Anthropic's S-1 Is a Take-or-Pay Contract the Size of a Country

Yesterday I wrote about what happens when the money for AI data centers gets tight (Stranded Compute). Today Anthropic filed its S-1 publicly, and it settles most of the open questions in one document. The filing is on EDGAR, 335 pages, dated October 26, 2026.¹ Strip out the IPO mechanics and what is left is a capacity-planning document: a company describing, in the mandatory language of a securities filing, how it buys compute, what that buying costs, and what it owes even when the machines sit idle.

The floors

Start with the minimums, because they are the part that survives any outcome. Anthropic has committed to spend at least $44.9 billion with Microsoft between November 2026 and May 2033 “regardless of actual usage.” At least about $114.5 billion with Google between April 2026 and July 2033, with a true-up payment if actual spend falls short. At least $110.0 billion with Amazon between May 2026 and April 2036, same true-up. Those three floors alone are about $269 billion owed no matter how many tokens anyone buys.² None of the three agreements can be cancelled except for the provider’s own uncured material breach.

Then the layer beneath the floors. Broadcom master equipment leases, covering TPU-based compute, add roughly $161.2 billion of rent over the life of the leases, excluding interest, non-cancelable during the lease period. The SpaceXAI arrangements run up to $120.5 billion into 2029, though after some initial locked periods either side can walk on 90 days’ notice. Nscale, up to $44.6 billion, fees accruing only as capacity is delivered and accepted, and Nscale is required to use best efforts to obtain the financing for the GPUs. Lambda, about $41.3 billion into 2034. AMD, no fixed price but expected to exceed $20 billion if the first tranche fills. Google’s additional TPU compute for 2028 to 2031 is estimated at $125 billion to $150 billion and is not even counted in the contractual obligations table. There is a hazard in adding these up, because they are not all the same kind of promise, but the named pieces alone clear $600 billion, and the filing itself leaves the grand total blank.

What the numbers do to the capacity plan

A capacity plan is a stack of numbers that behaves differently, and the filing shows the whole stack at once. Committed compute is bought 12 to 18 months before it is available, so the expense line is “generally fixed in the near term” and rises as contracted tranches land rather than as demand arrives. In the first half of 2026 compute and infrastructure expense was $8.9 billion, up 283 percent, while revenue was $16.3 billion, up 1,300 percent, so the ratio fell from 251 percent of revenue in the first quarter to 48 percent in the second. That ratio is doing the work the book assigns to utilization: it is not the discount that made the commitments rational, it is the demand that showed up on top of them.³

The revenue side is shaped differently from the cost side, and the filing says so plainly: revenue is “primarily consumption-based, and much of it is not subject to long-term contractual commitments, while our compute and infrastructure commitments generally extend over multiple years.” The top five customers are about a quarter of revenue, many without long-term contracts. That is the mismatch the book’s take-or-pay chapter is about: a floor shaped like years against demand shaped like tokens (chapter 20). The filing also prices the failure mode in its own words, as a risk-factor heading: “we may be committed to excess compute that remains underutilized.”

Demand is not the only inelastic side. The filing commits to the physical inputs too: twenty-year initial terms on some power and data center arrangements, with annual escalators, take-or-pay compute obligations, and “uncapped fuel cost pass-throughs that apply for the duration of the term regardless of actual utilization.” Anthropic has publicly committed to covering the cost of electricity price increases its data centers cause to consumers. Grid interconnections take years, water for cooling is scarce in some regions, and the chips come from a concentrated Taiwan supply chain Anthropic says it has “not always been able” to draw from on time. When demand moves faster than supply, a twenty-year fuel pass-through is the sound of a capacity planner buying certainty at any price available (chapter 2).

What the filing admits it cannot know

The most striking passage is the shortage one. Anthropic states it has “historically operated in a compute-constrained environment in which demand for compute has exceeded the supply available to us,” and that it has sought quantities of compute “in excess of what those partners were willing to provide.”⁴ The buyer of $269 billion in floors could not buy as much as it wanted. Its own growth ceiling is not customers, it is “limited principally by the availability of compute.”

The response is Anthropic Direct: leasing chips directly in dedicated facilities, splitting the commitment so data center capacity, which is slow to secure and a smaller share of cost, is locked early while accelerator purchases wait until demand is clearer. And the chain that delivers it is longer now: chips procured through one counterparty, leased by a financing entity, installed in facilities built by another, operated by a third, with cross-defaults meaning one missed commitment can accelerate the rest. The book’s warning that obligations grow their own management (chapter 23) is now a disclosure item in a public filing.

The take-or-pay break-even does not care what is being committed. At 70 percent utilization the discount wins; below it, the floor becomes a bill. Anthropic has signed floors into the 2030s on the belief that utilization stays above the line, and the filing itself prices what happens if it does not.

Sources