Dylan Patel – Two labs will soon control most of the world's workforce

I'm back with Dylan Patel, founder of SemiAnalysis. Our version of a family Thanksgiving dinner is this regular yearly podcast—though we're not actually related, so don't tell anyone; it would destroy the myth.

Where the world economy is headed is increasingly becoming a function of where lab economics and the compute market are headed. I want to understand where this crazy future ends up within a few years, but let's start with where we are today: walk me through lab compute and lab revenue right now, and maybe project out a year or two.

Last year, even by the end of the year, most of GDP growth in America was just AI infrastructure. Looking at this year, about a third of the compute coming online is for the labs—OpenAI and Anthropic. It may be built by others and then rented to them, but at the end customer, it's them.

Compute Spending and Lab Centralization

The numbers for compute are ballooning. We're at a little over a trillion dollars of CapEx this year, and by 2028 it's going to be more than $2 trillion. The labs are also taking an increasing percentage of this, so you've got a very interesting situation where the labs go from companies spending tens of billions of dollars a year to hundreds of billions, to forecasting trillions of dollars a year even towards the end of the decade — at least based on some of the contracts they've begun signing with their partners. This requires a big reshaping of their economics.

From venture-funded losses to profit

Up until now, they have been companies that mostly lost money. Anthropic started turning a profit in Q2, and it's believed that at some point in Q3, OpenAI could start turning a profit too, with the bigger rise of Codex and 5.6. A year ago — even at the beginning of this year — all their money was venture-funded losses. They've now turned the corner and are actually starting to profit. That doesn't mean they're not taking in new capital; the new capital is still coming in to accelerate growth further. But more and more of their business is being funded off their own revenue rather than capital injections.

Over the last year and a half, their margins have really skyrocketed. The base cost of compute tends to be around $10, $13, or $15 million per megawatt. The most interesting aspect of what's happening now is this: before, when they served a model — GPT-4 on Nvidia Hopper GPUs — it generated negative gross margin for OpenAI. But now, when OpenAI serves GPT-5.6 or Anthropic serves Opus 5 or Fable 5, their revenue generation has passed well beyond the incremental $10–15 million per megawatt. In Anthropic's case, revenue has gone as high as $50 million per megawatt. What that enables them to do is: "Hey, if I spend 10 bucks on inference capacity, I actually generate 50 bucks of revenue, and then I can turn around and incrementally spend all of that profit on training."

When do the labs take most of the world's compute?

One thing I'm very interested in understanding is how you see the centralization of compute happening at the labs. If a third of marginal compute is going to the labs right now, by when is over half of the incremental compute in the world going to the labs? By what point do the labs have basically a vast majority of the world's compute?

At the beginning of this year, OpenAI started at 2 gigawatts and Anthropic at less than 2. By the end of this year, they're both above 5 — they've 3–4x'd compute as a whole. Looking at incremental compute added, that's about 30% of the compute added this year. Next year, given what's already been signed and inked, it's even more dramatic: Anthropic and OpenAI are taking as much as 40% to 50% of compute. This centralization doesn't look like it's slowing down or stopping — in fact, it looks like it's only accelerating.

Who's building that compute for them will change. Next year, a big new entrant is, for example, SpaceX, which is building a ton of compute and will most likely lease quite a bit of it to Anthropic and OpenAI, because they're the ones with the marginal capability to pay the highest price. In addition, the labs are starting to build their own compute — OpenAI with their own chips, Anthropic with TPUs purchased from Google and deployed with Fluidstack.

So when does half of the world's incremental new compute go to just OpenAI and Anthropic? Really by the end of next year. And because compute is growing so fast, incremental compute is going to be basically most of compute. So it's very soon — maybe within a year and a half or two years — that most of the world's compute is owned by two labs, or at least serving the demand from two labs.

There's this trend where world compute in gigawatts doubles every year, but the compute at the frontier labs triples every single year. Keeping the current trend going: from 2 at the beginning of this year to close to 6 at the end of this year, then 18 by the end of 2027 and 54 by the end of 2028. At that point, can they simply not continue tripling given the amount of world compute? How do you see the world compute situation over the next few years?

Efficiency gains compound the concentration

If the incremental compute this year adds 30 gigawatts, next year 50 gigawatts, and the year after that roughly 70, you end up with a really interesting phenomenon: a new watt deployed this year is significantly more efficient than the watts deployed two years ago. A humongous percentage of the world's compute was deployed this year. Even though it didn't double the number of watts deployed, I'm deploying GB300s, TPUv7s, and Trainium3s, which are 3–5x more performance per watt than prior-generation chips.

So ultimately you've got a huge ladder here. If Anthropic and OpenAI take on 45% of compute next year, then by, say, December '27 they've taken on half of the world's incremental new compute — but that half is actually at a higher performance than everything else before it, so you've got another multiplier on that. By the time you're towards the end of 2028 — if this trend continues, and I see nothing that's stopping it — you've got them just controlling most of the usable flops in the world on their own.

The thing I'm confused about is why you think we only add 80 gigawatts in 2028 if we enter a world in which the value of compute increases so much. That's the upper bound, by the way — the "I'm so bullish" case. Okay, let's do some chain of thought here.

When I interviewed you a few months ago, you said that to make a gigawatt of Vera Rubins, you need 55,000 N3 wafers, 6K N5 wafers, and 170K DRAM wafers. Those numbers may have changed.

Anyway, that's for one gigawatt. I had an LLM run your wafer fab equipment model and figure out how much the tooling costs to produce a gigawatt of compute every single year. It said $3–4 billion. Now suppose you add in cleanrooms, shell, and everything else at the fab: $6 billion of fab capex produces a gigawatt every single year. A gigawatt produces $100 billion of revenue right now. So over five years, the first gigawatt the fab produces has generated five years of profits, the second gigawatt four years, and so on. $6 billion of capex at the fab level will have generated over a trillion dollars of end AI revenue.

Yeah. There's a lot of opex along the way, and a lot of other capex — the data center, the power, installation. And you had to pay OpenAI for the R&D. There are a lot of different people who need money here. Take away half of it for all these middlemen. That still means there's a 100x discrepancy between fab capex and end revenue generated — more than that, actually, but we're being very conservative. This is capitalism: you have this huge discrepancy where you can turn $1 into $100.

They're not going to figure out a way to make more mirrors? They are. It's just that these mirrors take some time to make. But the emergency is so big where Anthropic and OpenAI are like, "We could make a trillion dollars right now, but we're just bottlenecked on the mirrors that go into the ASML machines. How can we make more mirrors if we spend $100 billion on this?" That's the situation we're going to be in pretty soon.

We're not going to be able to solve that supply constraint? That just seems quite hard to imagine. You've seen people do funny arbitrages here where they buy turbines and then try to resell them, because the value of a turbine is way more since it's the thing bottlenecking your data center. I think if anyone had $400 million and the ability to convince ASML to sell them an EUV tool, they should totally just go buy one, wait, and then sell it for north of a billion dollars. But ultimately, yes, capitalism will cause these things to expand. But it's a whip — it takes a long time for the whip signal to get to the tail end. The supply chain doesn't react immediately. In fact, you go talk to someone at Carl Zeiss, they're like, "Yeah, yeah, we need to make 100 EUV tools by the end of the decade." When we had our episode earlier this year, they didn't even think they needed enough mirrors to make 100 EUV tools a year. Now they're like, "Okay, we need to do that." But in reality, given all the economics going on, it should be even more. It takes so long to get pilled.

Suppose every single company in the stack got private-equity-acquired — somebody came in who was super AGI-pilled and said, "We're going to maximize production." What do you think the physical constraints on making more things would be? The reason I ask is we're pretty soon going to be in a world where lab revenue, or just AI cash flows — because obviously the accelerators also have these huge cash flows — will be so big that you can just fund extreme expansion of all this production from the cash flows themselves.

I do agree generally. There are obviously some physical constraints. The way the supply chain is expanding currently, 100 is roughly still the right number. For 2030? 100 ASML tools for 2030. But if you said, "Carl Zeiss, here's $10 billion, please just expand production," that would change things. You would have to do this with every company in the supply chain.

But you don't think that's going to happen next year? I don't think it'll happen this year, next year, or the year after, because the world is capital constrained. But in a world where the top labs are generating even a combined trillion dollars in revenue next year, they're not able to take $10 billion of that? I don't think they're going to do that, but… Or hundreds of billions at least? It just seems like they realize where the world is headed — they could just make it happen.

The thing is, the labs are going to generate hundreds of billions of revenue next year. But ultimately, capex next year is like $2 trillion, so you've got this big mismatch. The wafer fabrication equipment supply chain will do something on the order of $200 billion. The data center market supply chain will do even more, the accelerator supply chain even more, and the energy supply chain a number too. You sum all this up, it's going to be well north of $2 trillion of capex. So the labs have not yet gotten to the point where their cash flows can fund this stuff. Obviously they will never get to that point, because you want to keep your capex higher than your returns — yeah, you reinvest.

The key question I want to understand is: if the current trend continues, it would be north of 50 gigawatts per lab by the end of 2028, so between them they'd have 100 gigawatts. Those gigawatts would drive many-fold more throughput or performance by 2028 than they do now, because the hardware has gotten better — not only have flops per watt increased, but the hardware also gets better at working with AI workloads.

So that's 100 gigawatts for the labs by the end of 2028. How much is world compute? That may be difficult to pin down, given that by 2028 the labs would have taken 70–80% of incremental compute. And I'm not sure what happens to markets then. How much does the price of compute skyrocket for them to actually be able to buy 70–80% of it? Is Google, Meta, or Amazon even willing to sell that much?

One caveat on these gigawatt numbers: when Amazon serves Bedrock Anthropic models, that counts as Anthropic compute in our worldview, because it is effectively counted as revenue for Anthropic, even with the revenue share and credit-back arrangements.

If the labs reach 100 gigawatts combined in 2028, they will have done really disruptive things to the market. Anyone can make money off of $10–15 million per megawatt of compute today — it's not that hard. Get a GB300 rack, download the Kimi weights, download vLLM or SGLang, set it up. Codex and Fable can actually help you do this. It's not trivial, but it's not rocket science. Put it on OpenRouter, and you'll start generating more revenue than you're paying for the compute.

This has already led to that $10–15 million per megawatt pricing starting to inflect up. To get to 100 gigawatts in 2028, you have to believe the labs can outpay everyone for compute, because anyone can make money at $10–15 million. Does compute get to $25 million a megawatt? $40 million? The labs already generate far more revenue per megawatt than everybody else, and if they stay as far ahead as they currently are, you'd expect that to continue. If there's some kind of recursive self-improvement — the labs being relatively uplifted, or running internal models they don't release that help make their next model better — you'd expect it to be even more the case.

Aren't you already seeing this, where SpaceX or whoever is slightly further behind will just sell compute to the highest bidder if they can't internally monetize it as well as the labs? You'd expect the labs to keep bidding for larger and larger shares of the compute market. That is my worldview: they will continue to gobble up more of the compute. But they can't do it at current pricing or anywhere close to it. They have to start paying $25, $30, $50 million a megawatt to really take 70% of the world's compute in 2028 and reach 100 gigawatts — a very aggressive goal.

The other challenging aspect is that we've already seen a huge slowdown for the AI labs. The regulation they advocate for actually slows the labs down far more than it slows the open-source Chinese language models. OpenAI not releasing Astra. OpenAI stopping training for two weeks. Anthropic not releasing what their safety assessment calls Model 2, widely believed to be the next version of Mythos. They're clearly not releasing their best models, in which case their revenue per megawatt stalls or can even start to decline again, because other models become competitive. It's not that they're falling behind — they're just not releasing their best stuff. If regulatory impact prevents them from releasing their best models, their revenue per megawatt doesn't climb as fast, their ability to buy incremental compute at a higher price than everyone else starts to diminish, and maybe they can't get to 100 gigawatts.

But in a world where safety doesn't matter, I do believe that's exactly what happens. They can start generating $100 million per megawatt or more and pay $50 million a megawatt, and no one else has any logical reason to do anything with their compute besides saying, "Please, Dario, take everything off of my hands." But there are forces at play, which we cannot describe, that could potentially slow this down.

A good intuition pump: what if AI models were literally as good as a fully automated software engineer? They're not there yet — I think they're far from being able to fully automate the job of a full white-collar worker. But white-collar workers earn six figures or more a year. A gigawatt could sustain a population of roughly a million white-collar workers, which would be $100 billion. That's actually surprisingly low — $100K per person, a million population. But it would be many hundreds of billions of dollars per gigawatt if you get full AGI.

One thing we've continued to see is that most of the value capture is not happening at the model companies. Most of the value these models generate does not go to OpenAI and Anthropic — thankfully, so far it mostly goes to the users.

Jane Street, with their exclusive contract for OpenAI's GPT-5.6 Ultrafast mode, or as one of Anthropic's biggest customers, is generating far more value out of the tokens they pay for than Anthropic generates in profit, because they make money off the market. Or take Meta, at one point rumored to be as much as 10% of Anthropic's business: they generate far more efficiencies by optimizing their ad algorithms, getting engagement time 5% longer, and so on. They make way more money from using these models than Anthropic does. That's what's required. Sure, if you had a million new software engineers, the cost of a software engineer would also fall.

One thing I'm confused about: does the market come into equilibrium? If it does, would you expect the price of compute to equal whatever Anthropic and OpenAI can generate from it, or be very close to it with a small markup? Right now it's really weird that there's a 4x or more difference between what compute sells for and how much money Anthropic can make from it. In a world where revenue per gigawatt keeps increasing — if Anthropic's ability to monetize a gigawatt doubles or triples — it would be weird if that gap kept growing. Anthropic, just by having some weights, can take something that cost them $10 and turn it into $100.

Where the value has gone over time

This is always a fun question: where does the value go in AI? The end user, I think we all agree, generates more value than anyone else — hence they're paying a lot for these models. Then there's the app layer, which so far has generated very little value. Then the model layer, which up until a year ago was generating negative gross margins and is now generating massive positive ones — it looks to be on a path to $100 million per megawatt, turning $10–15 into $100.

But go back a year and the hardware supply chain was generating all the gross margin while literally everyone else was losing money: OpenAI and Anthropic were plowing in VC money, as were many startups, and many hyperscalers were building infrastructure without knowing if there'd be a payoff. The model layer was effectively creating negative value, selling tokens for less than the infra cost them. All the value was captured at the chip and the fab. Initially in 2023, the memory makers were making no money off HBM or memory for AI, even though the value they were delivering was theoretically humongous. Now, actually, TSMC captures way less value than the memory guys. So value capture has shifted around a lot, which is very fun for people tracking or participating in the market — Jane Street, for example.

Compute built without a customer

So what happens going forward? Anthropic and OpenAI have slowly started to balloon in value capture. Do they balloon and take all of it? That was the thought — and then Elon showed that, actually, no: he could sell his compute for $25 million or $40 million a megawatt to Anthropic and Google. Even if it's short-term, he's sold it at that price and will recoup his entire CapEx in a year.

What's your prediction for how much the relevant tranche of compute — B300s or whatever SpaceX sold for $40B a gigawatt to Google — sells for at the end of next year? I think most compute will still transact at sub-$20 billion a gigawatt. Even at the end of next year? Because all of it has to be financed. If Meta, Microsoft, Amazon, or SpaceX can build compute without finding a customer — just saying, "I'm going to build this compute" and waiting until it's built — they now control what's going on. Most compute is contracted well before it's built. That's what Elon took advantage of: he actually had all this compute and could say, "Hey Anthropic, I know you're making $60-plus billion per gigawatt — why not buy my stuff for a crazy amount of money?" Obviously it's not that Elon or Anthropic decided this; the market figured itself out.

Normally, you go to a random cloud operator building a gigawatt or 100 megawatts of compute: they spend the CapEx, then need to find a customer. To find a customer, they need capital — and the customer has to sign a deal, which they then take to the credit markets to raise the capital. Meta operates on a completely different power structure: they're effectively hoarding compute. Meta and SpaceX are the only plausible #3, because they're hoarding compute — using their balance sheets and capabilities to build compute without an end customer, monetizing at a huge degree. They have actual balance sheets, so they can go to the credit market directly. You build a gigawatt, you make your margin — not a crazy margin, but a good one.

Now Meta and SpaceX have the optionality of deciding whether their internal use case makes them more money, or whether they should sell to Anthropic or OpenAI at crazy margins. We've entered a regime where they're saying: "I'm going to build the compute, and I can rent it out for not $13 — I can sell it for $25, $50, and more."

Actively searching for editors has been quite time-consuming because the vast majority of candidates don't fit the profile I'm looking for. So I created a recruiter in Grok Bot to see if it would help. I gave it a huge context dump where I monologued basically everything I wanted, and it spun up four other bots to narrow in on different parts of the search: one went through the last year of my email for relevant inbound, one searched my X feed and DMs, one went through the end credits on various documentaries I like, and the last one looked for editors who work for some of the YouTubers I follow.

Grok Bot then took all the candidates the subagents had found, filtered them against my criteria, and delivered a final shortlist to review. To be honest, the first batch had a few good candidates I wanted to see, but mainly a bunch of duds. After I gave Grok Bot more feedback about what it was missing, it came back with a new list of candidates I'm actually extremely excited about.

I saved this whole workflow as a routine, so every week Grok Bot now checks my inbound email and X DMs for promising new candidates to potentially interview. If you want to try Grok Bot yourself, go to x.ai/bot.

Asked what revenue per gigawatt would be for Anthropic or OpenAI by the end of 2027, the answer was that it depends heavily on who has the best model and whether they're allowed to keep releasing it — but there's no reason it wouldn't be $50-plus million a megawatt. By the end of '27, that's where it gets more challenging, though it could go higher, to $70–80 million a megawatt blended across the company, if not higher.

If that's the case, what happens to the price of compute? If I'm Anthropic, incremental compute is worth it — maybe I spend $40 million a megawatt on SpaceX compute. If I'm SpaceX, I look to the supply chain: I've struck this deal with Jensen (who's now suddenly using Twitter), and Elon is saying they're exclusive to Nvidia — so why doesn't Jensen raise his prices? Then SK Hynix, Micron, and Samsung ask why they shouldn't raise theirs. There's a bullwhip effect in value capture: just because one player raises prices doesn't mean the whole supply chain rebalances immediately, but over time it will, and things will cost more and more — to get incremental capacity, you sort of have to. TSMC is raising prices very slowly, but memory companies and substrate companies are raising them very quickly. Elon wouldn't have sold at $15, but he's selling at $25+, so he clearly raised his prices fast.

The pushback: it's surprising to think revenue per gigawatt doesn't increase way beyond even $100 million per megawatt by the end of next year. When does RSI happen? When does takeoff happen? Or even without RSI, if the current rate of progress continues — look at how much progress we've made in the last year and a half. What was the model a year and a half ago, Claude 3.5 or something? The counterargument is that the best model in the world was trained in February — so the claim is really that the labs just won't be allowed to release their best models. OpenAI says they're not training models for two weeks, man. What the hell?

There are two separate questions here. One is whether, internally, these companies are getting enough use out of their models that they'll bid up the price of compute. The other is whether AI progress as a whole slows down because of regulation. The complaint: they're not even allowed to use the new model internally — Astra isn't even widely deployed internally. But still, compare the model released at the beginning of last year — GPT-4o? — to a Mythos 2-size leap by the end of 2027. The response: Mythos 2 isn't out. Or even Mythos — that leap isn't allowed to happen either. They've neutered it: we can't use it to optimize inference performance, or all sorts of other things.

The concession: maybe there is some slowdown in AI progress or deployment that means revenue per gigawatt could be lower — but that's the only way to get to only $100 million per megawatt by the end of next year. As long as the model gets better, the value generated from it gets better. Who captures that value is still up for debate, but ultimately everyone will raise their prices, because they can, and it's super inflationary.

Especially given how regulation is evolving. So far it's just "don't release the models," but more and more it's becoming physical: New York is banning data centers, Texas is holding moratoriums, Ohio is trying to require paying everyone's property tax within a certain radius. These measures decrease supply and increase cost, and that gets passed on too. You end up in a spot where progress does slow, at least in the external sense, even if the models internally keep getting better and better.

In a takeoff scenario, why wouldn't Anthropic have their best model six months ahead of what's externally available? Because of safety and regulation, but also competitive advantage. And if progress accelerates, that six-month gap is actually a bigger differential — which is what would cap revenue-per-megawatt gains to much lower growth than we saw in the first half of this year.

Once these companies are public and accountable to investors, imagine that by the end of next year they have close to 20 gigawatts. Ten percent of that is 2 gigawatts. Suppose they want to move from 60% of compute on training to 70%. Their investors would object: "If you can generate $100 billion per gigawatt, you're basically saying no to $200 billion of revenue in order to increase your training compute." Investors would ask why they're spending even more on training when they're already spending so much. As a public company, what would happen if they simply said, "No, we will keep increasing the share of compute we spend on training to offset the increase in revenue that each gigawatt of compute is giving us"?

This is what I personally believe, and it's non-consensus: the labs are going to allocate less and less compute to inference over time. The standard belief is that most compute will go to inference. In fact, most of it will go to forward passes for training, not necessarily revenue-generating inference.

If they're generating $30–40 million per megawatt today and allocate 40% to inference, then when they reach $60–70 million per megawatt, do they still allocate 40% to inference, generate all this profit, and pay dividends and buy back shares? Or do they go build AGI? I think the obvious answer from Anthropic and OpenAI—not just at the executive level but also their boards—is to go build AGI, because it's way more profitable. So you'll see them ratchet up the percentage of compute dedicated to training, even as each increment of compute would be more profit-generating if dedicated to inference.

The whole point is: if I'm selling tokens—is OpenAI releasing Ultrafast mode only externally, or internally too? It turns out they're allocating it to both, because the internal value from super-fast AI, or the best AI model, is way more than what an external customer generates. Sure, I could generate $100 million per megawatt, but if I turn that toward AI research, what incremental progress do I get toward my future earnings potential, the discounted cash flows of what I've done? They're not going through that calculation explicitly, but ultimately it makes more sense to dedicate more and more compute internally. The only reason to have inference compute be so large is so you can grow your training fleet.

This is an interesting economics question the models could digest: what would have to be true about a world where they reduce the fraction of compute spent on inference? I think they have been doing so over the last three months already. Take it month by month: every month Anthropic has added more compute than the prior month—there might be some noise when they sign a SpaceX deal or whatever, but in general the compute curve goes up. In January they added less compute than December, yet their revenue additions skyrocketed; then they plateaued. They're not adding $25 billion of ARR every month now. That means the marginal megawatt they're getting is going as a higher percentage to R&D than to inference. So they are factually increasing their compute toward R&D today. I think this is self-evident if you look closely enough at what they're doing.

Adding up the figures discussed, global compute would exceed 200 gigawatts by the end of 2028. Asked how fast that can keep growing after 2028, the projection is 30 gigawatts added this year, 50 next year, 70 in '28, and on the order of 90–100 in '29. Beyond that, the slope can keep going upward, but it's hard to predict more than four years out — nobody knows whether we're in an RSI regime, or when the world economy grows at 10% a year, because at 100+ gigawatts added per year you're talking about absurd GDP growth.

China's share of new compute

Level-set back to 2022: the US was adding about 45–50% of the world's compute, China about 30–35%, and the rest of the world took the remainder. Since 2022, big regulations against China and a dramatic increase in America mean that today 70% of watts are deployed in the US, while China gets a very small number — sub-10% of watts deployed for data center AI compute. Their domestic production is quite small, their Nvidia purchases are still quite small, and a lot of those end up elsewhere, in Malaysia or the like. So China domestically continues to have sub-10% of incremental new compute, though in 2028 it might start to inflect up. It's pretty easy to say China will have 30 gigawatts of AI compute or less in 2028.

When the hockey stick starts

In 2028, China should see a big uplift in deployable compute. In 2026 they're still mostly relying on smuggled chips, chips TSMC made for companies that turned out to be Huawei, and HBM Samsung is shipping. But in '27 fabs start to go up, and especially in '28, with SMIC and CXMT and others, domestic production reaches many millions of units a year — incrementally adding 5–10 gigawatts in 2028 alone from domestically produced chips. Those chips are definitely worse than what Nvidia, Google, or OpenAI will have in 2028, so even the gigawatt number overstates things: it's 30 gigawatts, but of much worse chips.

If the world adds 100 gigawatts the following year, how much can China add? Will they hockey stick once they can ship large amounts of compute, or stay below the US plus allies? A lot depends on whether the US passes the MATCH Act, whether tools stay export-controlled, and how fast China can build the new equipment it's starting to produce domestically. But China will definitely hockey stick — if there's anything China is really good at, it's scaling manufacturing very quickly. They should be able to extract more purchasing of even foreign chips into domestic China, or at least close the gap in what the US allows Nvidia to sell them.

Could China add 50 incremental gigawatts in 2029? That's completely reasonable, partly from foreign purchases. But if most are domestic chips, there's a factor where 50 gigawatts is really worth as much as 20 gigawatts of American chips. So the projection is a world where the leading lab in 2028 may have more compute than all of China in '29 or even '30, weighted by quality — assuming nothing is done to slow the US labs. That's right, though the US government and politicians are starting to do exactly that, whereas China won't slow down AI; the only thing they'll do is accelerate it.

Did export controls work?

The interviewer, describing himself as libertarian, says that when he interviewed Jensen Huang about export controls he wasn't sure what he thought, and steelmanned the opposite view — that cooperating with China might be better, especially since China controls so much of the supply chain needed for robotics. But he hadn't realized the compute situation was this lopsided. If China ships only the amounts described, then by the time we have an automated coder and are getting into automated researcher, China is far behind on compute stock — and if that's the case, the export controls would have worked. He calls that a notable success.

The caveat: some of this is export controls, but some is financial systems. American financial systems are more willing to YOLO into startups than Chinese ones, but once Chinese financial systems pick an industry to focus on, they subsidize it far more heavily — the Chinese semiconductor industry gets significantly more subsidies than the rest of the world's combined. If takeoff is slower than implied, China will drastically catch up on the semiconductor side, which eventually means compute.

The other noteworthy aspect of this is that Chinese companies today are not that far behind in AI models, at least as perceived by the public, relative to the amount of compute they have. The leading Chinese labs have 100–200 megawatts of compute at most, with ByteDance Seed the one outlier having significantly more. Kimi is not running a gigawatt or anywhere close to it, whereas Anthropic will have more than 5 gigawatts by the end of the year.

So the question is, does it matter? Right now, this difference in compute doesn't matter that much. Breaking down a lab's compute budget, so far it's been 60% training and 40% inference, but training itself splits further: 50% of total compute goes to research, 10% to development, and 40% to inference. By research I mean researchers generating ideas and testing new architectures, data mixes, hyperparameters, and attention techniques. But when they actually do the training run — when Anthropic trains Mythos — the pre-training is sub-200 megawatts for roughly two months, and the RL is even less.

You think the RL used less compute than the pre-training? At least in terms of a single pre-training site, yes. Total compute was probably higher, but it's sequential: at most, they ever used about 200 megawatts at one point in time. In reality they had multiple gigawatts, so most of their compute was going to research, not the development of a model. There are reasons for this: it's hard to coordinate all these clusters, hard to co-locate them, hard to do multi-site training, and hard to do RL — generating more rollouts during RL doesn't necessarily make it better. There are all sorts of reasons why you may not be able to leverage all two gigawatts onto training when you can only actually leverage 200 megawatts.

As we get further into automated coding and automated researchers, I expect the split of the compute budget between research and training to become much fuzzier, or even shift toward training. Things like continual learning also mean more and more compute goes to actually training the model.

If you end up in a world of 100 gigawatts a year, at current prices that would be $5 trillion of CapEx every single year. Then stack on the fact that you have to build the power plants well before then — they're a 30-year asset — and the data centers are a 15–20 year asset that also must be built ahead of time. So the $5 trillion, once you account for future years' growth, is actually more like $7–10 trillion of CapEx.

Wait, I don't understand — that doesn't include the infrastructure for power generation and the data center itself? Right, exactly. When people talk about AI CapEx, they quote $40–50 billion, but that's really just the critical IT: servers, networking, fiber, transceivers, optical communications. It doesn't account for the data center itself or the power plants being built ahead of time. If you're building 100 gigawatts this year and 150 next year, all the buildings for that 150 gigawatts need to be built in this year's CapEx; if you're building 200 gigawatts the year after, you have to buy the turbines this year. So it's much bigger than even $5 trillion if you're building 100 gigawatts.

Very plausibly, incremental CapEx every year gets close to $10 trillion by the end of 2030 — close to a tenth of the world economy. If all of it is in the US… the US economy will have grown too, but at its current size, that would be like a third to a quarter of the US economy going toward data centers. As I say that out loud, I think: maybe you're right and we just won't allow it, and that's why this doesn't happen. For this exponential to continue, a quarter of America's economy would be building data centers. I believe in capitalism and reallocating resources toward the most profitable thing, but politics exist, credit markets exist, and capital markets exist.

To enable, say, 100 gigawatts by 2030 — or even pare it down to 2028, at $3–4 trillion of CapEx across all these items: over $2.5 trillion toward IT CapEx, another $1–2 trillion on data centers and energy, plus the downstream supply chain like semiconductors — where does all this cash come from? No one is generating that much cash from the business yet. Hyperscalers funded all the growth up until now — Google, Microsoft, Amazon, Meta funded a huge percentage, more than half of compute — but they now don't generate cash; they spend everything on CapEx. In addition, they raise debt and spend that on CapEx too. You've seen Meta do it, even Amazon, even Google; Microsoft will be there soon.

So who is the incremental payer that wasn't paying before? In Google's case, it was simple to stop buybacks — or Meta stopping buybacks — and buy compute infrastructure instead. That doesn't have a huge effect on the market, but it has some. But stepping forward to 2028, where the hyperscalers are raising hundreds of billions of dollars of debt and their entire supply chain is raising hundreds of billions more — who pays for this? There are a few different ways. One is semiconductor companies like Nvidia, Broadcom, and the memory companies turning around and deciding to fund some of this CapEx.

There are the traditional infrastructure investors who gather capital and invest in infrastructure — instead of bridges, it's data centers. Then there's everyone in the economy who's realizing, "Maybe I shouldn't buy a home, or invest in credit that helps people buy homes, or buy government debt. I should just buy hyperscaler debt, or this data center's debt, or Anthropic's debt. Because Anthropic is willing to pay 20% rates for the incremental billion dollars to build their capacity — they know the revenue from it will be huge, and paying 20% is still better than renting from SpaceX at $50 billion a gigawatt." So you've got all of this contention for capital. If this happens, the whole world economy really shifts around.

Most powerfully, Antithesis gives you a live terminal into your system that you can use to perturb anything: kill a node or disable a feature, hit play and see what happens, then go back and try something else. In production, you often only get one shot at this kind of destructive analysis — if you restart a deadlocked service, the exact deadlock you needed to study disappears. With Antithesis, the original timeline is always reproducible, so you can test as many hypotheses as you need. And if you don't want to do the time traveling yourself, your agents can do it via the Antithesis API. Go to antithesis.com/dwarkesh to learn more.

We've been debating whether AI will trigger a sovereign debt crisis. The logic is this: you have a situation where very little investment turns into a lot of money.

What a problem to have. No, it's a huge problem for everybody else who can't turn a little money into a lot of money. The rate of return is incredibly high. Even at the data center level, you build a data center and rent it out to an Anthropic or an OpenAI for 10x what it costs you on a depreciated basis to build — you turn $1 into $2 or $10 or something by the end of the year.

That pushes interest rates higher. If rates rise for the entire economy, people are borrowing more and more money, competing against the lending the government would have done, other companies would have done, or you as a consumer or mortgage buyer would have done. That makes it more expensive for everybody else to borrow, with huge implications for tons of people.

Why the US will probably be fine

I think the US will be fine at the end of the day, because if the data centers are built in America, you can fundamentally just tax them. But the current tax system isn't set up for that: corporate income is less than 10% of federal revenues, while 80%-plus is payroll and income taxes — which will shrink as automation increases.

On the spending side, currently 20% of tax revenue goes toward servicing the debt. A lot of the debt is short duration, so it rolls over every five years.

Why are you laughing? Because it's things you've learned in the last month. Like it's any different for you — like you got a degree in financial economics. I didn't; the internet thinks I'm a beekeeper. A million people listen to this guy who just learned about debt this month.

Suppose interest rates rise 1%. Over a five-year basis, the fraction of tax revenue going to debt service goes from 20% to 25%. If rates rise 5 percentage points, that goes north of 40%. And if you account for the government borrowing $2 trillion every year, it goes from 40% to north of 60% of tax revenue just paying interest.

So the US should be fine because the tax base will increase if we let data centers get built here. Other countries are absolutely fucked, in my opinion. I was just looking at which countries have a lot of debt, very little tax revenue, and frequently serviced debt. Countries like Pakistan or Nigeria will be very badly hit in this new interest-rate regime.

Crowding out and the gigawatt constraint

This crowding-out effect is the reason it's not YOLO 1 billion gigawatts. You've got industries and countries that use a lot of debt — the impoverished countries that will just default, consumer packaged goods companies making what you see at Trader Joe's, telecom companies, banks. If market interest rates go up — not necessarily the government-set rate, but the spread between the federal rate and what everyone else charges, because Amazon wants to raise $100 billion of debt next year or whatever the number is, probably less — you end up with a really challenging problem: where does the cash come from?

Some level is funded by cash flows, and cash flows keep going up. But the logical thing to do is invest way more than your cash flows, because the returns in future years will be amazing. So you have this delta, and what's pushing down on the delta is everything else: regulations against data centers, consumers getting mad, politicians getting mad, regulations against AI, AI labs not releasing their latest models for safety reasons. Interest rates going up influence all of these things, bending the curve from what capitalism wants in pure, simple economics to what the complex system we actually have wants — lower and lower, so that not as many gigawatts as should be built will be built.

Well, the interest rate is part of capitalism, right? Yeah, but in the simple economic model versus the more complex reality we have.

How much debt will the buildout need?

What rate do you think Amazon or Anthropic or whoever will be issuing bonds at next year? If they do hundreds of billions of dollars of debt, what's the average rate? I don't think Amazon will do hundreds of billions of dollars of debt. In total — the hyperscalers and all the clouds combined. In the modeling we do, we have about $11 trillion of CapEx from 2024 to 2029. Total? Total. If you fund as much of it as you can with cash flows, you still end up with north of $5 trillion of credit that needs to be issued for this $11-trillion-plus buildout.

So you don't think AI revenue continues even 3x-ing year over year? AI revenue does go up. I just don't think it can go up forever without certain constraints being hit. Labs will have certain incentives, and labs aren't the ones building all the compute in many cases, even though they're increasingly trying to go that way. But they'll have all this cash flow. How much did you say the revenue will be?

"You think they'll not have that much revenue?" — No, I'm just saying that through 2029 there's something on the order of $11 trillion of CapEx. $6 trillion of that is funded with cash, and $5 trillion is funded with debt. If that's the case, $5 trillion of debt being raised across the whole ecosystem does make interest rates go up.

"Then what prevents that?" There are a couple of things. One: do labs increase their revenue per megawatt and keep inference allocations large? In which case, they're accumulating all the profit across the S&P 500, because everyone is paying to reduce their costs. Of course, their profits will also go up, but cash has to come from somewhere. So there's an upper limit on how fast their revenue can grow relative to the value they deliver into the world — there's a diffusion aspect to the technology. But ultimately, if labs' revenues keep going up, they can't cash-flow fund everything. The optimal scenario is to use credit as much as possible, because even if lab cash flows fund a lot, you want to build more than that. So some amount of credit does get built. Our current modeling has $5 trillion of credit and $6 trillion of cash-funded infrastructure investment through '29.

Even with that, it's not enough compute relative to demand growth from the AI models. So you get the obvious answer: revenue per megawatt keeps going up.

"How much do you think interest rates will increase by 2029 as a result of all this?" This is vibing a number, but growth in the world economy is going up a lot, so why wouldn't interest rates for Amazon go up from where they are today? This will be extremely vibed out, but Meta recently raised at 5 to 6%. I don't see why they wouldn't pay 8% — they'd happily pay 8%, because the return on the compute they're going to build is humongous. The market won't want them to, but they'll want to pay 8%.

The flip side: if they pay 8% versus the 5–6% they pay today — a 250 basis point increase — everyone else in the economy also pays 250 bps more, which causes a lot of things. Banks will scream, because if their credit spread goes up, their debt reprices faster than their assets, and they end up losing tons of money if their credit spread blows up.

The other consequence — a point you made — is that if interest rates rise, the discount rate increases, which means the discounted cash flows of all equities crater. So even though the stock market as a whole might be fine — the S&P 500 will be fine — any individual stock will probably have cratered in value, especially the Buffett, Berkshire-type stocks that pay good cash flows for 30 years. It's like, "Why would I pay this much for Johnson & Johnson?" They're seen as a stable stock — good cash flows, returned over time. Or a railway company. Why would I invest that much if my discount rate isn't 3% or 5%? It's now 8% or 10%.

For developing countries: Basil Halperin, a good friend and an economist, made the point that we'll see a second Volcker shock. In the '80s, to fight inflation, Fed Chair Paul Volcker raised interest rates by more than 5% — something like an 8% real interest rate. That caused some 40 countries, mostly in Latin America, to default in that decade. I think that will probably happen again.

"Okay, now we're getting into singularity talk." I think all of this happens before the singularity, by the way. We were talking about pre-singularity: interest rates rise 2–3%, et cetera. At some point, I think it's very likely the world economy will be doubling every single year. This is not happening in five years, but it'll happen eventually. There's a researcher, Damon Binder, who's done great work on this. If you look at input-output tables in a fully automated economy — what would it take to double the entire stock of things in the economy every single year? If the economy grows at 3% a year, rule of 70, that's twenty-something years. But right now we're bottlenecked by the fact that there are people, and you can't double people every year. In a world where you can also double the labor force every year, how fast can the economy grow? I think it could double every single year — at the very least, tens of percent per year.

"Okay. The rate of interest should be pretty close to the growth rate." It won't be exactly that because of consumption, but it should be pretty similar. So we'll enter a world, I think in the 2030s, where the rate of interest is tens of percent. Part of my brain says it might be hundreds of percent, but let's say at least tens of percent. Then: every country not involved in the production of AI defaults. Every stock that is not an AI stock is worth basically zero, because discounted cash flows are worth nothing. If the federal government can't figure out a way to tax AI, servicing the debt exceeds current tax revenue. And there are all these other effects I'm sure we're not even pricing in — you can't get a mortgage, et cetera.

"Fundamentally, what is happening in this world? This is all nerd speak. Let's step back — what's happening?" We'd be entering a totally different growth regime. The economy is basically saying: the opportunity cost of the government borrowing money to pay people pensions is extremely high now, because that money could be spent building a robot factory that builds a robot factory that builds a robot factory. The opportunity cost of capital is going to increase a ton. That's fundamentally the cause of all of these things we're talking about. As interest rates go up, equity markets get pummeled.

Even AI companies get hit. Some people who really believe in AI ask, "Why do Micron, Hynix, or Kioxia trade at 2 or 3 times earnings?" The answer is that if you're really AI-pilled, everything in the economy should trade at 2 or 3 times earnings. If you're not AI-pilled, then sure, memory makers are over-earning.

It's an argument for why — I think memory is going to do great — memory stocks shouldn't 10x again. Because if there's that much demand for memory, which means AI has caused a drastic change in the economy, then everything should trade at 2 or 3x multiples and the stock market should crash. In a sense, Meta trading at — I think they're like a $1.5 trillion company — is silly. They're worth way more than that, at least in a logical sense. You just look at their cash flows, all the infrastructure they're hoarding, and all the compute they'll be able to sell for crazy amounts of dollars per watt, either as tokens because their lab works, or just to Anthropic and OpenAI.

Ultimately it becomes a question of reallocating all the capital to AGI, and you do that by pricing everyone else out. So the limiter on AGI is not how fast the research engineers can crank the gears — it's how much the rest of the world lets that happen. They're going to regulate. They're going to increase interest rates. They're going to say, "No data centers," "Stop building fabs," and "Every company's equity value is tanking, so how can I pay for AI to improve my business?" Then Anthropic and OpenAI have to start building their own stuff. They're already designing their own chips, and it'll expand out — they're contracting their own data centers and building their own infrastructure over the next couple of years.

There's a real question of how this reallocation of the economy happens. There's a lot of downward pressure preventing a straight takeoff, even if the models were capable of it — and I think we both believe we're in a world where they are. But slow takeoff is, at least my hope, possible because of everything in the economy and the regulatory world: the government saying, "Don't release your models," or, "You can't even use your models internally that much" — which is going to happen soon. They're already saying you can't release your models.

The thing I'm most worried about is a singularity that external deployment is actually helping to prevent. So the fact that we're preventing external deployment is stupid. Right now it would just lead to more revenue, because the models are incapable of recursive self-improvement. But I'm worried about a world where it's 2030 and the government says, "We're going to wait six months before you can release your newest model to the public." Six months at 100x — in that time, the lab does recursive self-improvement internally, with all kinds of crazy stuff happening inside the company, while the rest of us are stuck with models that, at current pace, are years behind.

Here's my thought. Suppose the whole world gets in on this effort to slow down AI — and I don't think it's a conspiracy; it's written openly by every politician. Suppose they slow AI down by a year. If compute is increasing 2 to 3x every year, they've prevented a whole year of AI deployment, leaving you a year behind where you would otherwise have been. During recursive self-improvement, you're getting 3 to 6 years of AI progress in a single year. And they don't just limit compute — they also limit the lab's ability to release the model internally. We saw that: Anthropic had to stop giving Mythos to foreign employees for a while. I didn't know that was true internally as well. That's what they claimed. I thought it was just a different checkpoint that wasn't quite Mythos, but it was basically Mythos. Stuff like that won't be allowed either. The government is dumb, but they're not that dumb, I would hope.

Governments — at least the US government, which holds the cards here — are not going to want Anthropic using Mythos 4 internally. They're going to say, "Hold on, slow down," for all these regulatory reasons. Everyone who's elected is going to hate AI — even the people who are elected already hate AI, and so do all the constituents. I bet at some point your parents are going to call you and say, "Dwarkesh beta, you're doing a terrible job. You're making AI progress happen faster."

Because of my podcast I'm accelerating AI progress? Maybe — you educate people, and maybe smarter people progress AI faster. Anyway, you're going to have real-world constraints on the progress, development, and deployment of AI. Even though it will happen eventually, we could tear ourselves apart before we get there.

The Jane Street team follows frontier LLM research closely, and a relatively common intern project is adapting a recent paper to financial markets, which come with their own set of gnarly problems. "Ultimately, we're trying to model thousands of interconnected irregular time series. The signal-to-noise ratios are extremely low because we have a lot of competitors trying to do the same. So we have this adversarial, non-stationary, extremely high-dimensional problem that we're trying to solve." To be clear, you don't need to know anything about finance to be a good fit — as long as you have a background in ML research, Jane Street can teach you the rest. Their 2027 internship applications are open now; apply at janestreet.com/dwarkesh.

One thing I find crazy about these scenarios

What I find crazy about these scenarios is just how much of the world's future labor supply ends up in very few companies, and how fast that labor supply grows year over year. If compute at the frontier in FLOP terms is growing 4–5x a year, and the compute required to achieve a given level of capabilities is decreasing 3x a year, then the effective AI population size at the frontier labs is increasing roughly 10x year over year. That doesn't matter much right now, because AIs aren't yet good enough to do full jobs or act as autonomously as people. But if the current trend continues, you get a world where OpenAI goes from, say, 10 million AI laborers this year to 100 million the next year, to a billion the year after that. Pretty soon, even if compute scaling slows down, it doesn't take many more years before each company individually has more labor equivalence than there are people on Earth. I think that's very plausible by the end of this decade: more AI labor, more effective population, within a single lab than there are people on Earth.

We talk often about centralization of power because of nationalization or whatever, but we don't think enough about the fact that we're moving very fast into a regime where most "people" — in terms of work output — are concentrated within two labs that are consuming more and more of the world's compute. If these AIs are misaligned, then most of the world is misaligned, basically, because most of the world's minds are there. But even if they're not, very few companies have a lot of influence or control. There was the whole spat recently where Gavin Baker said, "Dario believes there's only going to be one company in the world," and then Sholto and Dario came out and said, "No, no, no, we didn't say that." But ultimately, if you believe in RSI — if you believe the labs are the most effective users of compute and can generate the most value from it — the only thing that's going to happen is centralization of compute. If you believe in AI researchers, RSI, AGI, then all of this follows. And this is true even without RSI: the effective population of the frontier is currently increasing 10x year over year for a given level of capabilities. So once you reach the level of a very competent remote worker, software engineer, or researcher, the population of those is increasing 10x a year at the current rate of capabilities growth. Once you add RSI, it's even crazier — maybe 100x or 1,000x a year, or their intelligence increases while the population doesn't, or some mixture of the two.

What world do you see, Dwarkesh, where everything is not centralized? Because it seems to me that every force is screeching towards centralization, and that's scary as hell. I would love for it not to be centralized completely.

It's so hard to think about the future, but I agree with you. I think the fundamental problem is that AI training has huge economies of scale, because any effort spent training an AI for a specific skill or body of knowledge gets amortized across billions of sessions or users. That's one effect. The other is that if you're slightly ahead in the AI race and compute is in shortage, you can charge a much higher markup because you can better economize this scarce resource. So there are two effects that give more and more to whoever is ahead in the AI race. There may be more: if models learn from deployment, a model deployed much more widely gets much more real-world data. Whether it's user deployment and continual learning, training economies of scale, or incremental progress where the best AI model helps you make the next best one — RSI — all of these point to centralization.

I think one of the big intellectual projects we should spend time on — or at least I will — is: what is a vision of a decentralized, broadly empowered future after AGI that takes these economies of scale seriously? The alternative vision is that the government controls it, and maybe you think you can trust the government more because it's not a private corporation. I don't trust the government, and I don't trust Dario, and I don't trust Sam. That's a problem, right? Obviously it's very easy to be wrong about the future — you don't anticipate a key effect or something that changes everything. But ex ante, it's very hard to see how we avoid a scenario where we have to choose one source of centralization.

It's why capitalism worked: decentralized decision-making and decentralized power. It's why super-centralized capitalist economies actually grew slower than super-decentralized ones, to some extent — you have to have rule of law and all that. But AI flips this on its head: you end up concluding that private ownership is probably not the most efficient economy, and therefore grows slower than a centralized AI economy.

Well, it's still private ownership, but how many firms are really involved in this share of the economy? Maybe 2% right now — a trillion dollars divided by 30. Nvidia is a huge share of it, along with Anthropic, OpenAI, and the hyperscalers. Obviously other firms are involved, but a large share of the AI stuff is happening at very few companies. So it can be private property, but with very few companies involved — that's just what the market structure is doing.

So what can prevent it? I don't know. Unless AI progress slows down, unless governments regulate the hell out of it, this is what happens.

In that case, we're headed for a world where either we have super concentration of resources and we pray that one company gets everything right, or governments and people slow everything down, hopefully producing some slowdown of progress and more of a balance of power. Even as we go towards AGI, ASI, RSI, everything along the way will still lead to someone capturing more resources. So it's hard to find a framework in which AI doesn't lead to super concentration.

The one positive thing here is that today Anthropic does not capture most of the value. We can talk all we want about how they went from $20 million per megawatt to $100 million per megawatt, but they're still paying $13 million for a lot of the compute they buy. At the end of the day, the reason they've gone to $100 million per megawatt is that Jane Street is capturing $300 million or $500 million per megawatt. Or Dwarkesh, from researching his podcast and learning about credit — how many dollars per megawatt is he capturing? How much can you use? Tough call.

I think that's the one saving grace: the rest of the economy maybe profits so much more from Anthropic—

No, but the whole logic you were laying out earlier — reallocating inference to AI R&D — rests on the returns to labor inside AI labs being much higher than the returns outside.

Yes. This is my cope, I agree. In all scenarios of the world — there are 80,000 worlds and in only one of them Anthropic doesn't own the whole world — power concentrates because I don't want to send the tokens outside; they're more valuable inside. So it's the same thing. Why would I let Jane Street make all this money off of these degenerate options traders? Hey, they're a sponsor, come on.

Jesus Christ. No, I think it's great. It's a good value for the world to make it an efficient market.

Jane Street making all this money off of getting the worldview correctly, off of degenerate options traders, whatever it is — why would Anthropic allocate compute to that? If Jane Street's end monetization per megawatt is $200 million, so they're willing to pay Anthropic $100 million — well, what if Anthropic can just generate hundreds of millions of dollars per megawatt by using that compute internally? That's what's happening.

On that somber note, I guess we'll meet again when the RSI is officially kicked off. You're not going to have me on your podcast again for like two months? Alright, cool. Thanks, dude.