OpenAI and Anthropic will add “about 12 gigawatts” of compute in 2027 while the world adds “about 40”, SemiAnalysis chief executive Dylan Patel estimated in a talk at Modal’s Runtime 2026 conference. That puts the two labs at 30 per cent of the world’s new compute next year, by our arithmetic. Patel himself gave a range: “25, 30%, 40%”.
Modal’s site dates Runtime 2026 to 1 October in San Francisco; the video of the talk was posted on 2 November. Every figure below is Patel’s estimate, not a company disclosure.
Dylan Patel, SemiAnalysis, at Modal’s Runtime 2026. The gigawatt figures are at about 2:58 to 4:20; the signed-compute remark is at about 14:45.
Business Pill · Concentration
An evergreen Business Engineer explainer on concentration: when a few big pieces make up the total. It teaches the general idea only and does not discuss OpenAI, Anthropic or any company in this story.
The key insight: SemiAnalysis chief executive Dylan Patel estimates that OpenAI and Anthropic add about 12 gigawatts of compute in 2027 while the world adds about 40. By our arithmetic that is 30 per cent of the year’s new compute, from two companies that hold about 10 per cent or less of the installed base at end-2026 on his figures.
Patel opened on scarcity [2:41]. Anyone who has tried to buy or rent chips over the last year knows “it’s extremely tight”, he said, and the reason is “OpenAI and Anthropic primarily. They’re just buying everything out there.”
He then sized the world’s stock of compute, “including AI and non-AI” [2:58]. “In 2024 the entire world had 60 gigawatts of compute globally by the end of 2024”, he said, and “by the end of this year we’re at over 100 already.” He added: “It took us you know, 50 years to get to 50 and then you know, just a couple years to get to 100.”
For the two labs, Patel said that “by the end of last year they had about 4 gigawatts total between Anthropic and OpenAI. And this year they’ll have about 10 and then the year after that they’ll have over 20.” For 2028 he said that, “based on what we’ve seen signed and such, the AI labs are going to add on the order of uh 30 gigawatts in 2028 alone.”
He tied the buying to revenue. “Anthropic’s revenue started to go ballistic with Claude Code, and so they started buying compute like crazy”, he said [3:59], adding that Meta and OpenAI did the same: “Now it’s like impossible to buy compute uh for a lot of folks.”
Year-end compute in gigawatts as Dylan Patel of SemiAnalysis estimated it at Modal’s Runtime 2026: the world (AI and non-AI) at 60 GW at end-2024 and over 100 GW at end-2026; OpenAI and Anthropic combined at about 4, about 10 and over 20 GW at end-2025, 2026 and 2027. Right: about 12 GW added by the two labs against about 40 GW added worldwide in 2027. The 30% and the 28 GW for everyone else are our arithmetic; he gave a range of 25 to 40 per cent and no world total for end-2025 or end-2027.
The Arithmetic Behind the 30%
Patel’s figures put the two labs’ installed compute at about 4 gigawatts at the end of 2025 and over 20 at the end of 2027: more than five times in two years, by our arithmetic. Their yearly additions on his numbers are about 6 gigawatts in 2026, about 12 in 2027 and about 30 in 2028, each step at least doubling, by our arithmetic.
The world side moves more slowly. His 60 gigawatts at end-2024 and “over 100” at end-2026 give an average of more than 20 gigawatts added a year, by our arithmetic. His “about 40” for 2027 is roughly double that pace.
The share of the stock is smaller than the share of the flow. On his figures the two labs hold about 10 of the world’s more than 100 gigawatts at end-2026, about 10 per cent or less, by our arithmetic. In 2027 they take about 12 of the 40 gigawatts added, leaving about 28 for everyone else.
Who Gets the Rest
The remaining 28 gigawatts are not open to all buyers either, on Patel’s account. “Next year if you look at all the compute that’s been signed, it’s something on the order of 65, 70% of it is going to hyperscalers and the AI labs”, he said [14:44]. That leaves 30 to 35 per cent of signed compute for everyone else, by our arithmetic.
He described the mechanism through debt. A neocloud that has 100 megawatts and wants a gigawatt can borrow only if it sells that gigawatt to a hyperscaler, he said: “But then none of that compute gets onto the market for the general industry.”
How the Estimate Compares With August
Patel gave an earlier set of figures on the Dwarkesh Podcast on 25 August, as relayed by China’s 36Kr on 28 August. In that account, “the world’s newly built computing power this year is about 30 gigawatts, 50 gigawatts next year, and around 70 gigawatts in 2028”, and the two labs’ share of new compute was put at about a third this year and close to half next year.
The October talk’s 2027 world figure of about 40 gigawatts is about 10 lower than the 50 in that relay, by our arithmetic, and his share range of 25 to 40 per cent sits below “half”. We have not heard the podcast itself, so the August figures are 36Kr’s account.
What the Filings Show
Anthropic’s amended S-1, filed with the SEC on 2 November, states its commitments in dollars, not gigawatts: an aggregate committed spend of approximately $895.4 billion, “substantially all of which is related to compute and infrastructure”, with approximately $69.8 billion due in the 12 months after 30 September. A search of the filing’s text for gigawatt and megawatt figures returned no match.
Earlier press reports of Anthropic’s investor discussions, which we covered in September and could not verify, put its own compute at about 1.4 gigawatts at the close of 2025 and about 5 gigawatts by the end of 2026. Those are reported figures, not company disclosures.
As we read Patel’s figures, the gap between the share of the stock and the share of the flow is the point: the two labs are small in installed compute and large in what is being added.
As we read his remarks on debt, the remaining capacity is narrower than the 28 gigawatts suggest, because he said 65 to 70 per cent of compute signed for next year goes to hyperscalers and AI labs.
Dylan Patel, CEO, SemiAnalysis, at Modal Runtime 2026
“It took us you know, 50 years to get to 50 and then you know, just a couple years to get to 100”
Three Implications
FOR OTHER COMPUTE BUYERS Patel said 65 to 70 per cent of compute signed for next year goes to hyperscalers and AI labs, which leaves 30 to 35 per cent, by our arithmetic.
FOR NEOCLOUDS Patel said a neocloud scaling from 100 megawatts to a gigawatt can get a loan only by selling that gigawatt to a hyperscaler, so none of it reaches the general market.
FOR READING THE FILINGS Anthropic’s amended S-1 states about $895.4 billion of commitments in dollars, and a search of its text found no gigawatt figure to set against Patel’s estimates.
The framework argues that “The $1T capex cycle is being underwritten by a $1T concentrated demand commitment from two counterparties whose combined free cash flow is deeply negative”, naming OpenAI and Anthropic.
As we read it, Patel’s gigawatt estimates describe the same two counterparties in physical units. The framework was published in May 2026 and does not discuss his talk.
The figures are Patel’s spoken estimates, rounded, and he gave a range of 25 to 40 per cent rather than one share. He did not give a world total for end-2025 or end-2027, did not split the labs’ gigawatts between OpenAI and Anthropic, and did not say whether his gigawatts count IT power or whole-site power. Neither company has published a gigawatt total that matches his. The 30 per cent and every other ratio here is our arithmetic.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
On Patel’s estimates, OpenAI and Anthropic add about 12 of the roughly 40 gigawatts of compute the world adds in 2027, 30 per cent by our arithmetic, while holding about a tenth of the installed base at end-2026. He also said 65 to 70 per cent of compute signed for next year goes to hyperscalers and AI labs.
A note on sourcing. Gigawatt figures and quotes are Dylan Patel’s, from the auto-generated captions of his talk at Modal’s Runtime 2026, checked against the video; they are his estimates, not company disclosures. The August figures are 36Kr’s account of a podcast we have not heard. Anthropic figures are from its amended S-1. Shares, ratios and differences are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
