Two labs, most of the FLOPs: Dylan Patel's compute bet

Dylan Patel's latest conversation with Dwarkesh Patel is a market-structure argument about who can pay for the next watt. Patel, founder of SemiAnalysis, argues that OpenAI and Anthropic already take a large and rising share of new AI compute, that their revenue per megawatt now lets them outbid almost everyone else, and that by the end of 2028 those two labs could control most of the world's usable FLOPs. 1

The episode, published August 25, 2026, runs about 77 minutes. Dwarkesh frames the world economy as increasingly "a function of where lab economics are headed." Patel supplies the contracts, rental prices, fab bottlenecks, and financing underneath that claim. 1

At the beginning of 2026, Patel says, OpenAI had roughly 2 gigawatts of compute and Anthropic had less than 2. By year-end he expects both above 5 gigawatts — a three- to four-fold jump. About a third of the compute coming online this year is already for those labs as end customers, even when another company builds and rents the capacity. 2

Next year looks more extreme. Contracts already signed put Anthropic and OpenAI at 40% to 50% of incremental compute. By the end of 2027, half of new capacity is going to the two labs. If the trend continues into 2028, Patel has them taking 70% to 80% of incremental compute. 2

World capacity is growing fast, but frontier lab demand is growing faster. Patel sketches world additions of about 30 GW this year, 50 next year, and roughly 70 in 2028, with global stock over 200 GW by the end of 2028. New chips also do more work per watt: GB300s, TPUv7s, and Trainium3-class hardware are 3x to 5x more performance per watt than the prior generation. A lab that captures half of next year's new watts is capturing more than half of the new FLOPs. 2

Dwarkesh multiplies the frontier path — roughly 2 GW at the start of 2026, near 6 by year-end, 18 by end-2027, 54 by end-2028 if labs keep tripling. Combined, that would put the two labs near 100 GW by the end of 2028. Patel calls 100 GW "a very aggressive goal" and is unsure whether markets will sell them that much, but he sees "nothing that's stopping" the direction of travel. One bookkeeping note matters: when Amazon serves Anthropic models through Bedrock, Patel still counts that capacity as Anthropic compute because the end revenue is Anthropic's. 2

The mechanism is pricing power. Patel puts the base cost of compute around $10 million to $15 million per megawatt. Serving GPT-4 on Hopper GPUs once produced negative gross margin for OpenAI. Serving current frontier models is different: Anthropic's revenue has gone as high as $50 million per megawatt. Spend $10 on inference capacity, generate $50 of revenue, and plow the difference into training. 2

That gap turns SpaceX, Meta, and other builders into suppliers. Patel says SpaceX is already selling compute in the $25 million to $40 million per megawatt range, and Dwarkesh cites a SpaceX tranche sold to Google at about $40 billion per gigawatt. Meta and SpaceX are, in Patel's view, the only plausible third poles because they can build capacity on their own balance sheets without a pre-signed end customer, then choose internal use or a high-priced lease to OpenAI or Anthropic. 2

Most of the market still clears lower. Patel expects most compute to remain under $20 billion per gigawatt through the end of next year because projects have to be financed before they are built. To reach 70% of world incremental compute by 2028, though, the labs would need to pay $25 million, $30 million, or $50 million per megawatt. Anyone can still make money at $10–15 million per megawatt by renting a GB300 rack, loading open weights, and selling tokens on OpenRouter. That floor is already lifting. 2

By the end of 2027, Patel puts blended lab revenue at $50 million-plus per megawatt, and possibly $70–80 million, if the labs can keep releasing their best models. Dwarkesh thinks those figures are low if capability keeps compounding. Patel's reply is that regulation is already holding revenue per megawatt down. 2

Dwarkesh presses the supply-side puzzle. Using Patel's earlier wafer model, about $6 billion of fab CapEx can support a gigawatt of annual chip production once cleanrooms and shells are included. If that gigawatt supports roughly $100 billion a year of end AI revenue, five years of output can mean more than $1 trillion of end revenue against a few billion of fab tooling. Even after splitting the stack with data centers, power, and model R&D, Dwarkesh still sees roughly a 100x gap between fab CapEx and end revenue. 2

The signal should expand EUV mirrors, tools, and turbines. Patel agrees that Carl Zeiss is now planning enough mirrors for about 100 ASML EUV tools a year by 2030, and that a $10 billion blank check to Zeiss would change the path. He does not expect that kind of forced expansion this year, next year, or the year after. Labs may generate hundreds of billions in revenue next year while total AI CapEx runs near $2 trillion, with wafer equipment alone on the order of $200 billion. The whip takes time to reach the end of the supply chain. 2

Patel also changes the usual picture of what labs do with their fleets. Allocation is moving toward internal R&D. A rough split he gives is 50% research, 10% development, 40% inference, with the inference share falling over time. Even with multiple gigawatts on the floor, a single model run may only be able to use on the order of 200 megawatts because of multi-site coordination and RL limits. The actual Mythos pre-train he describes is sub-200 MW for about two months. 2

In the last three months, he says, revenue additions have plateaued while compute kept growing, so the marginal megawatt has been going more to research than to serving tokens. More inference capacity still funds the training fleet, but the highest return on the next watt is often another research experiment rather than another customer workload. 2

On geography, Patel's numbers are stark. In 2022 the US added roughly 45–50% of world compute and China 30–35%. Today he puts about 70% of new watts in America and China under 10% of incremental data-center AI compute. By 2028 he expects China at 30 GW or less of AI compute, with domestic fabs adding only 5–10 GW that year on weaker chips. Dwarkesh's summary is that a leading US lab in 2028 could have more effective compute than all of China in 2029 or 2030. Patel agrees. 2

The brake on the US side is release policy and local politics. Patel lists OpenAI holding back Astra, OpenAI pausing training for two weeks, and Anthropic holding back what safety assessments treat as Model 2 — widely believed to be the next Mythos — plus limits on who can use current systems. If the best internal models cannot ship or run widely inside the company, revenue per megawatt stalls and the labs lose the ability to outbid everyone else. Local data-center fights in New York, Texas, Ohio, and elsewhere raise costs and slow supply on a different track. 2

The second half of the episode turns from racks to balance sheets. Patel cites on the order of $11 trillion of AI CapEx from 2024 through 2029, with roughly $6 trillion cash-funded and $5 trillion debt-funded. Higher returns on AI projects raise the interest rate that other borrowers face. He thinks Meta or Amazon credit could move from about 5–6% toward 8% — still viable for hyperscalers, harder for long-duration equities and weaker sovereigns. 2

Dwarkesh extends the tax side. Most federal revenue is payroll and income tax, not corporate tax. If automation shrinks the wage base while debt service already takes a large share of tax receipts, higher rates make the fiscal math worse. Patel's historical analogy is a second Volcker-style shock: the 1980s rate spike contributed to roughly 40 country defaults, mostly in Latin America. He expects something like that pattern again for countries outside the AI production stack. 2

If the world were adding 100 GW a year at current prices, IT CapEx alone could approach $5 trillion a year, and power plants plus data centers built ahead of demand could push the total toward $7–10 trillion a year by 2030 — on the order of a tenth of world output. Patel and Dwarkesh leave open the possibility that politics simply refuses a buildout of that size. 2

Dwarkesh's closing frame is labor, not chips. Effective AI labor at the frontier is compounding roughly 10x a year if FLOPs grow several-fold and the compute needed for a given capability falls several-fold. On that path, a lab can move from tens of millions of AI-labor equivalents to hundreds of millions to a billion. Most work output would then sit inside two companies that also buy most of the world's new compute. 2

Patel still sees a temporary saving grace: end users capture more value than the labs. Jane Street or Meta may extract hundreds of millions of dollars of value per megawatt while paying Anthropic far less. Dwarkesh answers with the same logic Patel used earlier. If internal R&D returns more per watt than external inference sales, the rational lab pulls tokens inside. "There's 80,000 worlds," Dwarkesh says, "and in only one of them, Anthropic doesn't own the whole world." 2

Neither speaker claims that path is locked. Both treat slowdown through regulation, supply-chain friction, or weaker model progress as real brakes. What the episode makes hard to dismiss is the intermediate mechanism: when two buyers can monetize a megawatt several times better than everyone else, markets will keep offering them a larger share of the next rack until something breaks that spread.

Listen on Apple Podcasts, read the transcript on Dwarkesh's site, or watch the full episode: