Analysis
Anthropic has quietly signed the most expensive shopping list in tech history, and none of it shows up as revenue yet. Since the spring, the AI lab has committed to five separate multibillion-dollar compute contracts -- with Amazon Web Services, Fluidstack, Nscale, SpaceX, and Lambda -- that together add up to more than $275 billion in reserved cloud capacity, before the company has even priced its IPO.
The scramble traces back to a single constraint: Anthropic's Claude models are compute-bound, and every frontier lab is racing to lock in GPU and custom-silicon capacity years before it is needed. Amazon set the tone in April, expanding its Anthropic partnership to as much as 5 gigawatts of Trainium capacity as part of a commitment exceeding $100 billion over a decade, layered on top of the equity Amazon has separately invested directly into Anthropic. Everything since has been Anthropic diversifying away from depending on any single supplier.
Five Deals, One Company
Laid out together, the five contracts show how far that diversification has gone:
- Amazon Web Services -- $100B over 10 years (announced April 2026): up to 5 gigawatts of Trainium2-through-Trainium4 capacity, on top of Amazon's existing equity stake in Anthropic. Constellation Research
- Fluidstack -- $50B: a neocloud compute deal supplying GPU capacity outside the traditional hyperscalers, part of Anthropic's push to avoid over-concentration in any one vendor.
- Nscale -- $45B: capacity out of a West Virginia buildout -- the same neocloud that separately raised its own $3.5 billion pre-IPO round from Nvidia, a deal Pulse covered as this week's lead.
- SpaceX -- $45B: compute capacity from Elon Musk's SpaceX, notable given SpaceX's own record-setting IPO earlier this year and its parallel build-out of AI infrastructure.
- Lambda -- $35B (announced late August): a Nvidia-backed neocloud deal built around a roughly 350-megawatt Hut 8 data center in Nueces County, Texas, with Nvidia itself holding the lease on the facility. Bloomberg
OpenAI is running a smaller version of the same playbook: it has committed roughly 2 gigawatts of Trainium capacity through AWS starting in 2027, converting what had been an almost exclusively Azure-aligned compute relationship into a contested one. Nvidia, meanwhile, sits on nearly every side of this market at once -- it backs Lambda as chip supplier and landlord, it financed Nscale's pre-IPO round, and on September 3 it confirmed a $12.93 billion acquisition of Hugging Face, the model-distribution platform used by more than 200,000 companies, giving Nvidia a direct stake in where AI workloads get discovered and deployed, not just where they run. CNBC
The Bill Comes Due
Scale matters more here than any single figure. Anthropic's last disclosed private valuation was $965 billion as of its Series H in May, meaning its committed compute spending now runs close to 30% of its own valuation -- a ratio with no real precedent among prior tech IPO candidates, since even the most capital-intensive telecom and cloud buildouts of the 2000s rarely pre-committed capital at that multiple of enterprise value before going public.
None of this is cash Anthropic has already spent. These are multiyear, take-or-pay-style capacity reservations, not drawn balance-sheet liabilities, and the actual utilization -- how much of that reserved compute Anthropic's revenue can justify filling -- is undisclosed. A gap between reserved capacity and paying demand is exactly the risk that sank prior infrastructure buildouts in the dot-com era, and Anthropic's revenue, while growing fast, remains a small fraction of $275 billion even on aggressive multiyear projections.
For founders building on top of Claude, Gemini, or GPT-6, the practical read is that inference is likely to get cheaper and more available over the next 18 months as this capacity comes online, since Anthropic, OpenAI, and Google are all simultaneously over-provisioning against the same demand curve. For LPs evaluating AI infrastructure funds, the diligence question is which neoclouds -- Lambda, Nscale, Fluidstack, Crusoe -- have contracts with real termination penalties versus soft commitments that evaporate if a single anchor customer's growth slows.
Anthropic's IPO prospectus, expected in late September ahead of a marketing push that has already slipped once toward mid-October, will be the first place investors get an actual number for how much of this $275 billion is contracted revenue commitment versus optionality Anthropic can still walk away from.
