Anthropic has signed compute deals totaling about $517 billion over roughly 11 months, according to a report from The Information, reserving 14.8 gigawatts of data center capacity that may cost the company as much as $517 billion to use over ten years.
The umbrella covers several agreements reported through the year. Amazon Web Services committed $100 billion over ten years for up to 5 gigawatts of Trainium capacity. A Google and Broadcom deal from April added multiple gigawatts of next-generation TPU capacity, reported at around $200 billion over five years for 5 gigawatts. Neocloud deals with Fluidstack, Nscale, SpaceX and Lambda fill in the rest.
Five deals, one company
A tally compiled by Value Add Pulse in early September put Anthropic’s five cloud contracts at more than $275 billion: AWS at $100 billion, Fluidstack at $50 billion, Nscale at $45 billion, SpaceX at $45 billion and Lambda at $35 billion. The Information’s $517 billion figure counts the same deals plus the Google and Broadcom arrangement over a longer window.
Against a last disclosed private valuation of $965 billion from its Series H in May, committed compute spending runs close to 30% of the company’s own value. The Information noted that ratio has no real precedent among prior tech IPO candidates, pointing out that even the most capital-intensive telecom and cloud buildouts of the 2000s financed far less before going public.
Nvidia is on every side of the table
The wiring gets strange fast. Nvidia backs Lambda as chip supplier and landlord. It financed Nscale’s pre-IPO round. It agreed this week to buy Hugging Face for $12.93 billion, adding the open-model distribution platform to a portfolio that already touches most of the compute chain. A single vendor now sits inside the financing, supply and infrastructure layers of its biggest customer’s expansion.
Anthropic runs Claude on three chip platforms: AWS Trainium, Google TPUs and Nvidia GPUs. Company statements describe matching workloads to whichever chip fits best. That flexibility gives Anthropic negotiating leverage with each vendor, but it also means the company is spreading an enormous bill across three supply chains at once.
Demand is real, and so is the risk
None of this is speculative demand. Anthropic’s run-rate revenue passed $30 billion in April, up from about $9 billion at the end of 2025. More than 500 business customers each spend over a threshold the company disclosed during its February Series G fundraising, which valued the firm at $380 billion at the time. The compute is being bought to serve contracts that already exist, not to chase a market that has yet to appear.
The risk sits in the gap between revenue and commitments. $30 billion of run-rate revenue against $517 billion of potential spend means Anthropic is betting on years of steep growth to make the math work. If AI demand holds, the capacity is there and competitors are locked out of the same data centers. If it softens, the company carries obligations that dwarf its earnings.
There is also a concentration question. Data center power has become the binding constraint across the industry, and Anthropic’s 14.8 gigawatts is a large share of the capacity that grid operators can realistically add in the next three years. Amazon’s Northern Virginia region already had power issues in May, and drone activity disrupted AWS’s Bahrain region in March during the Middle East conflict. Physical infrastructure, not chip supply, is where the deals could slip.
| Partner | Value | Capacity |
|---|---|---|
| AWS | $100B over 10 years | Up to 5 GW Trainium |
| Google/Broadcom | About $200B over 5 years | 5 GW TPUs from 2027 |
| Fluidstack | $50B | GPU capacity |
| Nscale | $45B | Data center capacity |
| SpaceX | $45B | Data center capacity |
| Lambda | $35B | GPU capacity |
Most of the new compute is sited in the United States, deepening the company’s November 2025 pledge to invest $50 billion in American AI infrastructure. Amazon remains the primary cloud and training partner through Project Rainier. The Google deal starts delivering TPU capacity in 2027, which means the heaviest bills land after the company has priced an IPO that now looks inevitable.
discussion