Monday, September 7, 2026 | 4 min read
Key Takeaway: Anthropic has quietly committed up to $517 billion to compute over eleven months — roughly triple the $180 billion it once told investors it would spend — putting the industry’s loudest voice for spending discipline squarely inside the buildout race it warned about.
---
Eleven months ago, Dario Amodei was the AI industry’s designated skeptic on capital expenditure — the executive who told a podcast audience that his rivals were “YOLOing” hundreds of billions into data centers they might never fill. On Sunday, The Information reported that Amodei’s own company has signed compute contracts worth as much as $517 billion since October 2025, locking in at least 14.8 gigawatts of capacity on top of the one to two gigawatts Anthropic already had. The number is not a forecast or a letter of intent. It is the running total of what Anthropic has agreed to pay.
That figure represents an extraordinary escalation. Anthropic had previously told investors it expected to spend roughly $180 billion renting servers through 2029. The Information’s tally — assembled from public announcements and its own reporting — is nearly three times that, and the publication notes the company could have signed additional deals it never disclosed. Even so, Anthropic still trails OpenAI, whose commitments run to roughly $750 billion and whose stated target is 30 gigawatts by 2030. Many of Anthropic’s contracts extend well past that date, which makes a clean head-to-head comparison harder than the headline numbers suggest.
The counterparty list reads like a map of the entire compute economy. Google and Amazon Web Services together account for about 11 of the 14.8 gigawatts, with a $200 billion TPU arrangement on the Google side and a five-gigawatt AWS lease tied to Amazon’s $5 billion investment and its custom Trainium clusters. Beyond the two hyperscalers, Anthropic has signed a $50 billion partnership with Fluidstack for U.S. sites in Texas and New York, a reported $45 billion agreement with Nscale, $35 billion with Lambda, $18 billion with Akamai, a 191-megawatt, 20-year lease with Riot Platforms in Texas, and deals with CoreWeave, AMD, Microsoft, Volta Infra, and SpaceX. A SpaceX IPO filing disclosed that Anthropic is set to pay Elon Musk’s firm roughly $125 million a month to rent xAI data center space. The company has also filed more than a dozen letters of intent for data center leases with U.S. developers and is planning facilities of its own.
Why It Matters
The gap between what Amodei said and what Anthropic signed is the most interesting thing in this story — and it is not necessarily hypocrisy. In a February interview with Dwarkesh Patel, Amodei laid out an unusually precise version of the bear case, and the arithmetic he used still holds.
“I really do believe that we could have models that are a country of geniuses in the data center in one to two years,” he said. “One question is: How many years after that do the trillions in revenue start rolling in? I don’t think it’s guaranteed that it’s going to be immediate.”
His concern was never that compute was a bad investment. It was that the timing mismatch between a signed lease and the revenue meant to service it is unforgiving. “If my revenue is not $1 trillion, if it’s even $800 billion, there’s no force on Earth, there’s no hedge on Earth that could stop me from going bankrupt if I buy that much compute,” Amodei said. “If I’m just off by a year in that rate of growth, or if the growth rate is 5x a year instead of 10x a year, then you go bankrupt.”
Crucially, Amodei also said in that same conversation, “We’re buying an amount that’s comparable to what the biggest players in the game are buying.” Read against the $517 billion figure, his position looks less like a reversal than a bet with a specific shape: buy aggressively, but stay one tier below the leader, and lean on enterprise customers whose spending is stickier than consumer subscriptions. Anthropic was tracking around $10 billion in revenue for 2026 by Amodei’s own account in February, on a roughly 10x annual growth curve; The Decoder, citing Bloomberg, reports annualized revenue has since topped $65 billion. Neither Anthropic nor OpenAI can cover its commitments out of current revenue.
The wider context is a market where the caution is now bipartisan. Sam Altman, who was the implicit target of Amodei’s “YOLO” remark, told an interviewer last week that parts of the buildout amount to “unsustainable silliness,” singling out neocloud providers announcing capacity without customers to fill it. He conceded that OpenAI’s own efficiency gains could turn today’s expensive contracts into bad bets. Both CEOs are now warning about the same dynamic while participating in it — which is what a genuine arms race looks like from the inside.
Watch the IPO. Anthropic confidentially filed with the SEC in June, and no substantive updates have followed. A public listing would force the company to disclose the actual structure of these commitments — how much is firm take-or-pay obligation versus optionality, and how the $517 billion is phased across years. That disclosure would move this story from a compiled estimate to an audited liability, and it would give the market its first honest look at whether the buildout race is a balance sheet or a bubble. Anthropic did not respond to a request for comment from DatacenterDynamics.
---
Sources: - The Decoder - DatacenterDynamics - Fortune via Yahoo Finance - The Decoder — Altman on the buildout
“If my revenue is not $1 trillion, if it's even $800 billion, there's no force on Earth, there's no hedge on Earth that could stop me from going bankrupt if I buy that much compute.”— Dario Amodei, Cofounder and CEO, Anthropic