Anthropic has signed compute contracts worth up to $517 billion within eleven months, as reported by The Information. The commitments amount to an additional 14.8 gigawatts of capacity – on top of the one to two gigawatts the company already had. Anthropic has not officially confirmed the total sum so far.
Google and Amazon provide the largest individual items
The largest individual item is a five-year contract with Google and its chip partner Broadcom for $200 billion for about five gigawatts of TPU capacity. Amazon follows with a ten-year contract for $100 billion for an additional five gigawatts based on the new Trainium3 chips, in addition to the existing data center Project Rainier. Microsoft contributes $30 billion for Azure capacity. The rest is distributed among smaller infrastructure partners: $50 billion for a dedicated data center build with the startup Fluidstack, $45 billion for capacity with the British provider Nscale, about $45 billion for a monthly billed agreement with SpaceX, and $35 billion for compute power from provider Lambda. The commitments are so-called take-or-pay reservations that will be drawn down over the years – not immediate cash payments. The Information’s tally effectively consolidates nearly all previously separately reported capacity contracts of the company from the past eleven months into a single total. Anthropic justifies the capacity need with growing demand from corporate clients who increasingly use Claude for automated workflows, as well as with training future model generations.
Investors ask about revenue per gigawatt ahead of the IPO
The sum lands as Anthropic prepares for its IPO, now postponed to mid-October. Potential investors are demanding more granular metrics such as revenue per token and per gigawatt of compute, according to Crypto Briefing, to gauge the efficiency of the billion-dollar infrastructure bets. The annualized revenue rate jumped from $47 billion to $65 billion between May and the end of July, with some investors pricing in a valuation of up to $3 trillion. Anthropic had also projected annual revenue of up to $200 billion by 2028 for investors – more than three times the current rate. Despite the record sum, Anthropic still trails the competition: according to the Wall Street Journal, OpenAI plans to spend $750 billion on compute by 2030 – 25 percent more than estimated earlier this year. Observers read the gap as a sign of how much the race for compute is now shaping the balance sheets of every major AI provider.
Amodei had warned rivals about risky pace
The tally highlights a possible contradiction: Anthropic CEO Dario Amodei had warned competitors at the New York Times DealBook Summit in December 2025 against those who “pull the risk dial too far” when expanding their compute capacity. Anthropic itself consistently emphasizes a disciplined approach to capacity building. The company has not officially confirmed the total aggregated by The Information; the $517 billion figure is therefore not independently verified. Several of the individual contracts, including those with Lambda and Nscale, had previously not been officially confirmed by any of the companies involved. Also not included are further, separately negotiated projects such as the joint venture Theseus Infrastructure with Macquarie and GIC or the ten-billion-dollar deal with Volta running through a Norwegian data center. The trend extends beyond Anthropic: Google, Microsoft, and Meta are also announcing new billion-dollar contracts for data centers and chip capacity within weeks of each other, while Nvidia, as a supplier, is additionally helping finance some of these projects. Analysts read the buildup differently – some see forward-looking planning, others a growing financial risk for the still privately held company.
What will matter is whether the added capacity translates into measurable revenue before the first billion-dollar contracts come due. For the planned IPO, that question is likely to decide the valuation – especially since investors are for the first time demanding detailed efficiency metrics instead of pure growth figures. No specific date for the listing has been set yet.
