Anthropic's Compute Commitments Near $900 Billion as Operating Losses Resurface in Q3
nashnova research
Anthropic's updated SEC filing reveals $895.4 billion in cumulative compute commitments — 6.8× its cash on hand. Q3 revenue surged 12-fold to $16.96 billion, but a scramble for short-term compute flipped the quarter to an operating loss, laying bare the bill for the AI arms race.
A $900 billion compute tab — where does the money come from?
Through purchase commitments and lease agreements, Anthropic has locked in $895.4 billion in cumulative spending obligations, overwhelmingly tied to compute and infrastructure.
Cash and short-term investments stand at just $131.8 billion — the commitment is 6.8× the cash on hand. This means → years of compute spending already far exceed what the company can fund from its balance sheet; continuous fundraising is non-negotiable.
Roughly $447.5 billion comes due by end-2030; the rest stretches to 2045. About $69.8 billion is due within 12 months — 53% of current cash.
Revenue grew 12-fold — so why is it losing money?
Q3 revenue hit $16.96 billion, up roughly 12× from ~$1.38 billion a year earlier. Year-to-date revenue reached $33.242 billion versus $2.545 billion in the prior-year period.
But operating expenses rose to $17.5 billion, producing a $567 million operating loss — versus $974 million in operating profit just one quarter earlier. In plain terms = revenue is surging, but spending is surging faster, and costs overtook the top line.
The biggest line item: compute and infrastructure costs of roughly $9.9 billion, nearly double the prior quarter's ~$5.6 billion, eating about 58% of Q3 revenue.
The short-term compute scramble — how much did the "rush order" cost?
Customer demand outpaced Anthropic's pre-contracted compute supply, forcing the company to buy spot compute. Q3 short-term compute purchases totaled roughly $5 billion. This means → this was "rush-order" capacity, priced well above long-term contract rates.
A separate ~$2 billion non-cash charge tied to an employee donation-matching program also weighed on operating results.
Stripping out stock-based compensation and the donation match, adjusted operating margin was about 15%, down from 19% in Q2. Price cuts on new model versions and contract discounts offset roughly $600 million in incremental revenue.
Who are the contracts with?
The largest commitments cluster around a handful of tech giants: Google ~$121.3 billion, SpaceX-related agreements up to ~$120.5 billion (extending to 2036), Amazon ~$110 billion, Microsoft ~$44.9 billion; Broadcom equipment leases ~$161.2 billion.
Beyond these fixed commitments, Anthropic plans to purchase an additional $125–150 billion in Google chips and over $20 billion in AMD chips — neither included in the $895.4 billion total because the contract amounts are not fixed.
This reflects a tangled web: Amazon and Google are simultaneously investors, cloud and chip suppliers, and potential competitors. Evercore ISI analyst Mark Mahaney has described Anthropic's relationship with the big cloud providers as "customer, competitor, distribution channel, and investment target" all at once.
Can building its own data centers fix the problem?
Anthropic has begun hiring data-center construction and supply-chain specialists, purchasing servers and cooling systems directly, and exploring conversion of retired coal plants and industrial sites into data centers.
The company initially sought sites of at least 200 MW; it has since begun accepting ~50 MW projects, supplementing power with battery storage and natural-gas generation. The prospectus calls this self-built infrastructure "Anthropic Direct."
Sources say Anthropic expects to control over 3.5 GW of compute capacity by 2026 and plans to double that by 2027. In plain terms = the company is stitching together power and infrastructure project by project, across the globe, to assemble its own compute footprint.
Where is the biggest risk?
The prospectus states plainly: revenue is overwhelmingly consumption-based — pay-as-you-go with little long-term contract backing — while compute spending must be locked in years ahead. This means → on the revenue side it's "pay only for what you use"; on the cost side it's "pay whether you use it or not." If demand slows, the cash-flow gap could widen fast.
If actual revenue cannot cover contract payments — or the company cannot raise capital on acceptable terms — certain defaults could trigger cross-default clauses (one contract breach forcing other agreements into default as well) in credit facilities and convertible notes.
Whether the short-term procurement premium narrows as "Anthropic Direct" capacity scales up will be the critical test of whether this expansion logic can hold.
市场有风险，内容仅供研究参考，不构成投资建议。