, launching in the U.S. next quarter and additional countries next year. It features a streamlined user in base; AI capabilities with an AI agent, Anet; improved dashboards for day-to-day management and support for in-person card readers and Tap to Phone. It will help businesses analyze data, summarize insights and adapt to rapidly changing customer trends…

……I talked about the Authorize.net platform that we’ve relaunched and we’re relaunching. That’s a great example of enabling all different types of payments. And that’s going to be, we think, a really positive impact in the market specifically focused on growing our share in small business checkout.

Visa has an enhanced holistic fraud protection solution known as Adaptive Real-time Individual Change identification (ARIC) Risk Hub; ARIC Risk Hub uses AI to build more accurate risk profiles;

We also now provide an enhanced holistic fraud protection solution from Featurespace called the Adaptive Real-time Individual Change identification, or ARIC, Risk Hub. This solution utilizes machine learning and AI solutions to enable clients to build more accurate risk profiles and more confidently detect and block fraudulent transactions, ultimately helping to increase approvals and stop bad actors in real time. 

---

Disclaimer: The Good Investors is the personal investing blog of two simple guys who are passionate about educating Singaporeans about stock market investing. By using this Site, you specifically agree that none of the information provided constitutes financial, investment, or other professional advice. It is only intended to provide education. Speak with a professional before making important decisions about your money, your professional life, or even your personal life. I have a vested interest in Alphabet, Amazon, Apple, ASML, Coupang, Datadog, Mastercard, Meta Platforms, Microsoft, Netflix, Paycom Software, PayPal, Shopify, TSMC, Tesla, The Trade Desk and Visa. Holdings are subject to change at any time.

Warren Buffett is one of my investment heroes. On 3 May 2025, he held court at the 2025 Berkshire Hathaway AGM (annual general meeting).

For many years, I’ve anticipated the AGM to hear his latest thoughts. This year’s session holds special significance because it may well be his last – during the AGM, he announced that he would be stepping down as CEO of Berkshire Hathaway by the end of this year, ending an amazing 60-year run since becoming the company’s leader in 1965. Greg Abel is slated to be Berkshire Hathaway’s next CEO.

The most recent Berkshire meeting contained great insights from Buffett and other senior Berkshire executives that I wish to share and document. Before I get to them, I would like to thank my friend Thomas Chua for performing a great act of public service. Shortly after the AGM ended, Thomas posted a transcript of the session at his excellent investing website Steady Compounding. 

Without further ado, the italicised passages between the two horizontal lines below are my favourite takeaways after I went through Thomas’ transcript.

---

Buffett thinks his idea on import certificates is different from tariffs and that it’s important to have more balanced trade between countries; he also thinks that trade should not be wielded as a weapon, and that the more prosperous the world becomes, the better the USA would be

Becky Quick: Thanks Warren. This first question comes from Bill Mitchell. I received more questions about this than any other question. He writes, “Warren, in a 2003 Fortune article, you argued for import certificates to limit trade deficits and said these import certificates basically amounted to a tariff, but recently you called tariffs an act of economic war. Has your view on trade barriers changed or do you see import certificates as somehow distinct from tariffs?”

Warren Buffett: Well, the import certificates were distinct, but their goal was to balance imports against exports so that the trade deficit would not grow in an enormous way. It had various provisions to help third world countries catch up a little bit. They were designed to balance trade, and I think you can make very good arguments that balanced trade is good for the world. It makes sense for cocoa to be raised in Ghana and coffee in Colombia and a few other things…

…There’s no question that trade can be an act of war, and I think it’s led to bad things like the attitudes it’s brought out in the United States. We should be looking to trade with the rest of the world. We should do what we do best, and they should do what they do best…

…The main thing is that trade should not be a weapon. The United States has become an incredibly important country starting from nothing 250 years ago – there’s never been anything like it. And it’s a big mistake when you have 7.5 billion people who don’t like you very well and you have 300 million people crowing about how well they’ve done. I don’t think it’s right and I don’t think it’s wise. The more prosperous the rest of the world becomes, it won’t be at our expense – the more prosperous we’ll become and the safer we’ll feel and your children will feel someday.

Buffett did not look at macroeconomic factors in Japan when making the huge investments he did in five Japanese trading houses; Berkshire won’t be selling the Japanese investments for a long, long time, if at all; Berkshire would be happy to invest a lot more in Japan if there was capacity to do so; the fact that Berkshire could borrow in Japanese Yen to hedge the Japanese investments’ currency risk is merely a lucky coincidence

Question: Mr. Buffett and Mr. Munger did a very good and successful investment in Japan in the past five or six years. The recent CPI in Japan is currently above 3%, not far away from its 2% target. Bank of Japan seems very determined in raising rates while Fed, ECB, and other central banks are considering cutting them. Do you think BOJ makes sense to proceed with the rate hike? Will its planned rate hike deter you from further investing in the Japanese stock market or even considering realizing your current profits?

Warren Buffett: Well, I’m going to extend the same goodwill to Japan that you’ve just extended to me. I’ll let the people of Japan determine their best course of action in terms of economics. It’s an incredible story. It’s been about six years now since our Japanese investments. I was just going through a little handbook that probably had two or three thousand Japanese companies in it. One problem I have is that I can’t read that handbook anymore – the print’s too small. But there were these five trading companies selling at ridiculously low prices. So I spent about a year acquiring them. And then we got to know the people better, and everything that Greg and I saw, we liked better as we went along…

…Greg Abel: When you think of the five companies, there’s definitely a couple meetings a year, Warren. The thing we’re building with the five companies is, one, it’s been a very good investment, but we really envision holding the investment for 50 years or forever…

…Warren Bufett: We will not be selling any stock. That will not happen in decades, if then…

…It’s too bad that Berkshire has gotten as big as it is because we love that position and I’d like it to be a lot larger. Even with the five companies being very large in Japan, we’ve got at market in the range of $20 billion invested, but I’d rather have $100 billion than $20 billion…

…The Japanese situation is different because we intend to stay so long with that position and the funding situation is so cheap that we’ve attempted to some degree to match purchases against yen-denominated funding. But that’s not a policy of ours…

…Greg Abel: There’s no question we were fundamentally very comfortable with investing in the five Japanese companies and recognizing we’re investing in yen. The fact we could then borrow in yen was almost just a nice incremental opportunity. But we were very comfortable both with the Japanese companies and with the currency we would ultimately realize in yen.

Just the simple act of reading about companies can lead to great investment opportunities

Warren Buffett: It’s been about six years now since our Japanese investments. I was just going through a little handbook that probably had two or three thousand Japanese companies in it…

…I never dreamt of that when I picked up that handbook. It’s amazing what you can find when you just turn the page. We showed a movie last year about “turn every page,” and I would say that turning every page is one important ingredient to bring to the investment field. Very few people do turn every page, and the ones who turn every page aren’t going to tell you what they’re finding. So you’ve got to do a little of it yourself.

Berkshire’s current huge cash position is the result of Buffett not being able to find sufficiently attractive investment opportunities; Buffett thinks that great investment opportunities appear infrequently

Becky Quick: This next question comes from Advate Prasad in New York. He writes, “Today, Berkshire holds over $300 billion in cash and short-term investments, representing about 27% of total assets, a historically high figure compared to the 13% average over the last 25 years. This has also led Berkshire to effectively own nearly 5% of the entire US Treasury market. Beyond the need for liquidity to meet insurance obligations, is the decision to raise cash primarily a de-risking strategy in response to high market valuations?…

…Warren Buffett: Well, I wouldn’t do anything nearly so noble as to withhold investing myself just so that Greg could look good later on. If he gets any edge of what I leave behind, I’ll resent it. The amount of cash we have – we would spend $100 billion if something is offered that makes sense to us, that we understand, offers good value, and where we don’t worry about losing money. The problem with the investment business is that things don’t come along in an orderly fashion, and they never will. I’ve had about 16,000 trading days in my career. It would be nice if every day you got four opportunities or something like that with equal attractiveness. If I was running a numbers racket, every day would have the same expectancy that I would keep 40% of whatever the handle was, and the only question would be how much we transacted. But we’re not running that kind of business. We’re running a business which is very opportunistic.

Investing in stocks is a much better investment-bet than investing in real estate

Warren Buffett: Well, in respect to real estate, it’s so much harder than stocks in terms of negotiation of deals, time spent, and the involvement of multiple parties in the ownership. Usually when real estate gets in trouble, you find out you’re dealing with more than just the equity holder. There have been times when large amounts of real estate have changed hands at bargain prices, but usually stocks were cheaper and they were a lot easier to do.

Charlie did more real estate. Charlie enjoyed real estate transactions, and he actually did a fair number of them in the last 5 years of his life. But he was playing a game that was interesting to him. I think if you’d asked him to make a choice when he was 21 – either be in stocks exclusively for the rest of his life or real estate for the rest of his life – he would have chosen stocks. There’s just so much more opportunity, at least in the United States, that presents itself in the security market than in real estate…

…When you walk down to the New York Stock Exchange, you can do billions of dollars worth of business, totally anonymous, and you can do it in 5 minutes. The trades are complete when they’re complete. In real estate, when you make a deal with a distressed lender, when you sign the deal, that’s just the beginning. Then people start negotiating more things, and it’s a whole different game with a different type of person who enjoys the game.

Berkshire’s leaders think AI will have a massive impact on the insurance business, but they are not in a hurry to pour money into AI as they think there’s plenty of faddish capital in the space

Ajit Jain: There is no question in my mind that AI is going to be a real game-changer. It’s going to change the way we assess risk, we price risk, we sell risk, and then the way we end up paying claims. Having said that, I certainly also feel that people end up spending enormous amounts of money trying to chase the next fashionable thing…

…Right now the individual insurance operations do dabble in AI and try to figure out the best way to exploit it. But we have not yet made a conscious big-time effort in terms of pouring a lot of money into this opportunity.

Buffett prefers Ajit Jain to any kind of sophisticated AI systems when pricing insurance risks

Warren Buffett: I wouldn’t trade everything that’s developed in AI in the next 10 years for Ajit. If you gave me a choice of having a hundred billion dollars available to participate in the property casualty insurance business for the next 10 years and a choice of getting the top AI product from whoever’s developing it or having Ajit making the decisions, I would take Ajit anytime – and I’m not kidding about that.

Despite the political upheaval happening in the USA right now, Buffett still thinks the long-term future of the country is incredibly bright; in Buffett’s eyes, the USA has been through plenty of tumultuous periods and emerged stronger

Warren Buffett: America has been undergoing significant and revolutionary change ever since it was developed. I mentioned that we started out as an agricultural society with high promises that we didn’t deliver on very well. We said all men were created equal, and then we wrote a constitution that counted blacks as three-fifths of a person. In Article 2, you’ll find male pronouns used 20 times and no female pronouns. So it took until 1920, with the 19th amendment, to finally give women the vote that we had promised back in 1776.

We’re always in the process of change, and we’ll always find all kinds of things to criticize in the country. But the luckiest day in my life is the day I was born, because I was born in the United States. At that time, about 3% of all births in the world were taking place in the United States. I was just lucky, and I was lucky to be born white, among other things…

…We’ve gone through all kinds of things – great recessions, world wars, the development of the atomic bomb that we never dreamt of when I was born. So I would not get discouraged about the fact that we haven’t solved every problem that’s come along. If I were being born today, I would just keep negotiating in the womb until they said I could be in the United States.

It’s important to be patient while waiting for opportunities, but equally important to pounce when the opportunity appears

Warren Buffett: The trick when you get in business with somebody who wants to sell you something for $6 million that’s got $2 million of cash, a couple million of real estate, and is making $2 million a year, is you don’t want to be patient at that moment. You want to be patient in waiting to get the occasional call. My phone will ring sometime with something that wakes me up. You just never know when it’ll happen. That’s what makes it fun. So patience is a combination of patience and a willingness to do something that afternoon if it comes to you.

It does not pay to invest in a way that depends on the appearance of a greater fool

Warren Buffett: If people are making more money because they’re borrowing money or participating in securities that are pieces of junk but they hope to find a bigger sucker later on, you have to forget that.

Buffett does not think it’s important to manage currency risk with Berkshire’s international investments, but he avoids investments denominated in currencies that are at risk of depreciating wildly

Warren Buffett: We’ve owned lots of securities in foreign currencies. We do nothing in terms of its impact on quarterly and annual earnings. We don’t do anything based on its impact on quarterly and annual earnings. There’s never been a board meeting I can remember where I’ve said, “If we do this, our annual earnings will be this, therefore we ought to do it.” The number will turn out to be what it’ll be. What counts is where we are five or 10 or 20 years from now…

…Obviously, we wouldn’t want to own anything in a currency that we thought was really going to hell.

Buffett is worried about the tendency for governments to want to devalue their currencies, the USA included, but there’s nothing much that can be done about it; Buffett thinks the USA is running a fiscal deficit that is unsustainable over a long period of time; Buffett thinks a 3% fiscal deficit appears sustainable

Warren Buffett: That’s the big thing we worry about with the United States currency. The tendency of a government to want to debase its currency over time – there’s no system that beats that. You can pick dictators, you can pick representatives, you can do anything, but there will be a push toward weaker currencies. I mentioned very briefly in the annual report that fiscal policy is what scares me in the United States because of the way it’s made, and all the motivations are toward doing things that can cause trouble with money. But that’s not limited to the United States – it’s all over the world, and in some places, it gets out of control regularly. They devalue at rates that are breathtaking, and that’s continued…

…So currency value is a scary thing, and we don’t have any great system for beating that…

…We’re operating at a fiscal deficit now that is unsustainable over a very long period of time. We don’t know whether that means two years or 20 years because there’s never been a country like the United States. But as Herbert Stein, the famous economist, said, “If something can’t go on forever, it will end.” We are doing something that is unsustainable, and it has the aspect to it that it gets uncontrollable to a certain point….

…I wouldn’t want the job of trying to correct what’s going on in revenue and expenditures of the United States with roughly a 7% gap when probably a 3% gap is sustainable…

…We’ve got a lot of problems always as a country, but this is one we bring on ourselves. We have a revenue stream, a capital-producing stream, a brains-producing machine like the world has never seen. And if you picked a way to screw it up, it would involve the currency. That’s happened a lot of places.

Buffett thinks the key factors for a developing economy to attract investors are having a solid currency, and being business-friendly

Audience member: What advice would you give to government and business leaders of emerging markets like Mongolia to attract institutional investors like yourself?

Warren Buffett: If you’re looking for advice to give the government over there, it’s to develop a reputation for having a solid currency over time. We don’t really want to go into any country where we think there’s a significant probability of runaway inflation. That’s too hard to figure…

…If the country develops a reputation for being business-friendly and currency-conscious, that bodes very well for the residents of that country, particularly if it has some natural assets that it can build around.

Private equity firms are flooding the life insurance market, but they are doing so by taking on lots of leverage and credit risk

Ajit Jain: There’s no question the private equity firms have come into the space, and we are no longer competitive in the space. We used to do a fair amount in this space, but in the last 3-4 years, I don’t think we’ve done a single deal.

You should separate this whole segment into two parts: the property casualty end of the business and the life end of the business. The private equity firms you mentioned are all very active in the life end of the business, not the property casualty end.

You are right in identifying the risks these private equity firms are taking on both in terms of leverage and credit risk. While the economy is doing great and credit spreads are low, these firms have taken the assets from very conservative investments to ones where they get a lot more return. As long as the economy is good and credit spreads are low, they will make money – they’ll make a lot of money because of leverage.

However, there is always the danger that at some point the regulators might get cranky and say they’re taking too much risk on behalf of their policyholders, and that could end in tears. We do not like the risk-reward that these situations offer, and therefore we put up the white flag and said we can’t compete in this segment right now.

Buffett thinks Berkshire’s insurance operation is effectively unreplicable

Warren Buffett: I think there are people that want to copy Berkshire’s model, but usually they don’t want to copy it by also copying the model of the CEO having all of his money in the company forever. They have a different equation – they’re interested in something else. That’s capitalism, but they have a whole different situation and probably a somewhat different fiduciary feeling about what they’re doing. Sometimes it works and sometimes it doesn’t work. If it doesn’t work, they go on to other things. If what we do at Berkshire doesn’t work, I spend the end of my life regretting what I’ve created. So it’s just a whole different personal equation.

There is no property casualty company that can basically replicate Berkshire. That wasn’t the case at the start – at the start we just had National Indemnity a few miles from here, and anybody could have duplicated what we had. But that was before Ajit came with us in 1986, and at that point the other fellows should have given up.

Buffett thinks recent market volatility is not noteworthy at all; it’s nearly certain that significant downward moves in stocks will happen sometime in the next 20 years

Warren Buffett: What has happened in the last 30-45 days, 100 days, whatever this period has been, is really nothing. There have been three times since we acquired Berkshire that Berkshire has gone down 50% in a fairly short period of time – three different times. Nothing was fundamentally wrong with the company at any time. This is not a huge move. The Dow Jones average was at 381 in September of 1929 and got down to 42. That’s going from 100 to 11. This has not been a dramatic bear market or anything of the sort. I’ve had about 17,000 or 18,000 trading days. There have been plenty of periods that are dramatically different than this…

…You will see a period in the next 20 years that will be a “hair curler” compared to anything you’ve seen before. That just happens periodically. The world makes big mistakes, and surprises happen in dramatic ways. The more sophisticated the system gets, the more the surprises can come out of left field. That’s part of the stock market, and that’s what makes it a good place to focus your efforts if you’ve got the proper temperament for it and a terrible place to get involved if you get frightened by markets that decline and get excited when stock markets go up.

Berkshire’s leaders think the biggest change autonomous vehicles will bring to the automotive insurance industry is substitution of operator error policies by product liability policies; Berkshire’s leaders also think that the cost per repair in the event of an accident will rise significantly; the total cost of providing insurance for autonomous vehicles is still unclear; from the 1950s to today, cars have gotten 6x safer but auto insurance has become 50x pricier

Ajit Jain: There’s no question that insurance for automobiles is going to change dramatically once self-driving cars become a reality. The big change will be what you identified. Most of the insurance that is sold and bought revolves around operator errors – how often they happen, how severe they are, and therefore what premium we ought to charge. To the extent these new self-driving cars are safer and involved in fewer accidents, that insurance will be less required. Instead, it’ll be substituted by product liability. So we at GEICO and elsewhere are certainly trying to get ready for that switch, where we move from providing insurance for operator errors to being more ready to provide protection for product errors and errors and omissions in the construction of these automobiles…

…We talked about the shift to product liability and protection for accidents that take place because of an error in product design or supply. In addition to that shift, I think what we’ll see is a major shift where the number of accidents will drop dramatically because of automatic driving. But on the other hand, the cost per repair every time there’s an accident will go up very significantly because of the amount of technology in the car. How those two variables interact with each other in terms of the total cost of providing insurance, I think, is still an open issue…

…Warren Buffett: When I walked into GEICO’s office in 1951, the average price of a policy was around $40 a year. Now it’s easy to get up to $2,000 depending on location and other factors. During that same time, the number of people killed in auto accidents has fallen from roughly six per 100 million miles driven to a little over one. So the car has become incredibly safer, and it costs 50 times as much to buy an insurance policy.

There’s a tax now when American companies conduct share buybacks

Warren Buffett: I don’t think people generally know that, but there is a tax that was introduced a year or so ago where we pay 1%. That not only hurts us because we pay more for it than you do – it’s a better deal for you than for us – but it actually hurts some of our investee companies quite substantially. Tim Cook has done a wonderful job running Apple, but he spent about $100 billion in a year repurchasing shares, and there’s a 1% charge attached to that now. So that’s a billion dollars a year that he pays when he buys Apple stock compared to what you pay.

Buffett is very careful with the risks that come with derivative contracts on a company’s balance sheet

Greg Abel: I’ll maybe go back to the very first meeting with Warren because it still stands out in my mind. Warren was thinking about acquiring Mid-America Energy Holdings Company at that time, and we had the opportunity with my partners to go over there on a Saturday morning. We were discussing the business and Warren had the financial statements in front of him. Like anybody, I was sort of expecting a few questions on how the business was performing, but Warren locked in immediately to what was on the balance sheet and the fact we had some derivative contracts, the “weapons of mass destruction.”

In the utility business, we do have derivatives because they’re used to match certain positions. They’re never matched perfectly, but we have them and they’re required in the regulated business. I remember Warren going to it immediately and asking about the composition and what was the underlying risk, wanting to thoroughly understand. It wasn’t that big of a position, but it was absolutely one of the risks he was concerned about as he was acquiring Mid-America, especially in light of Enron and everything that had gone on.

The followup to that was a year or 18 months later. There was an energy crisis in the US around electricity and natural gas, and various companies were making significant sums of money. Warren’s follow-up question to me was, “How much money are we making during this energy crisis? Are we making a lot? Do we have speculative positions in place?” The answer was we weren’t making any more than we would have been six months ago because all those derivatives were truly to support our business and weren’t speculative. That focus on understanding the business and the risks around it still stands out in my mind.

Buffett spends more time analysing a company’s balance sheet than other financial statements

Warren Buffett: I spend more time looking at balance sheets than I do income statements. Wall Street doesn’t pay much attention to balance sheets, but I like to look at balance sheets over an 8 or 10 year period before I even look at the income account because there are certain things it’s harder to hide or play games with on the balance sheet than with the income statement.

Buffett thinks America’s electric grid needs a massive overhaul and it can only be done in via a partnership between the private sector and the government – unfortunately, nobody has figured out the partnership model yet

Warren Buffett: t’s very obvious that the country needs an incredible improvement, rethinking, redirection to some extent in the electric grid. We’ve outgrown what would be the model that America should have. In a sense, it’s a problem something akin to the interstate highway system where you needed the power of the government really to get things done because it doesn’t work so well when you get 48 or 50 jurisdictions that each has their own way of thinking about things…

…There are certain really major investment situations where we have capital like nobody else has in the private system. We have particular knowhow in the whole generation and transmission arena. The country is going to need it. But we have to figure out a way that makes sense from the standpoint of the government, from the standpoint of the public, and from the standpoint of Berkshire, and we haven’t figured that out yet. It’s a clear and present use of hundreds of billions of dollars. You have people that set up funds and they’re getting paid for just assembling stuff, but that’s not the way to handle it. The way to handle it is to have some kind of government-private industry cooperation similar to what you do in a war.

The risk of wildfires to electric utilities is not going to go away, and in fact, will increase over time

Greg Abel: The reality is the risk around wildfires – do the wildfires occur – they’re not going away, and we know that. The risk probably goes up each year.

Berkshire’s leaders think it’s important for utilities to de-energise when wildfires occur to minimise societal damage; Berkshire is the only utility operator so far that’s willing to de-energise; but de-energising also has its drawbacks; Berkshire may not be able to solve the conundrum of de-energising

Greg Abel: the one thing we hadn’t tackled – this is very relevant to the significant event we had back in 2020 in PacifiCorp – is we didn’t de-energize the system as the fire was approaching. Our employees and the whole management team have been trained all their lives to keep the lights on, and the last thing they want to do is turn those lights off and have a system de-energized. After those events and as we looked at how we’re going to move forward in managing the assets and reducing risk, we recognized as a team that we have to de-energize those assets. Now as we get fires encroaching at a certain number of miles, we de-energize because we do not want to contribute to the fire nor harm any of our consumers or contribute to a death. We had to take our team to managing a different risk now. It’s not around keeping the lights on, it’s around protecting the general public and ensuring the fire does not spread further. We’re probably the one utility or across our utilities that does that today, and we strongly believe in that approach.

Becky Quick: Doesn’t that open you up to other risk if you shut down your system, a hospital gets shut down, somebody dies?

Greg Abel: That’s something we do deal with a lot because we have power outages that occur by accident. When we look at critical infrastructure, that’s an excellent point and we’re constantly re-evaluating it. We do receive a lot of feedback from our customer groups as to how to manage that…

…Warren Buffett: There’s some problems that can’t be solved, and we shouldn’t be in the business of taking investors’ money and tackling things that we don’t know the solution for. You can present the arguments, but it’s a political decision when you are dealing with states or the federal government. If you’re in something where you’re going to lose, the big thing to do is quit.

Buffett thinks the value of electric utility companies have fallen a lot over the past two years because of societal trends and his enthusiasm for investing in electric utilities has waned considerably

Becky Quick: Ricardo Bri, a longtime shareholder based in Panama, says that he was very happy to see Berkshire acquire 100% of BHE. It was done in two steps: one in late 2022 – 1% was purchased from Greg Abel for $870 million implying a valuation of BHE of $87 billion, and then in 2024 the remaining 8% was purchased from the family of Walter Scott Jr. for $3.9 billion implying a valuation of $48.8 billion for the enterprise. That second larger transaction represented a 44% reduction in valuation in just two years. Ricardo writes that PacifiCorp liabilities seem too small to explain this. Therefore, what factors contributed to the difference in value for BHE between those two moments in time?

Warren Buffett: Well, we don’t know how much we’ll lose out of PacifiCorp and decisions that are made, but we also know that certain of the attitudes demonstrated by that particular example have analogues throughout the utility system. There are a lot of states that so far have been very good to operate in, and there are some now that are rat poison, as Charlie would say, to operate in. That knowledge was accentuated when we saw what happened in the Pacific Northwest, and it’s eventuated by what we’ve seen as to how utilities have been treated in certain other situations. So it wasn’t just a direct question of what was involved at PacifiCorp. It was an extrapolation of a societal trend…

…We’re not in the mood to sell any business. But Berkshire Hathaway Energy is worth considerably less money than it was two years ago based on societal factors. And that happens in some of our businesses. It certainly happened to our textile business. The public utility business is not as good a business as it was a couple of years ago. If anybody doesn’t believe that, they can look at Hawaiian Electric and look at Edison in the current wildfires situation in California. There are societal trends that are changing things…

…I would say that our enthusiasm for buying public utility companies is different now than it would have been a couple years ago. That happens in other industries, too, but it’s pretty dramatic in public utilities. And it’s particularly dramatic in public utilities because they are going to need lots of money. So, if you’re going to need lots of money, you probably ought to behave in a way that encourages people to give you lots of money.

Buffett thinks the future capital intensity of the USA’s large technology companies remains to be seen

Warren Buffett: It’ll be interesting to see how much capital intensity there is now with the Magnificent 7 compared to a few years ago. Basically, Apple has not really needed any capital over the years and it’s repurchased shares with a dramatic reduction. Whether that world is the same in the future or not is something yet to be seen.

Buffett thinks there’s no better system than capitalism that has been discovered so far

Warren Buffett: Capitalism in the United States has succeeded like nothing you’ve ever seen. But what it is is a combination of this magnificent cathedral which has produced an economy like nothing the world’s ever seen, and then it’s got this massive casino attached…

…In the cathedral, they’re designing things that will be producing goods and services for 300 and some million people like it’s never been done before in history. It’s an interesting system we developed, but it’s worked. It dispenses rewards in what seems like a terribly capricious manner. The idea that people get what they deserve in life – it’s hard to make that argument. But if you argue with it that any other system works better, the answer is we haven’t found one.

---

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 Alphabet, Amazon, Apple, Meta Platforms, Microsoft, and Tesla (they are all part of the Magnificent 7). Holdings are subject to change at any time.

The book mentioned in the title of this article is The Other Half of Macroeconomics and the Fate of Globalization written by economist Richard C. Koo (Gu Chao Ming) and published in 2018.

I first came across Koo’s book in March 2020 when I chanced upon a review of it in Mandarin, written by investor Li Lu. I can read Mandarin and I found myself agreeing to the ideas from the book that Li shared, so much so that I made a self-directed attempt at translating the review into English. But I only began reading the actual book near the start of this year and finished it about a month ago. There was even more richness in the book’s ideas about how economies should operate than what was shared in Li’s already-wonderful review.

Earlier this month, the US government, under the Trump administration, made sweeping changes to the global trading system by introducing the Reciprocal Tariff Policy, which raised tariffs, sometimes significantly so, for many of the US’s trading partners. Major driving forces behind the Reciprocal Tariff Policy ostensibly include the US’s sustained trade deficits (particularly with China) and a desire by the Trump administration to bring manufacturing jobs back to the country.

As I contemplated the Trump administration’s actions, and the Chinese government’s reactions, I realised Koo’s book explained why all these issues happened. So I’m writing my own notes and takeaways from the book for easy reference in the future and I would like to share them in this article in the hopes that they could be useful for you. I will be borrowing from my translation of Li’s Mandarin review in this article. Below the horizontal line, all content in grey font are excerpts from my translation while all italicised content are excerpts from the book. 

---

The three stages of economic development a country goes through

There are six important ideas that Koo discussed in his book. One of them is the concept that a country goes through three distinct stages of economic development over time. 

The first stage of development would be a country that is industrialising and has yet to reach the Lewis Turning Point (LTP). The LTP is the “point at which urban factories have finally absorbed all the surplus rural labour.” When a country starts industrialising, people are mostly living in rural areas and there are only a very few educated elite who have the knowhow to kickstart industrialisation. There is also a surplus of labour. As a result, the educated elite – the industrialists – hold the power and “most of the gains during the initial stage of industrialisation therefore go to the educated few.”  The first stage of economic development is also when income inequality widens – the gains from industrialisation continue to accumulate in the hands of the elite as they reinvest profits into their businesses because there continues to be a surplus of labour.

The second stage of development happens when an industrialising economy reaches the LTP. At this point, labour “gains the bargaining power to demand higher wages for the first time in history, which reduces the share of output accruing to business owners.” But business owners are happy to continue reinvesting their profits as they are still “achieving good returns, leading to further tightness in the labour market.” This dynamic leads to an economy’s “golden era”:

“As labor’s share increases, consumption’s share of GDP will increase at the expense of investment. At the same time, the explosive increase in the purchasing power of ordinary citizens means that most businesses are able to increase profits simply by expanding existing productive capacity. Consequently, both consumption and investment will increase rapidly…

…Inequality also diminishes as workers’ share of output increases relative to that of capital… 

…With incomes rising and inequality falling, this post-LTP maturing phase may be called the golden era of economic growth…

…Higher wages force businesses to look harder for profitable investment opportunities. On the other hand, the explosive increase in the purchasing power of ordinary workers who are paid ever-higher wages creates major investment opportunities. This prompts businesses to invest for two reasons. First, they seek to increase worker productivity so that they can pay ever-higher wages. Second, they want to expand capacity to address workers’ increasing purchasing power. Both productivity- and capacity-enhancing investments increase demand for labor and capital that add to economic growth. In this phase, business investment increases workers’ productivity even if their skill level remains unchanged…

…With rapid improvements in the living standards of most workers, the post-LTP maturing phase is characterised by broadly distributed benefits from economic growth.”

The golden era has its problems too. This is because this period is when “workers begin to utilise their newfound bargaining power” such as by organising strikes. But business owners and labour tend to be able to work out their differences.

The third stage of development is what Koo calls a “post-LTP pursued economy.” When a country is in the golden era, at some point ever-growing wages creates inroads for foreign competitors – and the country starts being chased by the foreign competitors that have lower wages. This is when businesses in the country find it very challenging to “find attractive investment opportunities at home because it often makes more sense for them to buy directly from the “chaser” or to invest in that country themselves.” This is also when “the return on capital is higher abroad than at home.” During the pursued stage, “real wage growth will be minimal” and “economic growth also slows.” Although a pursued country can continue to grow economically, a major problem is that inequality once again rears its ugly head:

“Japan’s emergence in the 1970s shook the U.S. and European industrial establishments. As manufacturing workers lost their jobs, ugly trade frictions ensued between Japan and the West. This marked the first time that Western countries that had already passed their LTPs had been chased by a country with much lower wages…

…While Western companies at the forefront of technology continued to do well, the disappearance of many well-paying manufacturing jobs led to worsening income inequality in these countries…

…Some of the pain Western workers felt was naturally offset by the fact that, as consumers, they benefited from cheaper imports from Asia, which is one characteristic of import-led globalisation. Businesses with advanced technology continued to do well, but it was no longer the case that everyone in society was benefiting from economic growth. Those whose jobs could be transferred to lower-cost locations abroad saw their living standards stagnate or even fall.”

Koo wrote that Western economies – the USA and Europe – entered their golden eras around the 1950s and became pursued starting in the 1970s by Japan. During the golden era of the West, “it was in an export-led globalisation phase as it exported consumer and capital goods to the world.” But as the West started getting pursued, they entered “an import-led globalisation phase as capital seeks higher returns abroad and imports flood the domestic market.”

The four states of an economy

Another important idea from Koo’s book is the concept that an economy has four distinct states, which are summarised in Table 1 below:

“An economy is always in one of four possible states depending on the presence or absence of lenders (savers) and borrowers (investors). They are as follows: (1) both lenders and borrowers are present in sufficient numbers, (2) there are borrowers but not enough lenders even at high interest rates, (3) there are lenders but not enough borrowers even at low interest rates, and (4) both lenders and borrowers are absent.”

Koo’s idea that an economy has four distinct states is important because mainstream economic-thought does not cater for the disappearance of borrowers:

“Of the four, only Cases 1 and 2 are discussed in traditional economics, which implicitly assumes there are always enough borrowers as long as real interest rates are low enough.”

There are two key reasons why an economy would be in Cases 3 and 4, i.e. when borrowers disappear. The first is when private-sector businesses are unable to find attractive investment opportunities (this is related to economies that are in the third-stage of development discussed earlier in this article, when attractive domestic investment opportunities become scarce):

“The first is one in which private‐sector businesses cannot find investment opportunities that will pay for themselves. The private sector will only borrow money if it believes it can pay back the debt with interest. And there is no guarantee that such opportunities will always be available. Indeed, the emergence of such opportunities depends very much on scientific discoveries and technological innovations, both of which are highly irregular and difficult to predict.

In open economies, businesses may also find that overseas investment opportunities are more attractive than those available at home. If the return on capital is higher in emerging markets, for example, pressure from shareholders will force businesses to invest more abroad while reducing borrowings and investments at home. In modern globalized economies, this pressure from shareholders to invest where the return on capital is highest may play a greater role than any technological breakthroughs, or lack thereof, in the decision as to whether to borrow and invest at home.”

The second reason for the disappearance of borrowers is named by Koo as a “balance sheet recession” which is described as such:

“In the second set of circumstances, private‐sector borrowers have sustained huge losses and are forced to rebuild savings or pay down debt to restore their financial health. Such a situation may arise following the collapse of a nationwide asset price bubble in which a substantial part of the private sector participated with borrowed money. The collapse of the bubble leaves borrowers with huge liabilities but no assets to show for the debt. Facing a huge debt overhang, these borrowers have no choice but to pay down debt or increase savings in order to restore their balance sheets, regardless of the level of interest rates.

Even when the economy is doing well, there will always be businesses that experience financial difficulties or go bankrupt because of poor business decisions. But the number of such businesses explodes after a nationwide asset bubble bursts.

For businesses, negative equity or insolvency implies the potential loss of access to all forms of financing, including trade credit. In the worst case, all transactions must be settled in cash, since no supplier or creditor wants to extend credit to an entity that may seek bankruptcy protection at any time. Many banks and other depository institutions are also prohibited by government regulations from extending or rolling over loans to insolvent borrowers in order to safeguard depositors’ money. For households, negative equity means savings they thought they had for retirement or a rainy day are no longer there. Both businesses and households will respond to these life‐threatening conditions by focusing on restoring their financial health—regardless of the level of interest rates—until their survival is no longer at stake.”

A balance sheet recession can be a huge problem for a country’s economy if it is unresolved as it can lead to a rapidly shrinking economy as a manifestation of the “fallacy of composition” problem:

“One person’s expenditure is another person’s income…

…The interaction between thinking and reacting households and businesses create a situation where one plus one does not necessarily equal two. For example, if A decides to buy less from B in order to set aside more savings for an uncertain future, B will have less income to buy things from A. That will lower A’s income, which in turn will reduce the amount A can save.

This interaction between expenditure and income also means that, at the national level, if one group is saving money, another group must be doing the opposite – “dis-saving” – to keep the economy running. In most cases, this dis-saving takes the form of borrowing by businesses that seek to expand their operations. If everyone is saving and no one is dis-saving on borrowing, all of those savings will leak out of the economy’s income stream, resulting in less income for all.

For example, if a person with an income of $1,000 decides to spend $900 and save $100, the $900 that is spent becomes someone else’s income and continues circulating in the economy. The $100 that is saved is typically deposited with a financial institution such as a bank, which then lends it to someone else who can make use of it. When that person borrows and spends the $100, total expenditures in the economy amount to $900 plus $100, which is equal to the original $1,000, and the economy moves forward…

…If there are no borrowers for $100 in savings in the above example, even at zero interest rates, total expenditures in the economy will drop to $900, while the saved $100 remains unborrowed in financial institutions or under mattresses. The economy has effectively shrunk by 10 percent, from $1,000 to $900. That $900 now becomes someone else’s income. If that person decides to save 10 percent, and there are still no borrowers, only $810 will be spent, causing the economy to contract to $810. This cycle will repeat, and the economy will shrink to $730, if borrowers remain on the sidelines. This process of contraction is called a “deflationary spiral.”…

…Keynes had a name for this state of affairs, in which everyone wants to save but is unable to do so because no one is borrowing. He called it the paradox of thrift. It is a paradox because if everyone tries to save, the net result is that no one can save.

The phenomenon of right behaviour at the individual level leading to a bad result collectively is known as the “fallacy of composition.””

Japan was the “first advanced country to experience a private-sector shift to debt minimization for balance sheet reasons since the Great Depression.” Japan’s real estate bubble burst in 1990, where real estate prices fell by 87%, “devastating the balance sheets of businesses and financial institutions across the country” and leading to a disappearance of borrowers:

“Demand for funds shrank rapidly when the bubble finally burst in 1990. Noting that the economy was also slowing sharply, the BOJ took interest rates down from 8 percent at the height of the bubble to almost zero by 1995. But demand for funds not only failed to recover but actually turned negative that year. Negative demand for funds means that Japan’s entire corporate sector was paying down debt at a time of zero interest rates, a world that no economics department in university or business school had ever envisioned. The borrowers not only stopped borrowing but began moving in the opposite direction by paying down debt and continued doing so for a full ten years, until around 2005…

…While in a textbook economy the household sector saves and the corporate sector borrows, both sectors became net-savers in post-1999 Japan, with the corporate sector becoming the largest saver in the country from 2002 onward in spite of zero interest rates.”

The Western economies experienced their own balance sheet recessions starting in 2008 with the bursting of housing bubbles in that year. When the “bubbles collapsed on both sides of the Atlantic in 2008, the balance sheets of millions of households and many financial institutions were devastated.” Borrowers also disappeared; the “private sectors in virtually all major advanced nations have been increasing savings or paying down debt since 2008 in spite of record low interest rates.” For example, “the U.S. private sector saved 4.1 percent of GDP at near-zero interest rates in the four quarters through Q1 2017” and the “Eurozone’s overall private sector is saving 4.6 percent of GDP in spite of negative interest rates.”

Prior to the Japanese episode, the most recent example was the Great Depression:

“Until 2008, the economics profession considered a contractionary equilibrium (the $500 economy) brought about by a lack of borrowers to be an exceptionally rare occurrence – the only recent example was the Great Depression, which was triggered by the stock market crash in October 1929 and during which the U.S. lost 46 percent of nominal GNP. Although Japan fell into a similar predicament when its asset price bubble burst in 1990, its lessons were almost completely ignored by the economics professional until the Lehman shock of 2008.”

The appropriate macroeconomic policies

The third important idea from Koo’s book is that depending on which stage (pre-LTP, golden era, or pursued) and which state (Case 1, Case 2, Case 3, or Case 4) a country’s economy is in, there are different macroeconomic policies that would be appropriate.

First, it’s important to differentiate the two policies governments can wield, namely, monetary policy and fiscal policy:

“The government also has two types of policy, known as monetary and fiscal policy, that it can use to help stabilise the economy by matching private-sector savings and borrowings. Themore frequently used is monetary policy, which involves raising or lowering interest rates to assist the matching process. Since an excess of borrowers is usually associated with a strong economy, a higher policy rate might be appropriate to prevent overheating and inflation. Similarly, a shortage of borrowers is usually associated with a weak economy, in which case a lower policy rate might be needed to avert a recession or deflation.

With fiscal policy, the government itself borrows and spends money on such projects as highways, airports, and other social infrastructure. While monetary policy decisions can be made very quickly by the central bank governor and his or her associates, fiscal policy tends to be very cumbersome in a peacetime democracy because elected representatives must come to an agreement on how much to borrow and where to spend the money. Because of the political nature of these decisions and the time it takes to implement them, most recent economic fluctuations were dealt with by central banks using monetary policy.”

Fiscal policy is more important than monetary policy when a country’s economy is in the pre-LTP stage, but the relative importance of the two types of policies switches once the economy enters the golden era; fiscal policy once again becomes the more important type of policy when the economy is in the pursued stage: 

“In the early phases of industrialisation, economic growth will rely heavily on manufacturing, exports, and the formation of capital etc. At this juncture, the government’s fiscal policies can play a huge role. Through fiscal policies, the government can gather scarce resources and invest them into basic infrastructure, resources, and export-related services etc. These help emerging countries to industrialise rapidly. Nearly every country that was in this stage of development saw their governments implement policies that promote active governmental support.

In the second stage of development, the twin engines of economic growth are rising wages and consumer spending. The economy is already in a state of full employment, so an increase in wages in any sector or field will inevitably lead to higher wages in other areas. Rising wages lead to higher spending and savings, and companies will use these savings to invest in productivity to improve output. In turn, profits will grow, leading to companies having an even stronger ability to raise wages to attract labour. All these combine to create a positive feedback loop of economic growth. Such growth comes mainly from internal sources in the domestic economy. Entrepreneurs, personal and household investing behaviour, and consumer spending patterns are the decisive players in promoting economic growth, since they are able to nimbly grasp business opportunities in the shifting economic landscape. Monetary policies are the most effective tool in this phase, compared to fiscal policies, for a few reasons. First, fiscal policies and private-sector investing both tap on a finite pool of savings. Second, conflicts could arise between the private sector’s investing activities and the government’s if poorly thought-out fiscal policies are implemented, leading to unnecessary competition for resources and opportunities. 

When an economy reaches the third stage of development (the stage where it’s being chased), fiscal policy regains its importance. At this stage, domestic savings are high, but the private sector is unwilling to invest domestically because the investing environment has deteriorated – domestic opportunities have dwindled, and investors can get better returns from investing overseas. The government should step in at this juncture, like what Japan did, and invest heavily in infrastructure, education, basic research and more. The returns are not high. But the government-led investments can make up for the lack of private-sector investments and the lack of consumer-spending because of excessive savings. In this way, the government can protect employment in society and prevent the formation of a vicious cycle of a decline in GDP. In contrast, monetary policy is largely ineffective in the third stage.”

It’s worth noting that an economy that is in the pre-LTP stage is likely to be in Case 4, where borrowers and lenders are both absent. Meanwhile, an economy that is in the Golden Era is likely to be in Case 1 (where both borrowers and lenders are present in abundance) and in Case 2 (where borrowers are present, but lenders are absent) during a run-of-the-mill recession, although a Case 1 Golden Era economy can also quickly be in Case 3 or Case 4. Once an economy is in the pursued stage, it is likely to be in Case 3 (where borrowers are absent but lenders are present) because of a lack of domestic investment opportunities or a balance sheet recession, or Case 4 (where borrowers and lenders are both absent) because of a balance sheet recession.

When a country’s economy is in Case 1 or Case 2, monetary policy is more important:

“Case 1 requires a minimum of policy intervention – such as slight adjustments to interest rates – to match savers and borrowers and keep the economy going. Case 1, therefore, is associated with ordinary interest rates and can be considered the ideal textbook case.

The causes of Case 2 (insufficient lenders) can be traced to both macro and financial factors. The most common macro factor is when the central bank tightens monetary policy to rein in inflation. The tighter credit conditions that result certainly leave lenders less willing to lend. Once inflation is under control, however, the central bank typically eases monetary policy, and the economy returns to Case 1. A country may also be too poor or underdeveloped to save. If the paradox of thrift leaves a country too poor to save, the situation would be classified as Case 3 or 4 because it is actually attributable to a lack of borrowers.

Financial factors weighing on lenders may also push the economy into Case 2. One such factor is an excess of non-performing loans (NPLs) in the banking system, which depresses banks’ capital ratios and prevents them from lending. This is what is typically called a “credit crunch.” Over-regulation of financial institutions by the authorities can also lead to a credit crunch. When many banks encounter NPL problems at the same time, mutual distrust may lead not only to a credit crunch, but also to a dysfunctional interbank market, a state of affairs typically referred to as a “financial crisis.”…

…Non-developmental causes of a shortage of lenders all have well-known remedies… For example, the government can inject capital into the banks to restore their ability to lend, or it can relax regulations preventing financial institutions from serving as financial intermediaries.

In the case of a dysfunctional interbank market, the central bank can act as lender of last resort to ensure the clearing system continues to operate. It can also relax monetary policy. The conventional emphasis on monetary policy and concerns over the crowding-out effect of fiscal policy are justified in Cases 1 and 2 where there are borrowers but (for a variety of reasons in Case 2) not enough lenders.””

When a country’s economy is in Case 3 or Case 4, fiscal policy is more important because monetary policy does not work when borrowers disappear, although the appropriate type of fiscal policy can also differ:

“It should be noted that in the immediate aftermath of a bubble collapse, the economy is usually in Case 4, characterized by a disappearance of both lenders and borrowers. The lenders stop lending because they provided money to borrowers who participated in the bubble and are now facing technical or real insolvency. Banks themselves may be facing severe solvency problems when many of their borrowers are unable to service their debts…

…In a financial crisis, therefore, the central bank must act as lender of last resort to ensure that the settlement system continues to function…

…Once the bubble bursts and households and businesses are left facing debt overhangs, no amount of monetary easing by the central bank will persuade them to resume borrowing until their balance sheets are fully repaired…

…When private-sector borrowers disappear and monetary policy stops working, the correct way to prevent a deflationary spiral is for the government to borrow and spend the excess savings of the private sector… 

…In other words, the government should mobilize fiscal policy and serve as borrower of last resort when the economy is in Case 3 or 4. 

If the government borrows and spends the $100 left unborrowed by the private sector, total expenditures will amount to $900 plus $100, or $1,000, and the economy will move on. This way, the private sector will have the income it needs to pay down debt or rebuild savings…

…It has been argued that the fiscal stimulus is essential when the economy is in Case 3 or 4. But there are two kinds of fiscal stimulus: government spending and tax cuts. If the economy is in a balance sheet recession, the correct form of fiscal stimulus is government spending. If the economy is suffering from a lack of domestic investment opportunities, the proper response would be a combination of tax cuts and deregulation to encourage innovation and risk taking… augmented by government spending…

…The close relationship observed prior to 2008 between central-bank-supplied liquidity, known as the monetary base, and growth in money supply and private-sector credit broke down completely after the bubbles burst and the private sector began minimizing debt. Here money supply refers to the sum of all bank accounts plus bills and coins circulating in the economy, and credit means the amount of money lent to the private sector by financial institutions…

…In this textbook world, a 10 percent increase in central bank liquidity would increase both the money supply and credit by 10 percent. This means there were enough borrowers in the private sector to borrow all the funds supplied by the central bank, and the economies were tin Case 1…

…But after the bubble burst, which forced the private sector to minimize debt in order to repair its balance sheet, no amount of central bank accommodation was able to increase private-sector borrowings. The U.S. Federal Reserve, for example, expanded the monetary base by 349 percent after Lehman Brothers went under. But the money supply grew by only 76 percent and credit by only 27 percent. A 27 percent increase in private-sector credit over a period of nearly nine years represents an average annual increase of only 2.75 percent, which is next to nothing.”

Fiscal stimulus equates to government-spending, which increases public debt. Koo suggests that (1) when an economy is in Case 3 or Case 4, rising and/or high public debt is not necessarily a problem, and (2) the limits of public debt should be determined by the bond market:

“Debt is simply the flip side of savings. Somebody has to be saving for debt to grow, and it is bound to increase as long as someone in the economy continues to save. Moreover, if someone is saving but debt levels fail to grow (i.e., if no one borrows and spends the saved funds), the economy will fall into the $1000 – $900 – $810 – $730 deflationary spiral….

…Growth in debt (excluding debt financed by the central bank) is merely a reflection of the fact that the private sector has continued to save. 

If debt is growing faster than actual savings, it simply means there is double counting somewhere, i.e., somebody has borrowed the money but instead of using it himself, he lent it to someone else, possibly with a different maturity structure (maturity transfer) or interest rates (fixed to floating or vice versa). With the prevalence of carry trades and structured financial products involving multiple counterparties, debt numbers may grow rapidly on the surface, but the actual debt can never be greater than the actual savings. 

Furthermore, the level of debt anyone can carry also depends on the level of interest rates and the quality of projects financed with the debt. If the projects earn enough to pay back both borrowing costs and principal, then no one should care about the debt load, no matter how large, because it does not represent a future burden on anyone. Similarly, no matter how great the national debt, if the funds are invested in public works projects capable of generating returns high enough to pay back both interest and principal, the projects will be self-financing and will not increase the burden on future taxpayers…

…Whether or not fiscal policy has reached its limits should be decided by the bond market, not by some economist using arbitrarily chosen criteria. 

During the golden era, when the private sector has strong demand for funds to finance productivity- and capacity-enhancing investments, fiscal stimulus will have a minimal if not negative impact on the economy because of the crowding-out effect. The bond market during this era correctly assigns very low prices (high yields) to government bonds, indicating that such stimulus is not welcome.

During the pursued era or during balance sheet recessions, however, private-sector demand for funds is minimal if not negative. At such times, fiscal stimulus is not only essential, but it has maximum positive impact on the economy because there is no danger of crowding out. During this period, the bond market correctly sets very high prices (low yields) for government bonds, indicating they are welcome…

…Ultra-low bond yields in economies in Cases 3 and 4 are also a signal to the government to look for public works projects capable of producing a social rate of return in excess of those rates. If such projects can be found, fiscal stimulus centered on them will ultimately place no added burden on future taxpayers.” 

The experience of the economies of the US, the UK, Japan, and Europe in the aftermath of the housing bubble bursting in 2008 which thrust them into balance sheet recessions is instructive on the importance of fiscal policy in combating balance sheet recessions:

“In November 2008, just two months after Lehman Brothers went under, the G20 countries agreed at an emergency meeting in Washington to implement fiscal stimulus. That decision kept the world economy from falling into a deflationary spiral. But in 2010, the fiscal orthodoxy of those who did not understand balance sheet recessions reasserted itself at the Toronto G20 meeting, where members agreed to cut deficits in half even though private-sector balance sheets were nowhere near a healthy state. The result was a sudden loss of forward momentum for the global economy that prolonged the recession unnecessarily in many parts of the world. After 2010, those countries that understood the danger of balance sheet recessions did well, while those that did not fell by the wayside…

…Bernanke and Yellen both understood this, and they used the expression “fiscal cliff” to warn Congress about the danger posed by fiscal consolidation, which the Republicans and many orthodox economists supported. The extent of Bernanke’s concerns about fiscal consolidation can be gleaned from a press conference on April 25, 2012, when he was asked what the Fed would do if Congress pushed the U.S. economy off the fiscal cliff. He responded, “There is . . . absolutely no chance that the Federal Reserve could or would have any ability whatsoever to offset that effect on the economy.”10 Bernanke clearly understood that the Fed’s monetary policy not only cannot offset the negative impact of fiscal consolidation, but would also lose its effectiveness if the government refused to act as borrower of last resort.

Even though the U.S. came frighteningly close to falling off the fiscal cliff on a number of occasions, including government shutdowns, sequesters, and debt‐ceiling debates, it ultimately managed to avoid that outcome thanks to the efforts of officials at the Fed and the Obama administration. And that is why the U.S. economy is doing so much better than Europe, where virtually every country did fall off the fiscal cliff…

…The warnings about the fiscal cliff set the Fed apart from its counterparts in Japan, the UK, and Europe. In the UK, then-BOE Governor Mervyn King publicly supported David Cameron’s rather draconian austerity measures, arguing that his bank’s QE policy would provide necessary support for the British economy. At the time, the UK private sector was saving a full 9 percent of GDP when interest rates were at their lowest levels in 300 years. That judgement led to the disastrous performance of the UK economy during the first two years of the Cameron administration…

…BOJ Governor Haruhiko Kuroda also argued strongly in favor of hiking the consumption tax rate, believing a Japanese economy supported by his quantitative easing regime would be strong enough to withstand the shock of fiscal consolidation. This was in spite of the fact that Japanese private sector was saving 6.2 percent of GDP at a time of zero interest rates. The tax hike, which was carried out in April 2014, threw the Japanese economy back into recession…

…ECB President Mario Draghi has admonished member governments to meet the austerity target imposed by the Stability and Growth Pact at every press conference, even though his own inflation forecasts have been revised downwards almost every time they are updated. He seems to be completely oblivious to the danger posed by fiscal austerity when the Eurozone private sector has been saving an average of 5 percent of GDP since 2008 despite zero or even negative interest rates.” 

Koo also noted that when Japan’s real estate bubble burst in 1990, the government was “quick to administer fiscal stimulus to stop the implosion” and that “the economy responded positively each time fiscal stimulus was implemented, but lost momentum each time the stimulus was removed.” The Japanese government was under enormous pressure to cut fiscal stimulus in the aftermath of the bubble, but the government did not completely cave, and the Japanese economy managed to fare better than it would otherwise:

“The orthodox fiscal hawks who dominated the press and academia also tried to stop fiscal stimulus at every step of the way, arguing that large deficits would soon lead to skyrocketing interest rates and a fiscal crisis. These hawks forced politicians to cut stimulus as soon as the economy showed signs of life, prompting another downturn. The resulting on-again, off-again fiscal stimulus did not imbue the public with confidence in the government’s handling of the economy. Fortunately, the LDP [Liberal Democratic Party] had enough pork-barrel politicians to keep a minimum level of stimulus needed in place, and as a result, Japanese GDP never once fell below its bubble peak. Nor did the Japanese unemployment rate ever exceed 5.5 percent.

That was a fantastic achievement in view of the fact that the Japanese private sector was saving an average of 8 percent of GDP from 1995 to 2005, and the Japanese lost three times as much wealth (as a share of GDP) as Americans did during the Great Depression, when nominal GNP fell 46 percent.”

The reason for US backlash against globalisation & the conflict between free-trade and free-capital

The fourth and fifth important ideas from Koo’s book are connected and they are respectively, (a) the possible reasons behind the backlash against globalisation that is seen from the current US government under the Trump administration, and (b) the possible conflict between free trade and free-movement of capital. Again, Koo’s book was published in 2018, so it was discussing Donald Trump’s first term as President. But the ideas appear to me to be very applicable to today’s context.

Koo advanced that the Western economies’ entrance into the third stage of economic development – pursued stage – is a reason for the backlash against globalisation:

“One reason for the frustration and social backlash witnessed in the advanced countries is that these countries are experiencing the post-Lewis Turning Point (LTP) pursued phase for the first time in history… 

…Many were caught offguard, having assumed that the golden era that they enjoyed into the 1970s would last forever. It comes as no surprise that those who have seen no improvement in their living standards for many years but still remember the golden age, when everyone was hopeful and living standards were steadily improving, would long for the “good old days.”…

…In the U.S. too, the Trump phenomenon, which has depended largely on the support of blue-collar white males, suggests that people are longing for the life they enjoyed during the golden era, when U.S. manufacturing was the undisputed leader of the world.

Participants in this social backlash in many of the pursued economies view globalization as the source of all evil and are trying to slow down the free movement of both goods and people. Donald Trump and others like him are openly hostile toward immigration while arguing in favour of protectionism and the scuttling of agreements such as the TPP that seek even freer trade.”

Koo described the mainstream view that free trade creates overall gains for trading partners, but cautioned that the view has a flawed assumption, in that imports and exports will be largely balanced as free trade grows, and it is that wrong assumption that also contributed to the backlash against globalisation:

“Economists have traditionally argued that while free trade creates both winners and losers within the same country, it offers significant overall welfare gains for both trading partners because the gains of the winners are greater than the losses of the losers. In other words, there should be more winners than losers from free trade…

…This conclusion, however, is based on one key assumption: that imports and exports will be largely balanced as free trade expands. When – as in the U.S. during the past 30 years – that assumption does not hold and a nation continues to run massive trade deficits, free trade may produce far more losers than theory would suggest. With the U.S. running a trade deficit of almost [US]$740bn a year, or about four percent of GDP, there were apparently enough losers from free trade to put the protectionist Donald Trump into the White House. The fact that Hillary Clinton was also nominated to be the Democratic Party’s candidate for president in the arena full of banners saying “No to TPP” indicates that the social backlash has grown very large indeed.”

Koo clarified that free trade is important and has its benefits, but the way free trade has taken place since World War II is hugely problematic because of (1) the way free trade is structured, and (2) the free movement of capital that is happening in parallel:

“Outright protectionism is likely to benefit the working class in the short term only. In the long run, history has repeatedly shown that protected industries always fall behind on competitiveness and technological advances, which means the economy will stagnate and be overtaken by more dynamic competitors…

…This does not mean that free trade as practiced since 1945 and globalism in general have no problems. They both have major issues, but these can be addressed if properly understood. A correct understanding is important here because even though increasing imports is the most visible feature of an economy in a pursued phase, trade deficits and the plight of workers displaced by imports have been made far worse by the free movement of capital since 1980…

…Once the U.S. opened up its massive markets to the world after 1945 and the GATT-based [General Agreement on Tariffs and Trade] system of free trade was adopted, nations belonging to this system found that it was possible to achieve economic growth without territorial expansion as long as they could produce competitive products. The first countries to recognize this were the vanquished nations of Japan and West Germany, which then decided to devote their best people to developing globally competitive products…

…By the end of the 1970s, however, the West began losing its ability to compete with Japanese firms as the latter overtook the U.S. and European rivals in many sectors, including home appliances, shipbuilding, steel, and automobiles. This led to stagnant income growth and disappearing job opportunities for Western workers.

When Japan joined the GATT in 1963, it still had many tariff and non-tariff trade barriers. In other words, while Western nations had been steadily reducing their own trade barriers, they were suddenly confronted with an upstart from Asia that still had many barriers in place. But as long as Japan’s maximum tariff rates were falling as negotiated and the remaining barriers applied to all GATT members equally, GATT members who had opened their markets earlier could do little under the agreement’s framework to force Japan to open its market (the same problem resurfaced when China joined the WTO 38 years later)…

…When U.S.-Japan trade frictions began to flare up in the 1970s, however, exchange rates still responded correctly to trade imbalances. In other words, when Japanese exports to the U.S. outstripped U.S. exports to Japan, there were more Japanese exporters selling dollars and buying yen to pay employees and suppliers in Japan than there were U.S. exporters selling yen and buying dollars to pay employees and suppliers in the U.S.

Since foreign exchange market participants in those days consisted mostly of exporters and importers, excess demand for yen versus the dollar caused the yen to strengthen against the dollar. That, in turn, made Japanese products less competitive in the U.S. As a result, trade frictions between the U.S. and Japan were prevented from growing any worse than they did because the dollar fell from ¥360 in mid-1971 to less than ¥200 in 1978 in response to widening Japanese trade surpluses with the U.S..

But this arrangement, in which the foreign exchange market acted as a trade equalizer, broke down with financial liberalization, which began in the U.S. with the Monetary Control Act of 1980…

…These changes prompted huge capital outflows from Japan as local investors sought higher-yielding U.S. Treasury securities. Since Japanese investors needed dollars to buy Treasuries, their demand for dollars in the currency market outstripped the supply of dollars from Japanese exporters and pushed the yen back to ¥280 against the dollar. This rekindled the two countries’ trade problems, because few U.S. manufacturers were competitive vis-a-vis the Japanese at that exchange rate.

When calls for protectionism engulfed Washington, President Ronald Reagan, a strong supporter of free trade, responded with the September 1985 Plaza Accord, which took the dollar from ¥240 in 1985 down to ¥120 just two years later. The dollar then rose to ¥160 in 1990 but subsequently fell as low as ¥79.75 in April 1995, largely ending the trade-related hostilities that had plagued the two nations’ relationship for nearly two decades…

…Capital transactions made possible by the liberalization of cross-border capital flows also began to dominate the currency market. Consequently, capital inflows to the U.S. have led to continued strength of the dollar – and stagnant or declining incomes for U.S. workers – even as U.S. trade deficits continue to mount. In other words, the foreign exchange market lost its traditional function as an automatic stabilizer for trade balances, and the resulting demands for protectionism in deficit countries are now at least as great as they were before the Plaza Accord in 1985.”

Specifically with regards to the belligerent relationship the US has with China today, Koo suggested that it was because of flaws in the free trade framework of the World Trade Organisation (WTO):

“…a key contradiction in the WTO framework: the fact that China levies high tariffs on imports from all WTO nations is no reason why the U.S.—which runs a huge trade deficit with China—should have to settle for lower tariffs on imports from China.

This problem arose because the developed‐world members of the WTO had already lowered tariffs among themselves before developing countries such as China, with their significantly lower wages and higher tariffs, were allowed to join. When they joined, developing countries could argue that they were still underdeveloped and needed higher tariffs to allow infant domestic industries to grow and to keep their trade deficits under control. Although that was a valid argument for developing countries at the time and their maximum tariff rates have come down as negotiated, the effective rates remained higher than those of advanced countries long after those countries became competitive enough to run trade surpluses with the developed world…

…Because the WTO system is based on the principle of multilateralism, with rules applied equally to all member nations, this framework provides no way of addressing bilateral imbalances between the U.S. and China. It is therefore not surprising that the Trump administration has decided to pursue bilateral, not multilateral, trade negotiations.

In retrospect, what the WTO should have done is to impose a macroeconomic condition stating that new members must lower their tariff and non‐tariff barriers to advanced‐country norms after they start to run significant trade surpluses with the latter. Here the term “significant” might be defined to mean running a trade surplus averaging more than, say, two percent of GDP for three years. If a country fails to reduce its tariffs to the advanced‐nation norm within say five years after reaching that threshold, the rest of the WTO community should then be allowed to raise tariffs on products from that country to the same level that country charges on its imports. The point is that if the country is competitive enough to run trade surpluses vis‐à‐vis advanced countries, then it should be treated as one.

If this requirement had existed when Japan joined the GATT in 1963 or when China joined the WTO in 2001, subsequent trade frictions would have been far more manageable. Under the above rules, Japan would have had to lower its tariffs starting in 1976, and China would have had to lower its tariffs from the day it joined the WTO in 2000! Such a requirement would also have enhanced the WTO’s reputation as an organization that supports not only free trade but also fair trade.”

Koo also noted that the term globalisation actually has two components, namely, free trade and free movement of capital, and that the former is important for countries to continue maintaining because of the benefits it brings, while the latter system needs improvement:

“The term “globalization” as used today actually has two components: free trade and the free movement of capital. 

Of the two, it was argued in previous chapters that the system of free trade introduced by the U.S. after 1947 led to unprecedented global peace and prosperity. Although free trade produces winners and losers and providing a helping hand to the losers is a major issue in the pursued economies, the degree of improvement in real living standards since 1945 has been nothing short of spectacular in both pursued and pursuing countries…

…The same cannot be said for the free movement of capital, the second component of globalization. Manufacturing workers and executives in the pursued economies feel so insecure not only because imports are surging but also because exchange rates driven by portfolio capital flows of questionable value are no longer acting to equilibrate trade.

To better understand this problem, let us take a step back and consider a world in which only two countries – the U.S. and Japan – are engaged in trade, and each country buys $100 in goods from the other. The next year, both countries will have the $100 earned from exporting to its trading partner, enabling it to buy another $100 in goods from that country. The two nations’ trade accounts are in balance, and the trade relationship is sustainable. 

But if the U.S. buys $100 from Japan, and Japan only buys $50 from the U.S., Japan will have $100 to use the next year, but the U.S. will have only $50, and Japanese exports to the U.S. will fall to $50 as a result. Earning only $50 from the U.S., the Japanese may have to reduce their purchases from the U.S. the following year. This sort of negative feedback loop may push trade into a “contractionary equilibrium.”

When exchange rates are added to the equation, the Japanese manufacturer that exported $100 in goods to the U.S. must sell those dollars on the currency market to buy the yen it needs to pay domestic suppliers and employees. However, the only entity that will sell it those yen is the U.S. manufacturer that exported $50 in goods to Japan.

With $100 of dollar selling and only $50 worth of yen selling, the dollar’s value versus the yen will be cut in half. This is how a surplus country’s exchange rate is pushed higher to equilibrate trade…

…If Japanese life insurers, pension funds, or other investors who need dollars to invest in the U.S. Treasury bonds sold yen and bought the remaining $50 the Japanese exporters wanted to sell, there would then be a total of $100 in dollar-buying demand for the $100 the Japanese exporter seeks to sell, and exchange rates would not change. If Japanese investors continued buying $50-worth of dollar investments each year, exchange rates would not change, in spite of the sustained $50 trade imbalances. 

Although the above arrangement may continue for a long time, the Japanese investors would effectively be lending money to the U.S. This means that at some point the money would have to be paid back. 

Unless the U.S. sells goods to Japan, there will be no U.S. exporters to provide the Japanese investors with the yen they need when they sell their U.S. Treasury bonds to pay yen obligations to Japanese pensioners and life insurance policyholders. Unless Japan is willing to continue lending to the U.S. in perpetuity, therefore, the underlying 100:50 trade imbalance will manifest itself when the lending stops.

At that point, the value of the yen will increase, resulting in large foreign exchange losses for Japanese pensioners and life insurance policyholders. Hence this scenario is also unsustainable in the long run. The U.S., too, would prefer a healthy relationship in which it sells goods to Japan and uses the proceeds to purchase goods from Japan to an unhealthy one in which it funds its purchases via constant borrowings…

…When financial markets are liberalized, capital moves to equalize the expected return in all markets. To the extent that countries with strong domestic demand tend to have higher interest rates than those with weak demand, money will flow from the latter to the former. Such flows will strengthen the currency of the former and weaken the currency of the latter. They may also add to already strong investment activity in the former by keeping interest rates lower than they would be otherwise, while depressing already weak investment activity in the latter by pushing interest rates higher than they would be otherwise.

To the extent that countries with strong domestic demand tend to run trade deficits and those with weak domestic demand run trade surpluses, these capital flows will exacerbate trade imbalances between the two by pushing the deficit country’s currency higher and pushing the surplus country’s currency lower. In other words, these flows are not only not in the best interests of individual countries, but are also detrimental to the attainment of balanced trade between countries. The widening imbalances then increase calls for protectionism in deficit countries.”

Prior to the liberalization of capital flows in the 1980s, “trade was free, but capital flows were regulated, so the foreign exchange market was driven largely by trade-related transactions.” This also meant that currency transactions could play the role they were meant to in terms of driving balanced trade:

“The currencies of trade surplus nations therefore tended to strengthen, and those of trade deficit nations to weaken. That encouraged surplus countries to import more and deficit countries to export more. In other words, the currency market acted as a natural stabilizer of trade between nations.”

As a sign of how free movement of capital has distorted the currency market, Koo noted that when the book was published “only about five percent of foreign exchange transactions involve trade, while the remaining 95 percent are attributable to capital flows.”

The problems with China’s economy

The sixth and last important idea from Koo’s book is a discussion of the factors that affect China’s economic growth, and why the country’s growth rate has slowed in recent years from the scorching pace seen in the 1990s and 200s. One issue, described by Koo, is that China no longer has a demographic tailwind to drive rapid economic growth, and is now facing a “middle-income trap” after passing the LTP around 2012:

“China actually passed the LTP around 2012 and is now experiencing sharp increases in wages. This means the country is now in its golden era, or post‐LTP maturing phase. However, because the Chinese government is wary of strikes, labor disputes, or other public disturbances of any kind, it is trying to pre‐empt such conflict by administering significant wage increases each year, with businesses required to raise wages under directives issued by local governments. In some regions, wages had risen at double‐ digit rates in a bid to prevent labor disputes. It remains to be seen whether such top‐down actions can substitute for a process in which employers and employees learn through confrontation what can reasonably be expected from the other party.

Just as China was passing the LTP, its working‐age population—defined as those aged 15 to 594—started shrinking in 2012. From a demographic perspective, it is highly unusual for the entire labor supply curve to begin shifting to the left just as a country reaches the LTP. Japan, Taiwan, and South Korea all enjoyed about 30 years of workforce growth after reaching their LTPs. The huge demographic bonus China enjoyed until 2012 is not only exhausted, but has now reversed… That means China that will not be able to maintain the rapid pace of economic growth seen in the past, and in fact growth has already slowed sharply. 

Higher wages in China are now leading both Chinese and foreign businesses to move factories to lower-wage countries such as Vietnam and Bangladesh, prompting fears that China will become stuck in the so-called “middle-income trap”. This trap arises from the fact that once a country loses its distinction as the lowest-cost producer, many factories may leave for lower-cost destinations, resulting in less investment and less growth. In effect, the laws of globalization and free trade that benefited China when it was the lowest-cost producer are now posing real challenges for the country.”

Koo proposed ideas for China to reinvigorate China’s growth, such as investing in productivity-enhancing measures for domestic workers. 

Another important factor affecting China’s economic growth involves the appropriate type of policy the government should implement to govern the economy. Since the country had passed the LTP more than a decade ago, and is in its golden era, fiscal policy – the act of the government directing the economy – is no longer the effective way to govern the economy. But is the government relinquishing control? To complicate matters, there are early signs now that China may already be in the pursued stage, in which case, fiscal policy will be important again. It remains to be seen what would be the most appropriate way for the government to lead China’s economy. 

---

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.