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

Summary

• Dylan Patel’s central forecast: Anthropic and OpenAI go from about 30% of compute added this year to 40%–50% next year, and “by the end of next year half of the incremental compute is already going to Anthropic and OpenAI.” OpenAI began this year at 2 GW and Anthropic below 2 GW; both end above 5 GW. Because GB300, TPU v7 and Trainium 3 deliver 3–5× more performance per watt, the two labs could control most usable FLOPs by the end of 2028 if the trend continues.
• The unit economics have flipped from negative gross margin to a 3–5× spread per megawatt. Compute costs $10–15M/MW; Anthropic’s revenue has reached as high as $50M/MW, it started turning profitable in Q2, and OpenAI is believed capable of turning profitable at some point in Q3 on Codex and GPT-5.6. The flywheel: “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.”
• Compute itself must reprice: anyone can make money at $10–15M/MW by deploying a GB300 rack with Kimi, vLLM or SGLang on OpenRouter, so labs may need to pay $25M–$50M/MW to take 70%–80% of supply. SpaceX and Meta are hoarding compute on their balance sheets and can choose between internal use and selling to labs at high prices. Even so, Dylan expects most compute to transact below $20B/GW through the end of next year because it is generally contracted and financed before it is built.
• Dylan’s non-consensus call: labs will allocate less compute to inference over time and more to research and training, including building AGI. Anthropic’s compute additions keep rising while its monthly revenue additions have plateaued after January’s spike, implying a greater share of marginal capacity is going to R&D. The current budget is roughly 50% research, 10% development and 40% inference; Mythos’s pretrain used less than 200 MW for about two months, while RL was lower at any one site or moment, though total sequential compute could be higher.
• Regulation and deployment constraints, not capability alone, may cap revenue per megawatt. Dylan cites OpenAI saying it is stopping training for two weeks and Anthropic not releasing what its safety assessment says is Model 2, widely believed to be the next Mythos. Dwarkesh says Mythos has been neutered and unavailable to “us” for inference optimization; Dylan says Astra is not widely deployed internally. Absent such constraints, Dylan says labs could generate $100M+/MW and pay $50M/MW.
• Export controls have so far widened the US–China compute gap: less than 10% of newly deployed data-center AI watts are in China today, and China could have 30 GW or less in 2028, on inferior chips. Dwarkesh speculates that even 50 GW of mostly domestic Chinese chips in 2029 might be worth only about 20 GW of US-quality compute. Dylan says 50 GW is reasonable, while export controls, the MATCH Act and domestic equipment production remain important; Dwarkesh notes that a slower takeoff would let China catch up.
• The financing math points to a rate shock: Dylan’s model has about $11T of CapEx from 2024–2029, with $6T cash-funded and north of $5T financed through credit. Meta raised recently at 5%–6% and would happily pay 8%; Anthropic would pay 20% on incremental debt if that is cheaper than renting capacity from SpaceX. Higher spreads reprice the economy, crush non-AI equity multiples, pressure debt-heavy countries toward default and could eventually produce interest rates in the tens of percent in the 2030s.
• Both discuss centralization as the likely endgame, but without a settled solution. Dylan estimates effective frontier AI labor grows about 10× annually, making it plausible that one lab has more labor-equivalent capacity than there are people on Earth by decade’s end. Dwarkesh emphasizes training scale economies and scarcity markups; Dylan adds deployment data and RSI. Dylan says, “I don’t trust the government, and I don’t trust Dario, and I don’t trust Sam,” proposes thinking about a decentralized post-AGI future, and concludes: “There’s 80,000 worlds and in only one of them, Anthropic doesn’t own the whole world.”

Deep dive

1. Lab economics flipped from venture-funded losses to $50M-per-megawatt machines

• Dylan’s baseline: about a third of compute coming online this year is ultimately for OpenAI and Anthropic; total AI CapEx is “a little bit over a trillion dollars” this year, more than $2T by ’28, with labs scaling from tens of billions of annual spend toward “trillions of dollars a year even towards the end of the decade” per contracts already signed.
• The margin flip is the story: base compute cost runs $10–15M/MW. GPT-4 on NVIDIA Hoppers was negative gross margin for OpenAI, but GPT-5.6, Opus 5 and Fable 5 monetize far past the incremental cost — Anthropic’s revenue has reached “as high as $50 million per megawatt.” Anthropic started turning profitable in Q2; OpenAI is believed capable of turning profitable at some point in Q3 on Codex and GPT-5.6.
• The self-funding flywheel, verbatim: “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.” New capital still comes in, but growth is increasingly funded by revenue rather than capital injections.

2. Two labs take half of the world’s incremental new compute by the end of next year

• OpenAI began this year at 2 GW and Anthropic below 2; both end above 5 — a 3–4× increase, according to Dylan, accounting for about 30% of compute added. Signed deals put them at 40%–50% next year: “it’s really by the end of next year when half of the incremental compute is already going to Anthropic and OpenAI.”
• The quality multiplier: the world adds roughly 30 GW this year, 50 next, 70 in ’28 and 90–100 in ’29 — but new watts using GB300s, TPU v7s and Trainium 3s deliver 3–5× more performance per watt. Thus, by the end of 2028 the labs could be “controlling most of the usable FLOPs in the world on their own,” if the trend continues.
• The builders change: SpaceX becomes a major entrant leasing to whoever pays most; OpenAI is designing its own chips, while Anthropic is purchasing Google TPUs and deploying them with Fluidstack. Accounting caveat: Amazon serving Anthropic models through Bedrock counts as Anthropic compute in Dylan’s worldview because it counts as Anthropic revenue.

3. The 100× fab arbitrage, and why the mirror bottleneck clears slowly

• Dwarkesh’s math, based on Dylan’s wafer-fab-equipment model: one gigawatt of Vera Rubins requires roughly 55,000 N3 wafers, 6,000 N5 wafers and 170,000 DRAM wafers. The model estimates roughly $3B–$4B in tooling, or about $6B including fab cleanrooms, shell and related CapEx, to produce a gigawatt every year.
• Each gigawatt generates roughly $100B of revenue annually. Over five years, the first gigawatt gets five years of revenue, the second four, and so on, producing more than $1T of end-AI revenue from about $6B of fab CapEx. After Dwarkesh conservatively removes half for other participants, he still sees roughly a 100× discrepancy.
• Dylan’s answer to “won’t capitalism fix this?”: “It’s a whip. It takes a long time for the whip signal to get to the tail end.” Carl Zeiss only recently accepted that it needs enough mirrors for 100 EUV tools a year by 2030 — “it takes so long to build.” His arbitrage: anyone who could buy an ASML EUV tool for $400M should, then resell it “north of a billion dollars.”
• Why the entire stack does not immediately expand: lab revenue is hundreds of billions next year against roughly $2T of total CapEx, so labs cannot yet fund the whole buildout from cash flow. Dylan does not expect a private-equity-style full-stack expansion this year, next year or the year after because “the world is capital-constrained.”

4. Compute repricing: anyone can make money at $10–15M/MW, so labs may pay $25–50M

• The floor, demonstrated: “Go get a GB300 rack, go download the Kimi weights, go download vLLM or SGLang… Go put it on OpenRouter… You’ll start generating more revenue than you’re paying for the compute.” That is already pushing $10–15M/MW pricing upward; reaching 100 GW combined by 2028 requires labs to pay $25M, $40M or even $50M/MW.
• SpaceX exploited the contract-first regime by having compute available before needing the usual customer commitment. Dylan says Elon could sell capacity at $25M–$40M/MW and recoup the entire CapEx in a year. Dwarkesh floated a separate example of “B300s or whatever” SpaceX might have sold for $40B/GW to Google.
• Meta and SpaceX are the “only plausible #3” candidates because they hoard compute on real balance sheets and can choose between internal use and selling to Anthropic or OpenAI at high margins.
• Dylan’s near-term sobriety despite all this: most compute still transacts below $20B/GW even at the end of next year because most compute is contracted before it is built — the customer commitment enables the credit-market financing. The supply chain rebalances as a bullwhip: memory and substrates reprice quickly, while TSMC moves more slowly.

5. Where the value actually goes: users, not labs

• The stack today: end users capture the most. Jane Street, with an exclusive OpenAI contract for GPT-5.6 Ultrafast mode, generates “way, way, way more value” from the tokens it buys than Anthropic generates in profit; Meta, once rumored to account for as much as 10% of Anthropic’s business, monetizes ad-algorithm and engagement gains far beyond what it pays. The app layer has generated very little value so far.
• The history: a year ago the model layer ran negative gross margins on VC money while the hardware supply chain captured the value. In 2023, memory companies earned little from HBM despite its importance; now, Dylan says, TSMC captures less value than the memory companies. “The value capture’s shifted around a lot.”
• The intuition pump for the ceiling: a gigawatt sustaining a million white-collar workers at $100K each represents $100B — “that’s actually surprisingly low.” Full AGI could imply “many hundreds of billions of dollars per gigawatt.”

6. Regulation and deployment constraints may cap revenue per megawatt

• Dylan cites OpenAI saying it is not training models for two weeks and Anthropic not releasing what its safety assessment says is Model 2, widely believed to be the next version of Mythos. Dylan’s broader complaint is that “the best model that exists in the world was trained in February.”
• Dwarkesh says Mythos has been neutered and unavailable to “us” for inference optimization and other uses. Separately, Dylan says Astra is not even widely deployed internally. The distinction matters: the transcript does not establish that Mythos is unusable internally.
• Dylan’s end-’27 revenue forecast is highly conditional: $50M+/MW, potentially $70M–$80M blended, “if not higher.” Dwarkesh pushes back by extrapolating from recent progress and asking what happens if a GPT-4o-to-Mythos-2-sized leap occurs by the end of 2027, while noting that Mythos and Mythos 2 have not been fully released.
• The unconstrained scenario, verbatim: “in a world where safety doesn’t matter… They can start generating $100 million per megawatt or more, and they can pay $50 million a megawatt. No one else has any logical reason to do anything with their compute besides say, ‘Please, Dario, take everything off of my hands.’” Supply-side restrictions add drag: New York banning data centers, Texas moratoriums and Ohio proposals requiring data centers to pay property taxes within a radius.

7. The non-consensus call: inference share shrinks — the marginal megawatt builds AGI

• Dylan, explicitly against the crowd: “The labs are going to allocate less and less compute to inference over time… very non-consensus.” At $60M–$70M/MW of monetization, the choice is dividends and buybacks or using the capacity to build AGI. Dylan says the obvious answer for Anthropic and OpenAI, including their boards, is the latter because it is more profitable.
• Evidence it is already underway: Anthropic adds more compute every month, yet revenue additions spiked in January and then plateaued. “The marginal megawatt they’re getting is going as a higher percentage to R&D than it is to inference.”
• Budget anatomy: roughly 50% research, 10% development and 40% inference. The Mythos pretrain ran below 200 MW for roughly two months; RL used less at any single site or moment, though total sequential compute could be higher. Multi-site coordination, colocation and rollout limits keep individual training runs smaller while research consumes the remaining fleet. Automated researchers and continual learning could push the mix further toward training.

8. Export controls have so far widened the US–China compute gap

• The reversal since 2022: the US was adding about 45%–50% of world compute in 2022 and now accounts for about 70% of newly deployed watts. China is below 10% of newly deployed data-center AI watts, relying on smuggled chips, chips TSMC made for companies it thought were not Huawei but that ended up being Huawei, and HBM shipped by Samsung.
• SMIC and CXMT fabs begin coming online in 2027–2028, with domestic production reaching many millions of units a year and adding roughly 5–10 GW of domestically produced chips in 2028 alone. Those chips are expected in the discussion to be worse than NVIDIA’s or Google’s 2028 chips, so raw gigawatts overstate China’s effective compute.
• Dwarkesh changes his mind after steelmanning cooperation in his Jensen interview: “I didn’t realize the compute situation was as fucked as you’re saying… I think that’s actually a notable success.” Leading Chinese labs have at most 100–200 MW total, with ByteDance Seed an outlier, versus Anthropic’s more than 5 GW by year-end.
• Dylan says 50 GW for China in 2029 is “completely reasonable.” Dwarkesh speculates that if most of it is domestic, 50 GW might be worth only about 20 GW of American-quality compute. The MATCH Act, further tool controls and China’s ability to build domestic equipment remain important; Dylan says Chinese semiconductor subsidies exceed those of the rest of the world combined. Dwarkesh adds that a slower takeoff would let China catch up drastically.

9. $11T of CapEx, north of $5T of credit, and a second Volcker shock

• Dylan’s model has about $11T of CapEx from 2024 through 2029, with roughly $6T funded by cash and north of $5T by credit. Headline critical-IT spending understates the total: 100 GW at current prices would cost about $5T annually, and prebuilding power plants, data centers and downstream capacity could make it more like $7T–$10T. Power plants are roughly 30-year assets; data centers are 15–20-year assets.
• Hyperscalers have funded much of the buildout through cash flow, redirected buybacks and debt. Google, Meta, Amazon and eventually Microsoft can raise debt, while semiconductor companies, infrastructure investors and the wider credit market become additional sources of capital; buybacks are a possible source of funding, not a universally stopped program.
• The rate mechanics: Meta recently raised at 5%–6% and “would happily pay 8%”; Anthropic would pay 20% on incremental billions because it is still cheaper than renting capacity from SpaceX at $50B/GW. A 250-basis-point spread rise reprices everyone — banks lose as liabilities reprice faster than assets, and higher discount rates crater non-AI equity valuations: “Why would I pay this much for Johnson & Johnson?”
• Dwarkesh’s sovereign math, with Dylan ribbing him for “things you’ve learned in the last month”: corporate income is under 10% of federal revenue, while payroll and income taxes could shrink with automation. In Dwarkesh’s hypothetical, a 1-point rate rise takes debt service from 20% to 25% of tax revenue over five years; a 5-point rise takes it above 40%, and continued $2T annual borrowing pushes it above 60%. He thinks the US could remain fine if it taxes American data centers, while Pakistan and Nigeria could be badly exposed.
• Basil Halperin’s “second Volcker shock” analogy points to a repeat of the 1980s pattern, when Paul Volcker raised real rates to roughly 8% and around 40 countries, mostly in Latin America, defaulted. The singularity extension, via Damon Binder’s input-output work: with an AI-doubling labor force, the economy could eventually double yearly, putting 2030s interest rates in the “tens of percent.”
• Dylan’s market kicker: if you are truly AI-pilled, “everything in the economy should trade at 2 or 3 times earnings.” Memory stocks should not necessarily 10× again, and Meta at roughly $1.5T is “silly” because its cash flows and hoarded compute could be worth much more.

10. Every force screeches toward centralization — and the trust problem has no answer

• Dylan estimates that frontier FLOPs are growing 4–5× annually while the compute required for a given capability falls 3× annually, implying effective AI population grows about 10× per year. OpenAI could go from 10 million AI laborers this year to 100 million next year and 1 billion the year after; it is plausible that one lab has more labor-equivalent capacity than there are people on Earth by decade’s end.
• Dwarkesh identifies the scale economies: training effort is amortized across billions of sessions, and a compute-short leader can charge a higher markup. Dylan adds that wider deployment supplies more learning data and that the best model can help produce the next-best model through RSI.
• Referencing the Gavin Baker–Dario dispute, Dylan says that if one believes in RSI and the labs’ superior ability to monetize compute, centralization follows. He says the limiter on AGI is not research engineers’ ability to “crank the gears,” but “how much the rest of the world lets that happen.”
• Dylan, not Dwarkesh, raises the fear of a 2030 six-month release moratorium: during those six months, a lab might conduct recursive self-improvement internally while the public remains years behind. He also says slow takeoff is possible — “at least, that’s my hope” — because of regulation, financing and deployment constraints.
• Dylan proposes an intellectual project to find a decentralized, broadly empowered post-AGI future that takes economies of scale seriously. He says he does not trust “the government,” Dario or Sam. His claim that users capture most value is explicitly his “cope”: Jane Street might capture $300M–$500M/MW while Anthropic captures $100M/MW, but if internal AI research is worth more, Anthropic has an incentive to keep the compute inside. The unresolved close: “There’s 80,000 worlds and in only one of them, Anthropic doesn’t own the whole world.”