Anthropic has committed to spend at least $518 billion on cloud computing, chips and data-center infrastructure over the next decade. The number, drawn from the company’s confidential IPO prospectus, dwarfs anything seen before in artificial intelligence. Roughly 80% of those obligations cannot be canceled or must be paid regardless of actual usage.

Google stands to collect at least $111.1 billion. Amazon gets $110 billion. Microsoft claims $31.4 billion. Add in $161.2 billion in largely non-cancelable equipment leases tied to Broadcom. The sums lock the startup into take-or-pay contracts that turn what once looked like variable costs into near-fixed obligations. Anthropic insists the deals secure scarce resources in a market where compute, not demand, will limit progress.

Yet the scale raises hard questions. Last year Anthropic generated $4.6 billion in revenue, a twelvefold jump from the year before. It still posted more than $8 billion in operating losses. Compute and infrastructure alone ate up $7.33 billion. By July 2026 its annualized revenue run rate had climbed above $65 billion. Growth looks explosive. The infrastructure bill does not.

The figure first surfaced in late September when Reuters reviewed the draft filing. It matches earlier tallies reported by Forkast and others that pegged commitments near $517 billion after an 11-month scramble for 14.8 gigawatts of additional capacity. Conservative estimates put verifiable deals above $441 billion. The rest flows from options, extensions and future generations of hardware not yet built.

Amazon’s piece traces back to an April 2026 agreement. The e-commerce giant invested another $5 billion in Anthropic, lifting its total stake to $13 billion with potential for $20 billion more tied to milestones. In return Anthropic pledged more than $100 billion over 10 years on AWS, targeting up to 5 gigawatts of capacity built around Trainium2, Trainium3 and yet-to-arrive Trainium4 chips. TechCrunch detailed the swap: cash and silicon for guaranteed spend. Andy Jassy, Amazon’s CEO, highlighted the custom chips’ price-performance edge at the time.

Google’s commitment runs from April 2026 through July 2033. The search company and parent Alphabet already invested heavily in Anthropic and supply tensor processing units through Broadcom. Microsoft locked in $31.4 billion through May 2033 on Azure, non-cancelable except in cases of material breach. These three hyperscalers also rank among Anthropic’s biggest investors and, in some cases, direct AI competitors. The circular flow of capital, compute and customers creates tight dependencies few outsiders fully map.

Broadcom’s $161.2 billion slice comes mostly through equipment leases tied to those Google TPUs. AMD agreed to buy up to $5 billion of Anthropic equity while supplying more than $20 billion in computing capacity. Arrangements with Elon Musk’s xAI offer more breathing room: up to $84.5 billion in Nvidia-based capacity through 2029, yet most of it cancelable on 90 days’ notice. Fluidstack landed a $50 billion deal to build data centers in Texas and New York. Nscale secured roughly $45 billion for its Monarch campus in West Virginia. SpaceX, Lambda, Akamai and others fill out the roster.

The spending surge coincides with explosive model progress. Claude’s capabilities have driven customer adoption at scale. Revenue routed through Amazon and Google hit 47% of total sales last year, up sharply from prior periods. But the prospectus warns that future AI systems will be constrained principally by compute availability. Dario Amodei, Anthropic’s CEO, has publicly called for the industry to pace development. The contracts suggest the company itself has chosen acceleration.

Power remains the unspoken bottleneck. Securing 14.8 gigawatts over 11 months required deals with utilities, specialized providers and even repurposed facilities. Some contracts bundle compute with energy infrastructure. Others rely on options that may never convert if power shortages persist. Data-center construction timelines stretch years. Chip supply chains face their own constraints. Non-cancelable payments offer suppliers visibility. They also saddle Anthropic with costs even if adoption slows or efficiency gains outpace expectations.

Investors appear unfazed so far. Speculation around a potential IPO has valued the company north of $2 trillion in some discussions. Amazon carries its Anthropic holdings at roughly $190 billion on its books. The circular economics benefit all parties while the growth story holds. Break the revenue trajectory, however, and those fixed obligations could weigh heavily.

Recent reporting adds color. On October 7 WebProNews framed the commitments as a locked-in gamble on AI’s power-hungry future, noting the shift from variable to fixed costs. The Motley Fool updated its coverage this week, emphasizing that more than $100 billion flows straight to Amazon and that the annual run rate already implies tens of billions in yearly spend against last year’s $7.33 billion actual outlay.

Meanwhile conversations on X highlight fresh angles. Lambda, a key partner, seeks up to $4 billion in funding at a $14.5 billion pre-money valuation with a $50 billion backlog that includes a reported $35 billion commitment from Anthropic. Industry pledges of compute for public science efforts and questions around power grid upgrades surface regularly. The numbers keep climbing.

Anthropic’s bet reflects a broader industry pattern. OpenAI’s Stargate project carries a similar half-trillion-dollar price tag shared across partners. Every frontier lab faces the same math: bigger models require orders of magnitude more compute. Efficiency improvements help at the margin. They have not yet changed the fundamental equation. So the contracts multiply. The capital flows. And the power draw grows.

Whether the economics close remains the open variable. Revenue must scale dramatically to service these obligations without eroding margins or forcing dilution. Customer willingness to pay for ever-larger models will decide the outcome. So will regulatory scrutiny, energy availability and the pace of hardware innovation. For now Anthropic has placed its chips. The table is set for the rest of the decade.