What it is
“Non-cancelable compute obligations” is not a technical AI term — it’s a contract and accounting concept describing a company’s binding commitment to buy a fixed or minimum amount of computing capacity (cloud servers, chips, data-center space) over a set period, regardless of whether it ends up needing that much. The obligation is “non-cancelable” because the buyer cannot exit the contract without penalty, and in many cases must keep paying even if it uses less capacity than contracted — a structure known in accounting as “take-or-pay.”
This pattern came into public view on September 28-29, 2026, when Reuters reported on Anthropic’s IPO prospectus (filed confidentially with the SEC and not yet public on EDGAR), which disclosed roughly $518 billion in infrastructure commitments over the next seven to ten years. According to that reporting, about 80% of the total is either non-cancelable outright or requires Anthropic to pay the contracted amount “regardless of usage.” The largest pieces: $111.1 billion with Google (April 2026-July 2033), $110 billion with Amazon (May 2026-April 2036), $31.4 billion with Microsoft (non-cancelable except for an uncured material breach), and $161.2 billion in largely non-cancelable equipment leases tied to Broadcom. A separate, more flexible arrangement of up to $84.5 billion with xAI reportedly allows cancellation with 90 days’ notice — one of the few carve-outs in the portfolio.
The underlying accounting concept is older than the AI boom. The SEC has required companies to disclose “purchase obligations” — enforceable agreements specifying fixed or minimum quantities and pricing — in the contractual-obligations section of their financial filings since a 2003 rule on off-balance-sheet arrangements. What’s new is the scale: AI labs are using these instruments to lock in chip and data-center capacity years in advance, in a market where compute supply, not model ideas, is treated as the binding constraint.
Why it matters for AI governance and narratives
The phrase matters to the observatory because it marks a shift in how the “AI buildout is unsustainable” narrative gets contested. Earlier rounds of skepticism about Anthropic’s prospectus, per the editorial’s framing, focused on the company’s own disclosed “existential risk” language — a story about safety and governance. The compute-obligations detail redirects that same document toward a financial-plumbing story: whether the buildout is being financed through structures that transfer risk from cloud providers to AI labs (and, some argue, back to the providers themselves, since Google, Amazon, and Microsoft are simultaneously investors, vendors, and competitors of the companies they’re locking into these contracts). That circularity — one industry newsletter reportedly tallied $252 billion of such commitments concentrated among a small number of counterparties — is itself a narrative fault line: is the AI capital buildout evidence of durable demand, or of a small number of interlocking firms manufacturing the appearance of demand for each other’s balance sheets? Non-cancelable obligations are the mechanism that makes this question answerable from public filings rather than speculation, which is exactly why financial and industry press seized on the term once the numbers surfaced.
Key facts and dates
- Anthropic confidentially submitted a draft S-1 to the SEC on June 1, 2026; as of late September 2026 it had not yet appeared publicly on EDGAR, meaning press reporting is based on details that circulated to reporters rather than a document available for independent verification by the observatory.
- Reuters’ reporting, published September 28-29, 2026, is the primary source for the $518 billion figure and the ~80% non-cancelable share.
- Individual commitments reported: Google $111.1B, Amazon $110B, Microsoft $31.4B, Broadcom equipment leases $161.2B, xAI up to $84.5B (cancelable), AMD over $20B plus a $5B stock purchase commitment.
- The SEC’s own definition of a “purchase obligation” — binding agreements specifying fixed or minimum quantities and pricing — has governed this category of corporate disclosure since 2003, independent of the AI industry.
- Analysts caution that “non-cancelable” is not synonymous with “risk-free”: actual enforceability depends on contract-specific ramp schedules, cure periods, and material-breach clauses that are not fully visible from summary reporting alone.
Where to learn more
- Reuters, via TechTarget: “Anthropic locks in $518B AI infrastructure commitments”
- Reuters, via The Express Tribune: “Anthropic’s $518b AI buildout hinges largely on deals that cannot be canceled, filing shows”
- Anthropic: “Anthropic confidentially submits draft S-1 to the SEC”
- SEC: Disclosure of Management’s Discussion and Analysis about Off-Balance Sheet Arrangements and Aggregate Contractual Obligations