Product2 publishers3 min readPublished
Anthropic's $518 billion compute pledge swaps demand risk for counterparty risk at four suppliers
Anthropic has committed at least $518 billion to compute over a decade, about 80% of it non-cancelable, Reuters reported from its IPO prospectus. Most of that sits with Google, Amazon, Microsoft and Broadcom, whose risk is now whether one customer can keep paying into the 2030s.
The Product Desk · Product desk
What happened
- Anthropic must pay Google $111.1 billion, Amazon $110 billion and Microsoft $31.4 billion under long-term service obligations over seven to ten years, regardless of usage.
- The Google payments run from April 2026 to July 2033, Amazon's from May 2026 to April 2036 and Microsoft's from November 2026 to May 2033.
- Anthropic also carries about $161.2 billion of Broadcom-related equipment leases that neither party can cancel except in the case of a default.
- A separate xAI arrangement worth up to $84.5 billion in Nvidia-based capacity through 2029 is largely cancelable with 90 days' notice.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- exposure Anthropic carries counterparty risk of its own, because Amazon, Google and Microsoft are also its investors, customers, distributors and competitors, with incentives it says may not be fully aligned.
- constraint Transformers that take three to five years to obtain and gas turbines that take more than five leave operators unable to energize all the contracted capacity on the contracts' own schedule.
- precedent With take-or-pay moving upstream, as Sopko describes it, lenders can expect developers to hold turbine and transformer commitments signed before their own end customer has fully signed.
"If our actual spend falls short, we must pay Google the difference," Anthropic said in the prospectus Reuters reviewed, adding that similar terms apply to its Amazon agreement [7]. For a data center developer, or the bank financing one, that clause addresses one of the biggest risks in funding new AI capacity: whether the customer shows up and pays [3].
Anthropic's pitch to investors is about scarcity. It says future demand for advanced AI will be "limited principally by the availability of compute" [8]. Stephen Sopko, practice lead for semiconductor and deep tech at HyperFrame Research, described what the contracts do for the suppliers. "It reduces demand risk and replaces it with counterparty risk," he said. "Non-cancelable is only as strong as the balance sheet behind the signature." [9]
Anthropic's Google, Amazon, Microsoft and Broadcom obligations add up to $413.7 billion [1]. That is about 80% of the $518 billion total [2], close to the share Anthropic calls non-cancelable or payable regardless of usage [2]. Broadcom's leases alone are about 31% [3]. Add the xAI capacity [6] and AMD's supply deal, expected to exceed $20 billion [21], and the six partners' figures come to about $518.2 billion [4].
Anthropic also describes the risk from its own side. "If the compute we have access to from third parties is curtailed, repriced, or terminated ... our business, financial condition, and results of operations could be adversely affected," it said [11]. Its Microsoft commitment "is non-cancelable except in the event of Microsoft's uncured material breach" [12].
The 80% figure invites a shortcut. "Reading '80% non-cancelable' as '80% risk-free' means you haven't read the contract," Sopko said [13]. How firm the money is depends on ramp schedules, cure periods, material-breach provisions and what happens if capacity arrives late [14].
Power is the largest constraint on turning the contracts into running capacity, Sopko said, and advanced packaging, memory, liquid cooling and skilled electrical labor can also slow deployment [24]. "The money is committed. The copper and concrete are not," he said [16].
His working rule is to treat the first two to three years of named-site, dated capacity as close to firm demand [17]. The later years of a ten-year commitment look more like an option, he said [17]. That gives an operator two tests for any single contract. The first is whether the buyer can walk away: mostly not on the Google, Amazon, Microsoft and Broadcom paper [4][5], and yes for the xAI arrangement [6]. The second is whether the capacity is named, dated and due inside three years.
Put the two tests together and every contract lands in one of four boxes. Paper the buyer cannot exit, for named capacity inside three years, is demand and can be underwritten as demand. The same paper in the later years is a credit exposure to one company, and belongs in the model at Anthropic's credit. Cancelable capacity in the near term is short-notice revenue, set against power and data center assets that can be financed over 25 years or more [18]. Cancelable capacity in the out-years is an option, and I would not build for it alone. "The smart builds are power-first, modular and phased to match contract tranches, so each phase stands on its own if the next one doesn't come," Sopko said [19].
What to watch
- A public version of Anthropic's confidential June IPO filing, letting lenders read the contract terms directly.
- The first period in which Anthropic's spend at Google or Amazon falls below its commitment and the shortfall clause becomes a cash payment.
- Whether Anthropic's move toward its own data centers and directly leased chips pulls capacity away from the cancelable xAI arrangement before 2029.