Anthropic's Updated S-1: $33.24 Billion in Nine-Month Revenue, and an $895.4 Billion Compute Commitment
One prospectus, two numbers side by side: revenue up twelvefold, and a promise to spend $895.4 billion — a figure roughly an order of magnitude larger than a full year of that revenue.
Three key facts
Anthropic updated its IPO prospectus with full third-quarter results. According to The Verge, the company reported $33.24 billion in revenue through September 30, up 1,200% year over year, alongside a $104 billion net loss that is mostly accounting charges. For the quarter alone, revenue was roughly $17 billion with a net loss of $23.29 billion.
The most striking figure is the commitment: $895.4 billion, "mostly on compute and infrastructure." That's roughly a decade of revenue pledged in advance to hardware. The Verge's author notes that after backing out accounting effects and items like employee compensation, adjusted income for the first half of 2026 was $1.62 billion — meaning, at current prices, the business is actually profitable.
The enterprise customer base is its hardest asset. The report notes net revenue retention of 900%, with about three-quarters of first-half 2026 customers spending at least $100,000 a year. Anthropic lists only OpenAI and Google as competitors in the filing.
Wang Dou's Take
Start with that $895.4 billion. It isn't a "plan," it's a commitment — and in a prospectus, a commitment is a legal obligation. A company with just over $30 billion in annual revenue pledging to spend $895 billion can only close the gap two ways: raise more money, or grow revenue explosively.
That explains why The Verge titled its piece "build God or die trying." Read the numbers together and the business plan requires recursive self-improvement — AI dramatically improving the efficiency of building AI — otherwise those compute commitments become a pile of depreciating hardware nobody buys. That's not pessimism; it's arithmetic.
The second thing worth pondering: this is actually a remarkably disciplined company. No image generator, no short-video clone, no companion product — just coding and cybersecurity, the things computers are easiest to point at. The payoff shows: 900% net revenue retention, three-quarters of customers paying $100,000+ a year. In an AI bubble, the path that actually makes money isn't building the coolest product — it's serving the most boring enterprise need.
Third, that "we list only OpenAI and Google as competitors." In a prospectus that's a legal statement, not marketing. It means Anthropic itself doesn't consider other model labs or open-source models to be peer-level threats. That confidence is either real, or the market will recalibrate it with a valuation after listing.
