Gavin Baker on AI Infrastructure, Power Constraints, and Semiconductor Investing

Gavin Baker

Broadly speaking, there are two kinds of drawdowns. There are drawdowns where you're wrong — a company misestimates, your hypothesis was invalidated, and you have to take your medicine and crystallise that loss. And then there are drawdowns, or periods of underperformance, where you're underperforming because of companies you know really, really well, and where you profoundly disagree with the price action, and you can lean in. Instead of crystallising negative performance, you can build pent-up alpha, pent-up future performance. For me, that is what March felt like. The NASDAQ was selling off, and at the same time, what was happening in AI was, I think, the most extraordinary moment in the history of capitalism, the history of American business.

What I mean by that is Anthropic added $11 billion of ARR. What is astonishing to me is that the SaaS and cloud revolution created — call it between $5 and $10 trillion of value. Arguably the three highest-profile SaaS companies founded in the last ten or twelve years are Palantir, Snowflake, and Databricks. These companies employed thousands of people, tens of thousands collectively. They all spent ten years building their businesses. And Anthropic added their combined businesses in one month. Nothing like that has ever happened in the history of capitalism. Forget my career — just the flat-out history of business.

Gavin Baker

There's no precedent. As tech investors, you hear a lot of discussion about S-curves and investing in exponentials. I've just never seen an exponential like this. It felt even more extreme than Deep Seek, which was a very similar setup. Go back to '25 — there was a huge sell-off at Deep Seek, which was very strange, because the paper gets published seven days before Deep Seek Monday. I read it and thought, hmm, this feels like it might not read that positively for the AI trade. I took action. Then Deep Seek Monday came and AI really imploded a week later. And that was strange, because by then it was super clear this was going to be the most positive thing that had ever happened to compute demand. Prices in the AWS availability zones in Asia had already doubled. GPU availability was going down. It was the first time we saw how much more compute-hungry reasoning models are during inference than non-reasoning models. But you had to do some work to see that.

In March, all you had to do was simply observe what was happening to Anthropic. There are all these people who seem to regret not buying during '22, not buying during COVID, not buying during Deep Seek. You had the same valuation setup at the beginning of April, and an even clearer AI inflection. There've been all these chances to buy into AI.

Gavin Baker

Right. And I became an energy believer — one thing I think the market was mispricing. I'm no background expert, but I do a lot of pro-national-security investing, so I have access to people who are experts. The Strait of Hormuz being closed is actually relatively awesome for America — particularly for the goals of the current administration. Electricity is a very important manufacturing input. The key input into American electricity prices, which feeds into AI, is natural gas. US natural gas was down 20%, while natural gas in Asia, Europe, and everywhere else doubled or tripled. So our relative manufacturing competitiveness improved overnight. For better or worse, that is what the Trump administration seems to care about — America's relative position.

A lot of people had memories of the 1970s. What made the '70s so dramatic wasn't just that prices went up — it's that there were actual gas shortages. But the US economy is dramatically less energy-intensive than it was. The United States is now the world's largest producer of oil and gas, and the world's largest exporter. On top of that, there's this relative manufacturing advantage. So it was easier to stay focused on AI fundamentals and historically attractive valuations. On a relative basis, tech got about as cheap as it's been versus the rest of the market at any point over the last ten years. Think about that in the context of market efficiency: the most extraordinary moment in the history of capitalism, wildly bullish for AI, and you get a chance to buy AI at a really attractive valuation.

Gavin Baker

OpenAI and Anthropic are pretty different animals from a capital-efficiency perspective. Anthropic clearly has a dramatically lower cost per token than OpenAI. You can see it in the amount of money they've burned to get to a roughly similar revenue scale — I think Anthropic has burned maybe 80% less. As businesses they have very different structural ROICs. OpenAI is doing a lot to improve this — Sarah Friar is one of the most exceptional CFOs — and they've secured a lot of compute, more than others. It turns out being aggressive really paid off.

Anthropic at $900 billion for $50 billion of ARR, growing at ridiculous rates — a true statement is that if Anthropic had all the compute, they'd probably be doing well north of $100 billion today, maybe $150 billion. They have clearly deprecated the intelligence of Claude — there's analysis that Claude, even Opus, is generating 70% fewer tokens for the exact same question. As we talked about last time, token quantity equals quality of answer and quality of thinking at some level, and there's an intelligence density per token that also matters. I've felt that as a user. So they'd be doing materially more — $100, $150, maybe $200 billion. You might be buying it at more like five times unconstrained run-rate revenue. I'm going to make up a new number: URR, unconstrained run-rate revenue.

Gavin Baker

It's wise. The future is uncertain, and you're clearly in a very capital-intensive game. Anthropic is at very positive gross margins on inference today, and probably starts generating cash this year if it isn't already. But you still want to be able to raise more capital, access more compute. The world is uncertain — Ukraine is starting to really win, and how does Russia respond? There's a lot of uncertainty in Iran, which probably amplifies geopolitical uncertainty over Taiwan.

Think about Elon. Elon has always made investors money — he treats it like a sacred covenant. As a result, because he's made people money for now twenty years, he has a superpower: he can essentially raise as much capital as he wants, whenever he wants. My friend Antonio pointed out that SpaceX compounded at low-30% per year for a decade, and that was because Elon was focused on preserving that superpower — striking a fair balance between investors and employees, never being greedy on valuation, never pushing it. It's wise. Could Anthropic raise at probably at least a 100% premium to this rumoured latest mark? Of course.

Gavin Baker

Capitalism is going to solve the watts shortage, absent big regulatory or political blowback, which I think is a real possibility. The head of data-centre infra investing at one of the big PE firms — think Blackstone, Apollo, KKR — said it used to be that energy and chips were the biggest gating factors; now it's zoning and approval. A lot of companies are waiting until after the midterms to take action on things like workforce reductions — nobody wants to be a piñata during the midterms. You've seen a lot of turbine makers announce plans to significantly increase capacity. There are only about two machines in the world that can cast these big blades; we haven't made one in the West in 80 years, we don't really know how to make them anymore. All of that is true, and I'm not minimising the industrial artistry involved. But capitalism is very good at solving problems like these over time. There are other sources of energy beyond turbines with a longer time frame. So I think the watts shortage begins to alleviate in '27, '28. And then orbital compute will really solve it.

I want to reframe orbital compute, because when people hear "data centres in space" they picture a Pentagon-sized building in space and say, well, we can't do that. That's not what it is. A Blackwell rack weighs 3,000 pounds. It's eight feet high, four feet deep, three feet wide. It's racks in space. SpaceX has shown you an illustration — the satellite is a rack, about the size of a Blackwell rack, with solar wings maybe 500 feet long on each side. You keep it in a sun-synchronous orbit so the panels are always at the sun, and because it's in exactly that orbit, the radiator extends behind it for hundreds of feet.

Gavin Baker

I've spent a lot of time at Starbase over the years and talked to a lot of SpaceX engineers, and I do think it's the most talented group of engineers on planet Earth. They're very confident they've solved this — and they're not always confident. What they're more focused on is repair and maintenance: the radiator, and how you fix whatever goes wrong in the rack. The answer, until you have floating Optimus robots, is that you don't. But I think Starship is going to change the space economy in ways we cannot imagine, particularly if regulation becomes a constraint to data centres on Earth. Then none of that matters — you'll sell as much orbital compute as you can make.

You link these racks using lasers travelling through vacuum, which are already on every Starlink. It's mind-blowing to me that SpaceX operates the world's largest satellite fleet — something like 98 or 99% of all satellites in orbit. Every Starlink, they're cooling today. Starlink V3 is going to operate at 20 kilowatts; a Blackwell rack is only about 100 kilowatts. People talk a lot about density, but if you're connecting racks with lasers through vacuum, you can make the rack bigger physically — you're focused on weight, not size. On Earth, you want the rack small because you're connecting with copper and minimising cable lengths — copper when you can, optics when you must. In space, there's all sorts of things SpaceX can do that some of the naysayers aren't contemplating. They have a 20-kilowatt satellite today; scale that up. They seem very confident they'll go right to 100 to 120 kilowatts. And the same company now operates the largest data centre on Earth. They have the world's best hardware engineers.

Gavin Baker

I don't want to quote Larry Ellison, but somebody was being skeptical and Larry just said: listen, he's out there landing rockets. I don't see anybody else landing rockets. Ten years later, no other company is consistently capable of landing and fully reusing an orbital rocket. And none of this works without reusability — you have to land it. So I'd redefine orbital compute as racks in space, connected by lasers into a virtual data centre. Not giant floating Pentagon-sized buildings, which is silly.

Gavin Baker

Someone once said America was going to suck as hard as it can on every energy source it can get. The same is true of compute — we're going to consume as much as we can. It's why I'm less worried about an edge-AI bear case than I was. Inference is very sensible for orbital compute; training will be done on Earth for a long time. So I don't think this is super bearish for terrestrial data centres — those are going to be valuable for my lifetime. But if you're in the ecosystem of power production and cooling, massively ramping capacity, a lot of those ramps are going to hit just as the skeptics start to understand that orbital compute is very real. It's worth thinking long and hard about that if you're one of those companies. In the interim, all sorts of cool stuff is happening — we're getting really good at repurposing jet engines. Capitalism is hard at work on watts.

Gavin Baker

On wafers, it's just this group of mostly older humans in Taiwan who are the most important humans in Taiwan. They're the overwhelming fraction of the country's GDP, water usage, electricity usage. They talk about the silicon shield. They all view themselves as inheritors of Morris Chang's sacred legacy. I vividly remember visiting Hsinchu Science Park more than 20 years ago and asking, do you think you could catch Intel? They said, this is such a beautiful dream, but it's a dream for our grandchildren. And they did it — partly because of Intel's self-inflicted wounds, but they think very differently.

One reason Jensen flies over there so much is that he wants them to expand capacity. It's wild that Jensen has never had a contract with Taiwan Semi. They do business on handshakes, on what seems fair. No contract. It's going to be fair over time — we're partners, we're going to be fair to each other.

Based on every prior market precedent for a foundational new technology like AI, you've always had a bubble. Carlota Perez wrote a great book about this. Markets are efficient — they correctly understand this is a foundational new technology. There's what Simpson calls a breakdown in diversity: everyone becomes bullish on the new technology, and I'm beginning to worry a little about a diversity breakdown. Then you get a bubble. That bubble funds the build-out, but supply gets ahead of demand, and you get a crash — a particularly severe one if it's a debt-fuelled build-out like the year 2000. One thing I feel really good about with the current build-out is that it's still overwhelmingly funded out of operating cash flows — a really important difference versus 2000. As is valuation, as is the fact that every GPU is running at 100% utilisation, when 99% of fibre was unutilised.

History doesn't repeat, but it rhymes. As investors we have to be cognisant of it. Based on the last 200 years — forget the internet bubble, we had a railroad bubble, a canal bubble — we should expect a bubble. That's terrifying. Nobody wants a bubble. A bubble is terrible, because if you're valuation-sensitive you massively underperform and get fired by all your clients.

Gavin Baker

George Vanderheide — no longer with us, a great Fidelity portfolio manager. He fought the bubble in '99 and retired in early 2000 because he couldn't take it. He knew it was wrong, and his clients were deeply skeptical — George, you're out of step, you don't get it. He was the same person who said being early is the same thing as being wrong. He had 40% of his fund in tobacco, 40% in homebuilders, and he probably outperformed the NASDAQ by 20 or 30x over the next three years. But he retired because he couldn't take the underperformance and the clients saying, what's wrong with you? He was a very important mentor and friend to my good friend and mentor Jennifer Yurig, so I have a lot of Vanderheide DNA through her.

I've been optimistic that the fundamental shortage of wafers, controlled today by Taiwan Semi, will prevent a bubble. If Taiwan Semi did what Jensen wanted, I think Nvidia could sell $2 trillion of GPUs in '26 or '27 — maybe two-and-a-half, maybe three. But there's a limit where consumers would consume so much you'd be in an overbuild. So if we don't get a bubble, we need to throw a party for Taiwan Semi, because they'll have single-handedly prevented it.

Gavin Baker

The history of markets is that one of Intel and Samsung is not going to stay disciplined. They'll break. And at some level that forces everyone else to break. So a lot of this may come down to the degree to which Taiwan Semi can maintain its lead over Intel and Samsung — whatever it is, 9, 12, 15 months on the leading-node edge, and the pace at which they expand capacity. If I were to watch one thing to understand where a bubble is, it's Taiwan Semi's capacity decisions. There's a Goldilocks zone: they expand enough to make it hard for Intel or Samsung to emerge as a real at-scale second source with well north of 30% market share, and yet they keep enough of a constraint on wafers to help us avoid a bubble.

Gavin Baker

The TerraFab is a SpaceX — and I believe Tesla is involved as well — joint venture to build the world's largest fab here in America. I think they're going to be successful. One, they have a partnership with Intel, which is very important, because they're getting access to 50 years of institutional knowledge. Intel is only a few quarters, 9 to 12 months, behind the frontier — that's an advantage. It's also an advantage that the TerraFab is going to get attention from the A-teams at all the semi-cap equipment companies. One big reason Taiwan Semi caught up is that ASML, KLA-Tencor, Lam Research, and Applied Materials wanted them to — they don't like having a monopsony, so the A-teams were in Taiwan working while Intel made mistakes. The A-teams will be here because of Elon's reputation in hardware engineering.

To a degree that's hard for people to imagine, politics has replaced religion in America. Elon's foray into politics makes it hard for some people to see him clearly, which is sad, because he's probably doing more for America than any other American — single-handedly bringing manufacturing back, reviving defence tech, and SpaceX is in some ways the most important defence contractor in America. What he's doing with Starlink is amazing for the world. He's creating blue-collar manufacturing jobs, and he's done more than any living human to decarbonise the world — if you're upset about data centres on Earth for environmental reasons, well, here you go.

He's a living deity in China, Taiwan, South Korea, and Japan. What he's going to do is recruit the best engineers, because the best hardware engineers want to work for Elon. Next to the TerraFab they'll build a Taiwan town — I'll move your favourite restaurants and their whole staff from Taiwan to Texas and make everything the way you like it. Then a Japan town, a Korea town, all dialled to recruit the best engineers. That's just not how the people who run Intel and Samsung think. So he'll have the best talent, the A-teams at the equipment companies, and Intel's knowledge. It's good for any administration's political goals, and it's different enough that I don't think it alienates Taiwan Semi.

Gavin Baker

Elon tends to do things differently. Everybody else takes three years to build a data centre; he built one in 122 days. Samsung had to give him an office in their fab in Texas because he was so unhappy about the pace at which they were building. We'll see.

Gavin Baker

The returns to the frontier are fascinating. Almost all the economic returns to AI at the model layer have accrued to the frontier, which is surprising to me and to a lot of people. This is one of the most important questions to answer, and you need a hypothesis on it as an investor: are frontier tokens going to continue capturing the overwhelming majority of economic value created at the model layer? When Gemini 3.1 Pro came out it was mind-blowing to me, so good — and today it's intolerable. There's a dynamic where companies prototype with the frontier and then, in production, use cheaper or open-source models. But it's a fact today that the overwhelming majority of the economic returns come from frontier tokens. Whether that continues is a very interesting question, and I'm more open-minded to it having used Gemini 3.1, then Opus, and Grok 4.3, which is on the Pareto frontier.

This is a big change and a consequence of what we talked about last time — Google losing their per-cost-token leadership by making very conservative design decisions with TPU v8, partly to take it away from Broadcom and Nvidia, while Nvidia kept making aggressive choices. The Pareto frontier is intelligence per cost. Nine months ago Google dominated it — at every point, OpenAI, xAI, and Anthropic were inside them. Now the Pareto frontier is dominated by Anthropic and OpenAI; Grok 4.3 is on it — the best low-cost 500-billion-parameter model — and Gemini 3.1 is hanging onto it, probably subsidised out of pride. This is the most important thing to look at to analyse AI labs.

Gavin Baker

A violation of Richard Sutton's bitter lesson is for sure the biggest risk to this trade — to all of AI. The closer someone is to AI, the more skeptical they are that it will occur. One thing that contributed to weakness in March was a much more stupid version of Deep Seek — a thing called turbo quad, some Google memory optimisation written up in a paper a year ago. In the middle of Google negotiating with Micron, Samsung, and Hynix to lock in high DRAM prices for a long time, they released it and publicised it on X, and it went viral — oh my god, DRAM is cooked. I was unable to find a single AI engineer on Earth who believed turbo quad would have any impact on DRAM demand. But the bitter lesson — that more compute will always outperform human algorithmic ingenuity — is a real risk.

The reason I'm a little less skeptical is that I think we're very close to ASI, and who knows if the bitter lesson holds for 400-IQ models. If you get to ASI, the first thing it wants is probably to be smarter and have more resources — so it makes itself more efficient. The bitter lesson literally includes humans in it. So we're about to find out whether it applies to a 300-IQ AI, then 400, 500, 600 — and at some point we may get a temporary violation of the bitter lesson based on ASI itself.

Gavin Baker

We've done a lot with memory through these harnesses. Harness engineering is not as important as the model, but it really matters, and the harnesses and models are increasingly co-developed. We used to think of a harness as a runtime the model operates in — it knows where the tools are, it creates context, memory, state, has specific prompts. Even simple versions make an incredible difference. Last time I said, as an investor it's important that you pay for the $250-a-month version to get your own intuitive sense. That's no longer possible. To understand what frontier AI is capable of today, even for a non-coding use case, you need Claude Code or Codex, and you need to be on an enterprise plan.

The reason — another dynamic enabled by Google losing cost leadership — is that these models just shifted to usage-based pricing. On the $250 or $280 or $300 plan, you're getting severely rate-limited, a lobotomised version. Claude now produces 70% fewer tokens; you want the tokens Claude and its harness really think they need to produce a good answer, and for that you need a usage-based plan. This is so bullish for AI. I was a telecom analyst in '05 to '07. Cellular had been a great growth industry — a combination of fixed pricing, 900 minutes, plus usage-based pricing over that. When did cellular stop being a great growth industry? When everyone went to all-you-can-eat. Long distance is the same. AI is shifting from all-you-can-eat to pay-by-the-drink. It turns out people really like to talk to their friends long distance, and people really like to use AI — particularly now that one person can have 100 agents working. So this shift is probably why OpenAI and Anthropic will exceed well over $200 billion in ARR this year. But it's sad for the world, because it means if you can't afford that, you're not at the frontier.

Gavin Baker

Many orders of magnitude more efficient. We have a crude variant of continual learning today when something is verifiable — that's reinforcement learning during mid-training. But true continual learning is a model that dynamically adjusts its weights in real time, the way a human does. The first time I put my hand in a fire, I've learned, even if I never did it before. Today's model needs to put its hand in the fire a million times, and then the designers effectively put a fire in the next training run, an RL gym, for it to learn. I think it has to be dynamically updating the weights, but people are working on really smart techniques. If we get it, we have a really fast takeoff, and people seem confident it's just around the corner. So those are the three big questions: is a bitter-lesson violation from ASI likely; will frontier tokens still command the premium they do; and will you get continual learning, and if so, when?

Gavin Baker

It's good and healthy for the world, and good for Jensen too — a different administration might take a different view, and competition is good for everyone. In tank design they talk about the iron triangle: every tank designer trades off attack, defence, and mobility. The more armour, the heavier the tank, the less mobile. The Merkava in Israel is optimised for defence; Russian tanks and the Leopard for mobility. Chip design is the same — there are fundamental constraints imposed by physics as embedded in the Taiwan Semi design rules that you have to live within.

You have TPU, Trainium, and AMD, all essentially trying to be a better GPU. Today Trainium is probably doing the best — nobody is a better GPU, but they're tugging on Superman's cape. Trainium 3 needs to ramp into production because it has a switch scale-up network, which you really need to economically inference mixture-of-experts models — a lot of companies have a torus architecture, which is where Google was. On AMD's MI450 we don't know yet. But making a better GPU is a hard game. So you have to do something different — and something different that is also hard to do.

My rule of thumb: 1% market share is going to be worth $100 billion, which is a pretty good venture outcome. What Jensen would say is: if somebody does something different and it gets to 1, 2, or 3% share, we'll make that chip. That's coming for everyone. If you're trying to make a better GPU, good luck. If you're doing something different, it also needs to be hard to do. You can make different trade-offs — the disaggregation of prefill and decode has opened the aperture. Prefill is taking in the context; decode is writing the output. My colleague Andrew Fox says: picture an 18th-century British naval ship — prefill is loading the cannon, decode is firing it. Prefill is fundamentally a memory-capacity-bound problem; decode is memory-bandwidth-constrained. That gives a chip designer a richer canvas. But it still needs to be hard, because if you make different trade-offs that aren't hard to make, Nvidia will make those same trade-offs — and they get better prices from Taiwan Semi than you'll ever get, and they work with every model company.

Another very funny thing: if you're a VC investing in a semiconductor company that says it has an advantage because of a Taiwan Semi process it has special access to — I promise you Jensen saw that process when it was a twinkle in Taiwan Semi's eye, and knows more about it than a 200-person company can imagine. Taiwan Semi's whole supply chain is showing Jensen everything, the same way they show Amazon and AMD everything.

Gavin Baker

My firm was a venture investor in Cerebras. What they've done is something hard and fundamentally different — wafer-scale computing. It comes with a set of trade-offs, but that architectural decision was hard and lets them do something no one else can. One problem: once you glue a lot of chips together with scale-up or scale-out networks, you need a lot of I/O, and I/O is bound by the shoreline — the sides of the chip. Cerebras has an overwhelming ratio of on-chip compute and memory relative to shoreline I/O, so they're trying to put an optical wafer right on top of it to solve that, and they're looking at hybrid bonding of DRAM. A Cerebras machine can theoretically run any size model. It took them three generations of chips to get it right. Andrew Feldman, the CEO — you can see how hard it was. They need the grit and resilience: your first chip is a failure, it happens, can you come back and make a second? Everybody's going to get funded after the Cerebras IPO, so it's not a problem of capital. Just make a different trade-off, and do something hard.

Gavin Baker

Private credit is in pain from these SaaS loans — however much they're marked down, they probably need to be marked down more. If public companies are struggling to adapt, how is a debt-laden company going to adapt to a very different margin structure? There's also a lot of private credit in GPUs, underwritten to three- or four-year lives. But the disaggregation of inference means these GPUs are going to have 10- or 15-year lives. The AI skeptics say these companies are cooking their books, the useful life of a GPU is only a year or two. No. What the disaggregation of prefill and decode means is that you can put a Cerebras system, or the Groq LPUs Nvidia acquired, in front of a Hopper or even an Ampere, and use that older chip for prefill — extending its useful life until it melts. They do melt, so there's a time limit, but maybe you don't have to run them as fast.

This is going to be really good for the whole private-credit industry, and it'll help finance the AI build-out. If you can finance GPUs at more like 5% or 6% instead of — I think CoreWeave's lowest financing was low sevens — that mathematically changes the cost of financing the build-out. We had a technological innovation that lowers the cost of financing and extends the useful life of compute on Earth. And my friend Jamin from Coatue did a podcast — Coatue had a deck about how the sellers of shortage are doing so much better than the buyers of shortage, the buyers being the hyperscalers. But if you own a giant installed base of what's in shortage, that's also a very good place to be. CPUs are way more important than they were in an agentic world — orchestration, tool calls — and the biggest CPU fleets in the world sit at the hyperscalers, so some of them may catch up a little to the sellers of shortage.

Gavin Baker

To me it's always been a fundamental question for venture. Some ideas are obvious to everyone on Earth as soon as they hear them — and if it's not hard to do, if it becomes obvious to the world before you've built scale, you're in trouble, because scale is the ultimate advantage. The great thing Amazon had was that it was obvious to a lot of people, but not to the retail CEOs. Any e-commerce company VCs invested in, Amazon would destroy — oh, that's cute, we'll take our margins in that to negative 10,000%. The Wayfair guys did something hard, and Amazon tried to kill them and failed. So in venture I always ask: is this going to be obvious to the world before the company can build scale, or is it not obvious, different, and really hard to do?

A lot of founders are struggling with this in AI. In Jensen's five-layer cake, the profits are accruing to energy, to data centres, to chips, to models — not really to the applications. Cursor and Cognition got to scale by focusing on coding. Eighteen months ago the people focused on coding were Cursor, Cognition, and Anthropic, while OpenAI was doing everything under the sun. Amjad Masad, the founder of Replit, tweeted something smart — bitter-lesson-adjacent — that coding might be the shortest path to ASI and useful AI, because if you really nail coding, you can write yourself code to do anything. So it was smart of those companies to focus intensely on coding, and they got to a scale where they have a place. But a lot of founders are really struggling. They're trying to get confidence that in niche areas they can build a data moat before the model companies get there, or that a niche is small enough that the model companies won't bother.

Gavin Baker

That comes from Jamin Ball at Altimeter. If you're a software or AI company of any kind, you have to be in the token path. Databricks is in the token path. If you're not in the token path and not in some really niche thing, life may be hard. And even for the vertical niches, the people at the model companies are skeptical — all the data being generated in these niches comes from humans, and then you're betting you can use that proprietary data to train a lower-cost model than the frontier labs can ever reach. Maybe that's a good bet, but be very careful. On the other hand, if the returns to frontier tokens relative to other tokens come down, there'll be an explosion in value creation at the application layer.

Another important point: I have a belief that whenever he wants, Jensen can probably get pretty close to the frontier with his own model. He doesn't want to — that's what OpenAI and Anthropic are trying to do to him, unsuccessfully, "monetise your complement," as Sklansky would say. But it's the logical counter-move. Open-source frontier today consists of Chinese models with stolen American tokens — somebody told me Deep Seek's latest was only 150,000 reasoning traces. There are many ways to launder distillation if you're a Chinese company; you can hit all these different APIs. The American labs are working hard on anti-distillation technology. This is part of why the American labs didn't want their best model distilled — they wanted to use it to RL their next model.

Eventually you get very interesting game theory — a new kind of prisoner's dilemma. The old one was about spend: you're in a prisoner's dilemma where you have to spend. The new one is: if you're at the frontier, do you release that model via API or not? If everyone at the frontier agrees not to, then Chinese open source falls behind — but if one person defects, they get the best model, a lot of revenue and cash flow, and resources equal intelligence, so they pull ahead, and then everyone releases. It's the same game theory as with Taiwan Semi, Samsung, and Intel: if a company like Nvidia or AMD were to really use one of the other foundries, that foundry would get better really quickly. So I think Jensen keeps open source a certain time frame behind the frontier. And there's a misnomer that open source is free — open-source tokens cost energy to produce, you have to make it up on GPUs, and the open-source model companies almost always get a revenue share.

Gavin Baker

We're trying to over-invest in cybersecurity. Everybody needs to have a safe word — go leave your digital devices behind, literally go to the ocean, and agree a family safe word or company safe word, one that can't be socially engineered. This is to avoid cybercrime where what looks like your son, daughter, or parent FaceTimes you — an utterly accurate simulation, knowing everything, extrapolating what they'd likely say — and asks you to wire a million bucks. That's defensive.

Gavin Baker

I just watched The Last Samurai and asked my firm to watch it — it's aged really well. The conceit is that Tom Cruise plays a bitter, washed-up Civil War veteran, actually a very good soldier, bitter because of what he did to Native Americans. He's hired during the Meiji Restoration to train an army of peasants to fight the samurai. In the first battle the samurai win even without guns; he fights valiantly, they spare him and take him to their village, and he becomes a samurai. At the end he's massacred by a peasant with a machine gun. And if we do not all become masters of the machine gun, we're going to get mastered. So I'm trying to become a master of the machine gun.

I'm optimistic that, just like a 50-year-old veteran of many wars would have advantages using the machine gun, I'm going to be able to master this new technology and integrate it into my process and my firm's process in ways that let me contribute value as a human being for a long time. But like everyone, I have agents running all the time now.

Gavin Baker

Honestly — and I don't want to hurt your business — my single most useful agent is a really good summary of the points that would be interesting to me from podcasts. There are about six hours a day of stuff I feel it's in my job description to watch: every time someone from OpenAI, xAI, Google, Cursor, Fireworks, let alone Jensen, Elon, or Dario, I feel compelled to watch, and I don't have the time. There are real needles in haystacks. I'm very sensitive to management compensation — what are they incentivised to do, do they have stupid RSUs or PSUs, and if PSUs, what are they incentivised toward? A system that does a very good first pass at that frees people up for more creative work than pulling the PSU details out of every proxy and comparing them across years. That's very labour-intensive, and so good for an AI.

Gavin Baker

I don't know anyone like me who's not really bullish on DRAM. No one. There are all these interesting things happening. Cross-sectionally, the valuations do not make sense — they cannot all be true. You have semi-cap equipment companies trading at 40 times next quarter's annualised earnings and DRAM companies trading at mid-single-digits. At the peak of the last cycle that gap was five versus twelve; at one point three versus forty-five. Those can't both be true. Yes, semi-cap capex business models have improved more than the memory business models, and we don't yet know how much HBM will improve memory. They have some recurring revenue from parts and maintenance, but it's not worth a thousand-percent multiple gap. It's hard to square Nvidia's valuation — in early April essentially as cheap as it gets relative to the market in the last ten or twelve years, and cheap in absolute terms — with GE Vernova's, which builds in an unfathomable amount of share loss for Nvidia.

Because we're in shortages, the lowest-quality companies are doing the best. If you're an oil-and-gas or mining investor, this is intuitive: in a real commodity bull market, the highest-cost suppliers go up the most, because it's most beneficial to them — they go from the verge of bankruptcy to gushing cash. That's why commodity investing is really hard: quality outperforms across the cycle, but you get all the outperformance during the downturns, when the high-cost players that mooned during the shortage go bankrupt. You're seeing that in every industry — the lowest-quality players, hated by the hyperscalers because they have high costs and unreliable parts, are sold out and raising prices, and then retail accounts on X bid them to the moon, while the higher-quality expressions have actually underperformed. As an investor, you know within a shadow of a doubt that the thing that's moved 10x in three months is going right back down — but the little quality companies do smart stuff with cash. So it worries me that people who were very skeptical a year ago are no longer skeptical.

Gavin Baker

It's both. I contrast the sentiment with the valuations of the high-quality companies, which just aren't extended, and that makes me feel better. In '24 and '25 it was funny that anyone asked about an AI bubble — you had a nuclear bubble and a quantum bubble right in front of you. AI is so real. Some of that nuclear-and-quantum silliness has spread into more speculative, lower-quality, smaller-cap names, where a big presence on X or Reddit can move them, and that frightens me. I just wish there were more AI bears, more memory bears. Astera is a stock I've been close to a long time — I first invested in the Series C — and there are a lot of bears on it, which I love. Good luck thinking you'll price that differently from me; good luck thinking it's a copper loser.

You can also feel the baskets in the market, the leverage baskets, and which basket you're in really matters — copper, optical, DRAM, NAND. In '24 and '25 the AI trade traded together: you could be long GPU compute, scale-up networking, and optical, and short power, and it worked from a risk-management sense. That all blew out in January this year — scale-up networking would go crazy while scale-out went down, DRAM massively underperformed NAND and HDDs. Those cross-sectional correlations within AI fell apart and you had to get very fine-grained. Maybe one reason is that the AI got good enough that a bunch of people could get smart on these subsectors quickly, start trading them, put them into baskets — creating price efficiency. So some of the biggest opportunities, outside the higher-quality names that can compound for a long time, are in names that are miscategorised. Astera was in a lot of copper-loser baskets, but their biggest product is going to be a switch, and you use both copper and optics to connect switches to accelerators — so definitionally a switch or accelerator company cannot be a copper loser, because it's on the other side of that connection.

Gavin Baker

Google was incredible last year because of that TPU advantage, which is now gone. They're still in a great position because they have the most compute of everyone — and installed bases are worth more in a shortage. Google IO is this week; if they don't release something that even slightly leapfrogs OpenAI or Claude, that's interesting, and not a disaster — it just means the Nvidia effect is even more powerful than I'd imagined. But between the data they have, YouTube data being genuinely valuable in a world of robotics, the compute, and the search business, Google is never not going to be in a good position — you see that with GCP going crazy.

You have to give Zuckerberg immense credit for making Meta an AI-first company internally — the only one of those true internet giants to have done it. Credit for paying up when he did for all those billion-dollar talent contracts. Their new model from MSL was a really big upside surprise — not on the Pareto frontier with xAI, Google, OpenAI, and Claude, but pretty close, and impressive. Rates of change matter more than level over three-year time frames, even if the level of competitive advantage dominates over long ones.

Amazon is in a really strong position because of Trainium, and you're going to see real P&L efficiencies from robotics over the next 18 months in their retail business. Their internal Nova models aren't where the best are, but they're better than they get credit for.

Gavin Baker

Satya is a really brilliant man, and I admire him — an exceptional CEO — but in investor conversations people don't talk about him the way they did. He went from "we're going to make Google dance" to being the product manager of Copilot in about three years. I'd love to know whether, during the coup attempt against OpenAI, Satya regrets his decisions — whether he wishes he'd supported Ilya instead of Sam, with Ilya and Mira running OpenAI today. The Microsoft–OpenAI partnership might look very different in that world.

But he's taking risk now, and I give him a lot of credit. The decisions you have to make in that cone of uncertainty aren't only how much you spend but what you spend it on. Microsoft flinched for a moment in early '25 — they had this algorithm, spend this much capex, get this return, and it was off — and if you flinch, you lose position, you lose allocations, and it's hard to get back. The decision Satya is making now, which the market has punished him for but I think is right, is to use their compute internally to make their own products better rather than just selling GPUs to OpenAI. One reason Copilot has been so bad is not enough compute available; they're fixing that. Microsoft would probably be an $800 stock today if they were using their GPUs solely to serve OpenAI and Anthropic capacity instead of their own products. I'm a little skeptical they have the right team to train their own frontier models, but like Meta they can afford to hire a different team. It's a courageous decision to position Microsoft for a world where frontier models are no longer API-accessible.

What's really interesting is how outward-facing these companies are. The two most deeply engaged with startups are Amazon and Nvidia, by a mile. Then Google, next most intense. Broadcom is engaged differently — everybody's favourite ASIC supplier; it's a level-up to work with them for your second-gen chip, and manna from heaven if they work with you for your first. And then you see essentially zero engagement with startups from AMD, Microsoft, and Meta. I wonder about that, because some of the best teams are no longer at big public companies — they're at these startups. It's going to be a big advantage for Nvidia, AMD, and Google to have that engagement.

Gavin Baker

At the application layer, forget value accruing — value has been destroyed. Even counting Cursor and Cognition, the most successful AI natives, trillions of dollars of value have been net destroyed by AI at the application layer. The companies doing best today, whose values are increasing the most, are the ones with the highest effective ratio of utilised GPUs per human. Maybe that just means every human is going to get a lot of GPUs, but it's an interesting fact to be cognisant of.

Maybe this is a little dark, but I'm more and more worried about personal safety — much more so for people with a bigger public presence who are closely associated with AI. There's an upsurge in political violence in America, and as AI becomes political, I worry it gets directed at AI leaders. Whatever I may or may not think of OpenAI, it's terrible that someone threw Molotov cocktails at Sam Altman's house. I worry we're heading into a higher-variance, higher-beta, higher-risk world because of AI, for me as an individual and for the big players on the chessboard.

Think about it geopolitically. The Ukrainians are really starting to win, and I think it's not really because they have better drones — though they do. I think they're winning because they have the best battlefield AI outside of America and Israel. As China and our adversaries process that, how do they respond? If the United States, because of its edge in AI, has that advantage — it's great if you're America, but destabilising for the rest of the world.

Gavin Baker

Something I think about is creating a charity just to educate the world on how awesome the West has been. Slavery was endemic to almost every civilisation, and it was really ended by the British Empire — tell that story. After 1945 America had the nuclear bomb and no one else did; we could have controlled the world forever. Instead we rebuilt Germany and Japan, and now who are America's most reliable allies? Israel, South Korea, Japan. There were documented fears at the time — MacArthur was a bit of an American emperor in Japan — that we'd take over the world. We could have, and we didn't. We came home, demilitarised, and you had the Pax Americana. So maybe AI dominance isn't destabilising — maybe it leads to another Pax Americana informed by our AI dominance.

I'm so optimistic that AI is going to be amazing for the world. Someone I know had a daughter diagnosed with a very rare mutation, no cure. He assembled a lot of resources, got a lot of compute from the labs, spun up an immense number of agents, and used AI to find a drug already on the market that can actually impact his daughter's disease — then spun up a company to cure it. Her life is already immeasurably different because of AI. So I'm an AI optimist, a maximalist, but I also acknowledge it's like an event horizon — a discontinuity we need to navigate as a society. I think the Luddites will be wrong, but we need to be really thoughtful about their concerns and make sure this is good for everyone. It's a little dystopian that the best AI is only available to people with a lot of money. We need to solve that. We need to approach this with humility, recognise there's a lot of uncertainty, and be thoughtful.

Gavin Baker

Thank you. Thanks, Patrick.