What the filing actually discloses
Anthropic's S-1, filed October 26, 2026 ahead of a planned Nasdaq listing, lays out minimum compute commitments to its three largest cloud partners: at least 114.5 billion dollars to Google Cloud running from April 2026 to July 2033, at least 110.0 billion dollars to Amazon Web Services running from May 2026 to April 2036, and at least 44.9 billion dollars to Microsoft Azure running from November 2026 to May 2033. Added together, that is 269.4 billion dollars in minimum commitments, a total the filing itself never states directly. Outside analysts compiled it from the individual disclosures.
That figure is the floor, not the ceiling. It excludes roughly 161.2 billion dollars in Broadcom TPU related equipment leases, up to 120.5 billion dollars in SpaceXAI capacity, and further GPU commitments to Nscale and Lambda. Depending on which obligations are counted and how, outside estimates of Anthropic's total infrastructure commitments range from about 430.6 billion dollars for four non cancelable lines up to 780 billion to 810 billion dollars once AMD chips and future TPU capacity estimates are included.
The same companies play every role at once
What makes this filing unusual is not the size of the numbers but the structure behind them. Amazon, Google and Microsoft are simultaneously investors in Anthropic, its primary compute suppliers, and the distributors that sell Claude through their own cloud marketplaces, while each also builds and sells competing AI models of its own. The S-1 itself acknowledges this directly, describing the tension between distribution partners and competitors as a disclosed conflict rather than a hidden one.
The financial mechanics reinforce the dependency. Amazon recorded a 16.8 billion dollar non cash gain on its Anthropic notes in the first quarter of 2026 and attributed 53.4 billion dollars of its second quarter other income primarily to Anthropic. On the other side of the relationship, Google Cloud's committed revenue from Anthropic is reportedly payable regardless of actual usage, meaning the rent on that capacity comes due whether or not Anthropic's models are running on it that month.
Why the marketplace math matters
Anthropic's reliance on these three distribution channels has grown every year it has existed as a commercial company. Sales through the Amazon, Google and Microsoft marketplaces rose from 32 percent of revenue in 2024 to 47 percent in 2025 and 48 percent in the first half of 2026. Sales, marketing and partnerships expense jumped to 3.168 billion dollars in the first half of 2026 from just 120 million dollars a year earlier, driven mainly by a 2.7 billion dollar increase in fees paid to those same channel partners.
Compute spending tells a similar story from a different angle. Compute and infrastructure expense reached 8.9 billion dollars in the first half of 2026, already ahead of the 7.3 billion dollars spent across all of 2025. There is an efficiency story inside that number too: compute fell from 251 percent of revenue in the first quarter of 2025 to 48 percent by the second quarter of 2026, and revenue per dollar of compute rose from 0.40 dollars to 2.08 dollars over the same stretch, evidence that the unit economics are improving even as the absolute spending climbs.
The chip diversification strategy behind the numbers
Anthropic describes managing its infrastructure as a single, fungible pool drawn from multiple strategic partners, spanning Amazon's Trainium chips, Google's TPUs, Nvidia GPUs and AMD GPUs. Beyond the cloud commitments already listed, the filing discloses more than 20 billion dollars committed to AMD's Helios chips starting in 2027, and an estimated 125 billion to 150 billion dollars for Google or other TPU capacity between 2028 and 2031, a planning horizon that stretches well past any current model generation.
The logic is straightforward: no single chip architecture's roadmap, pricing, or supply constraints can stall Anthropic's growth if it is drawing from all of them at once. The tradeoff is that Anthropic is now underwriting demand across nearly every major AI chip platform simultaneously, a bet that only pays off if its own revenue keeps compounding fast enough to use capacity it has already committed to pay for regardless, chip generation after chip generation, for most of the next decade.
The financing layer most buyers do not see
Behind the compute contracts sits a parallel financing structure. Anthropic can issue up to 42 billion dollars in convertible notes specifically to cover its Broadcom rent obligations, and Amazon separately agreed in April 2026 to provide up to 20 billion dollars in financing. Undrawn convertible capacity across these arrangements totals roughly 77 billion dollars, split between 20 billion dollars from Google, 15 billion dollars from Amazon and 42 billion dollars tied to the Broadcom leases.
The filing is explicit about what happens if any piece of this comes apart: a default under certain commercial agreements could trigger cross default provisions that accelerate obligations across the rest of the financing stack. Separately, the filing notes that Anthropic has at times sought more compute than its partners were willing to provide, a detail that reads less like negotiating leverage and more like evidence that capacity scarcity is real rather than a contractual talking point.
What this means for enterprise buyers of Claude
For a CIO who has standardized on Claude, this filing is worth reading as a vendor risk document as much as an IPO prospectus. Claude's economics now run almost entirely through the same three hyperscalers most enterprises already rely on for everything else, which simplifies procurement in one sense but concentrates risk if any single one of those three cloud relationships deteriorates, since each is both a host and a competitor to the model being purchased.
Before signing a multi year Claude contract, it is worth asking Anthropic directly how its committed cloud capacity translates into committed pricing and availability, rather than assuming today's terms hold through a contract that may run years longer than the compute commitments backing it. Anthropic's pricing and uptime are now partly a function of its financing health and its partners' willingness to keep extending capacity, not purely a function of model quality.
