Anthropic and OpenAI will absorb half of new compute by 2027

1. Margin flip funds training — ▶ 0:37

   CapEx is a little over $1 trillion this year and more than $2 trillion in 2028, and a third of new compute is for OpenAI and Anthropic. Base compute costs $10-15 million per megawatt. GPT-4 on Hopper ran at negative gross margin. Anthropic's revenue has now reached as high as $50 million per megawatt: spend $10 on inference, make $50, and put the profit into training. Anthropic turned profitable in Q2, and OpenAI could in Q3, so growth is increasingly funded by their own revenue.

2. Two labs take half — ▶ 3:50

   OpenAI began the year at 2 gigawatts and Anthropic below 2. Both end above 5, a 3-4x jump. That is about 30% of compute added this year, 40-50% next year, and half of all incremental compute by the end of 2027. SpaceX will lease capacity to them, OpenAI is building its own chips, and Anthropic buys TPUs deployed with Fluidstack. New chips deliver 3-5x performance per watt, so by late 2028 the two labs control most of the world's usable flops.

3. Fab capex math — ▶ 7:18

   One gigawatt of Vera Rubin needs 55,000 N3 wafers, 6,000 N5 wafers and 170,000 DRAM wafers. Tooling to produce a gigawatt every year costs $3-4 billion, or about $6 billion with cleanrooms and shell. A gigawatt generates roughly $100 billion of revenue a year. Over five years, $6 billion of fab CapEx yields over $1 trillion of end AI revenue. Even after halving it for middlemen, the gap between fab CapEx and end revenue is 100x. This is the "turn $1 into $100" problem.

4. Supply responds slowly — ▶ 9:39

   Supply is a bullwhip: it takes a long time for the signal to reach the tail. Carl Zeiss now says it must make enough mirrors for 100 EUV tools a year by the end of the decade, up from what it believed earlier this year. Patel says 100 is still right for 2030 on the current path. Next year's CapEx exceeds $2 trillion, while wafer-fab equipment does about $200 billion. Nobody hands Zeiss $10 billion to expand, because the world is capital constrained. An EUV tool bought for $400 million could be resold for over $1 billion.

5. Compute must reprice — ▶ 13:53

   Reaching 100 gigawatts across both labs by 2028 means taking 70-80% of incremental compute. Patel calls that very aggressive and is unsure how markets would behave. Anyone can profit at $10-15 million per megawatt: rent a GB300 rack, download Kimi weights and vLLM or SGLang, list on OpenRouter. That is already pushing prices up. To win 70% of world compute, the labs must pay $25, $30 or $50 million per megawatt. Bedrock-served Anthropic models also count as Anthropic compute in his accounting.

6. Regulation as the brake — ▶ 16:15

   Revenue per megawatt only climbs if the best models ship. OpenAI is not releasing Astra and stopped training for two weeks. Anthropic is withholding what its safety assessment calls Model 2, believed to be the next Mythos. If the best models stay internal, revenue per megawatt stalls or even declines as other models become competitive, and the labs lose the ability to outbid others. The regulation labs advocate slows them more than it slows open-source Chinese models. Without safety constraints, they would earn $100 million per megawatt or more and pay $50 million.

7. Value capture rotates — ▶ 20:16

   Most value still goes to users. Jane Street and Meta, rumored at one point to be up to 10% of Anthropic's business, earn more from tokens than Anthropic does. A year ago hardware took all the margin while model layers sold tokens at a loss. In 2023 memory makers earned nothing on HBM. Now memory captures more than TSMC. SpaceX showed that Anthropic and Google would pay $25-40 million per megawatt. Meta and SpaceX build without customers using their balance sheets, then choose between internal use and selling at "crazy margins".

8. End-2027 pricing dispute — ▶ 25:40

   Patel puts Anthropic or OpenAI revenue per megawatt at $50 million-plus by the end of 2027, perhaps $70-80 million blended. Dwarkesh calls that low given the jump from GPT-4o to Mythos in 18 months. Patel replies that the best model was trained in February and is not even used internally to optimize inference. Only a slowdown caps it near $100 million per megawatt. Price rises cascade: Nvidia, SK Hynix, Micron and Samsung raise, memory and substrates quickly and TSMC slowly. New York's ban, Texas's moratoriums and Ohio's property-tax rule cut supply.

9. Inference share shrinks — ▶ 30:31

   A non-consensus call: labs will allocate less compute to inference over time, with most going to forward passes for training. At $30-40 million per megawatt a lab puts 40% into inference. At $60-70 million it will not keep 40% to fund dividends and buybacks. Executives and boards will choose to build AGI. The evidence is already visible. Anthropic adds compute every month, yet no longer adds $25 billion of ARR monthly, so marginal megawatts are going to R&D. Running the best model internally is worth more than selling it.

10. China's compute gap — ▶ 34:37

    In 2022 the US added 45-50% of world compute and China 30-35%. Today 70% of deployed watts are American and China is under 10%. Patel expects China at 30 gigawatts or less in 2028, with 5-10 gigawatts of domestic chips that year, all worse than Nvidia's or Google's. A 50 gigawatt year in 2029 is plausible but worth about 20 gigawatts of American chips. Export controls look like a notable success. The caveat is that China's semiconductor subsidies exceed the rest of the world's combined, so a slower takeoff lets it catch up.

11. Training can't absorb everything — ▶ 40:50

    Lab budgets split about 60% training and 40% inference, with training being 50% research and 10% development. Mythos pre-training used under 200 megawatts for about two months, and RL used less at a single site, while the lab held multiple gigawatts. Most compute goes to research experiments, because multi-site training and RL are hard to coordinate and more rollouts do not guarantee better models. Leading Chinese labs run 100-200 megawatts, and for now the gap barely matters. Patel expects automated researchers and continual learning to blur research and training.

12. A $10 trillion bill — ▶ 42:51

    At 100 gigawatts a year, current prices mean $5 trillion of CapEx annually. Power plants and buildings are paid years ahead, so the real figure is $7-10 trillion, near a tenth of the world economy by 2030 and a quarter to a third of the US economy. His model has $11 trillion of CapEx from 2024 to 2029, with $6 trillion from cash and $5 trillion from credit. Hyperscalers already spend all their cash and raise debt. The marginal funders become Nvidia, infrastructure investors and households that skip homes and government bonds for hyperscaler debt.

13. Rates and crowding out — ▶ 48:42

    $5 trillion of debt lifts rates. Patel guesses Meta, which recently raised at 5-6%, would happily pay 8%, a 250 basis point rise that reprices everyone. Banks' debt reprices faster than their assets, and long-duration stocks like Johnson & Johnson fall. If rates rise 5 points while the US borrows $2 trillion a year, interest rises from 20% to over 60% of tax revenue. He thinks the US survives by taxing data centers. Pakistan and Nigeria default, in a second Volcker shock. He flags the rate number as vibed.

14. The world as limiter — ▶ 1:02:45

    The limiter on AGI is "how much does the rest of the world let that happen", not research speed. Rates, data center bans and fab stoppages all bend the curve. Patel hopes for slow takeoff but fears a 2030 where the government delays release by six months while labs self-improve internally and everyone else sits years behind. Dwarkesh argues a one-year slowdown costs 3-6 years of progress during RSI. Patel adds that governments will also restrict internal use, as when Anthropic briefly stopped giving Mythos to foreign employees.

15. Centralization has no exit — ▶ 1:07:54

    Frontier compute grows 4-5x a year while compute per capability falls 3x, so effective AI population at the frontier grows about 10x a year, even without RSI. OpenAI could go from 10 million AI laborers to 100 million to a billion, exceeding the human population within one company by decade's end. Economies of scale in training, markups for the leader and learning from deployment all point one way. Patel's saving grace: users like Jane Street capture $300-500 million per megawatt. Dwarkesh counters that labs will rationally keep tokens inside.

Compute is priced far below what the best buyers can earn from it, and almost everything else in this market follows from that gap. A megawatt of compute costs $10-15 million. A year ago, serving GPT-4 on Hopper GPUs earned OpenAI a negative gross margin. Now Anthropic's revenue has reached as high as $50 million per megawatt. Spend $10 on inference, make $50, and put the profit into training. Dylan Patel sees the path running to $100 million per megawatt and beyond. That is a gap of 4x or more between what compute sells for and what the labs extract from it. It is why Anthropic turned profitable in Q2, why OpenAI may follow in Q3, and why both are shifting from venture-funded losses to funding growth from their own revenue.

The surplus is not all the labs'. Most of the value still goes to users. Jane Street makes far more from the tokens it buys than Anthropic earns on them. Meta, rumored at one point to be 10% of Anthropic's business, gains more from better ad algorithms and engagement time 5% longer than Anthropic does from selling to it. Value capture has also rotated. A year ago the hardware chain took the margin while model companies lost money on tokens. In 2023 memory makers earned nothing from HBM despite the value it delivered. Now memory captures more than TSMC. The labs' share has ballooned, but the users' share is still larger.

A lab that earns more per megawatt can outbid everyone else for it. That is the centralization mechanism, and the numbers are already moving. OpenAI began the year at 2 gigawatts and Anthropic below 2. Both finish above 5. They take roughly 30% of the compute added this year, 40-50% next year, and half of all incremental compute by the end of 2027. Raw watts understate it. GB300s, TPUv7s and Trainium3s deliver 3-5x the performance per watt of the prior generation, and the labs hold most of the new ones. By late 2028 Patel expects the two to control most of the usable flops in the world. The market has already confirmed the price signal. SpaceX showed that Anthropic and Google would pay $25 million or $40 million a megawatt, letting it recoup CapEx in a year. Meta and SpaceX can build without a customer because they have balance sheets, then choose between internal use and renting at wide margins. Most clouds cannot. They need a signed customer before they can raise the credit.

The obvious objection is that supply will respond. It will, but slowly, and the arithmetic shows how big the prize is. One gigawatt of Vera Rubin needs 55,000 N3 wafers, 6,000 N5 wafers and 170,000 DRAM wafers. The tooling to make a gigawatt of chips every year costs $3-4 billion, or about $6 billion with cleanrooms and shell. A gigawatt earns about $100 billion a year. Over five years, $6 billion of fab CapEx produces more than $1 trillion of end revenue. Halve that for middlemen and the discrepancy is still 100x. Yet the signal travels like a whip, and it takes "a long time for the whip signal to get to the tail end." Carl Zeiss now says it must build 100 EUV tools a year by the end of the decade, more than it thought necessary earlier this year. Patel thinks 100 is still about right for 2030. Next year's total CapEx exceeds $2 trillion, while the wafer-fab-equipment chain does about $200 billion. Nobody writes Zeiss a $10 billion check, because the world is capital constrained. So scarcity persists, and price increases propagate backward. Memory makers and substrate suppliers raise prices fast. TSMC raises them slowly.

Where does the surplus go? Patel's non-consensus answer is that labs will shrink the share of compute spent on inference. Most people assume inference dominates. He expects most compute to go to forward passes for training instead. Today a lab might put 40% into inference. If revenue per megawatt doubles, why keep that share and pay out dividends and buybacks, when the alternative is to build AGI? He says executives and boards will choose the latter, and that it is already visible. Anthropic adds compute every month, yet it no longer adds $25 billion of ARR monthly, so each marginal megawatt is going to R&D. There is a physical limit. Mythos pre-training peaked under 200 megawatts for about two months, though the lab had multiple gigawatts. Most compute goes to research experiments, because multi-site training and RL are hard to coordinate. Automated researchers and continual learning, he expects, will blur that line and pull more into training.

Which raises the question of who pays. Patel's model has $11 trillion of CapEx from 2024 to 2029, with $6 trillion from cash flow and $5 trillion from credit. At 100 gigawatts a year, annual CapEx is $5 trillion at current prices, and $7-10 trillion once power plants and buildings are paid for years ahead. By 2030 that approaches a tenth of the world economy, or a quarter to a third of the US economy. Hyperscalers already spend everything they make and raise debt. Debt on that scale lifts rates. Patel guesses that Meta, which recently raised at 5-6%, would happily pay 8%, and he calls the number vibed. A 250 basis point rise reprices everyone. Banks' liabilities reprice faster than their assets. Stable cash-flow stocks like Johnson & Johnson fall as discount rates rise. If rates rise 5 points while the US keeps borrowing $2 trillion a year, interest climbs from 20% to over 60% of tax revenue. He thinks the US survives, because data centers expand the tax base. Pakistan and Nigeria do not. This crowding-out is why the buildout does not simply run to infinity.

The second brake is regulation, and here the two disagree on the forecast. Patel puts end-2027 revenue per megawatt at $50 million-plus, perhaps $70-80 million blended. Dwarkesh thinks that is low, given how far models moved from GPT-4o to Mythos in a year and a half. Patel's reply is that the best model in the world was trained in February and is not being deployed, even internally. OpenAI is holding back Astra and paused training for two weeks. Anthropic is withholding what is believed to be the next Mythos. State bans and moratoriums on data centers cut supply and raise costs. If the best models stay inside, external revenue per megawatt stalls and the labs cannot outbid anyone. Without safety constraints he sees $100 million per megawatt or more. Meanwhile everyone up the chain raises prices because they can, which he calls super inflationary.

China is the clearest test of whether any of this can be slowed. In 2022 the US added 45-50% of the world's compute and China 30-35%. Today 70% of deployed watts are American and China is under 10%. Patel expects China at 30 gigawatts or less in 2028, with 5-10 gigawatts of domestic chips that year, all worse than Nvidia's or Google's. A 50 gigawatt year in 2029 is plausible but worth perhaps 20 gigawatts of American chips. On that weighting, the leading American lab in 2028 holds more compute than all of China in 2029, if nothing slows the US labs. Dwarkesh, who pressed the opposite case with Jensen Huang, now calls export controls a notable success. Patel adds a caveat. China subsidizes semiconductors more than the rest of the world combined, so a slower takeoff lets it catch up. And Chinese labs run 100-200 megawatts yet stay close to the frontier, so compute gaps do not matter much yet.

The last problem has no answer in the conversation. Frontier compute grows 4-5x a year while the compute needed for a given capability falls 3x, so the effective AI workforce at a frontier lab grows about 10x a year, with no recursive self-improvement required. That could take OpenAI from 10 million AI laborers this year to 100 million, then a billion, more than there are people on Earth inside a single company by the end of the decade. Training has economies of scale, the leader can charge more for scarce compute, and deployment feeds back learning. Each force points the same way. Patel says the limiter on AGI is "how much does the rest of the world let that happen". Governments could slow it by raising rates, banning data centers or delaying releases. He hopes for a slower takeoff. He also fears a 2030 in which a government holds the newest model back for six months while the lab self-improves internally and everyone else sits years behind. His one consolation is that users like Jane Street still capture more than Anthropic, $300 million or $500 million per megawatt. Dwarkesh's reply undercuts it: if returns inside the lab are higher, the labs will keep the tokens. Dwarkesh trusts neither the government nor the lab chiefs, and neither man can name a decentralized path that takes the economies of scale seriously.