Anthropic’s business plan: Build God or die trying

Anthropic's numbers are real. Revenue grew fourteenfold to $16.3 billion in the first half of 2026, and The Verge's Elizabeth Lopatto calculates about $1.62 billion in adjusted income once a large accounting loss and other charges are set aside. AI can pay for itself, at least for this company and at least for now.

The commitments are just as real. Lopatto adds up at least about $850 billion in compute and infrastructure obligations over as long as 20 years, rising toward a trillion if pending deals close. PitchBook's Harrison Rolfes models roughly $90 billion a year in payments from 2027 to 2030. The company has to keep growing to meet them.

We want to give Anthropic its due. It skipped the image generators and companion bots and stuck to coding, cybersecurity, and enterprise customers. Its net revenue retention is 900 percent, and about three quarters of its customers spend at least $100,000 a year. That is discipline, and it is the kind of useful AI we have argued for: tools businesses pay for because they work.

The weak spot is the cost structure. Revenue jumps after each new model and then flattens, so the company can never stop training. Competitors include Google, which can bundle AI into products customers already pay for, plus open models that run a few months behind. Every new Claude also cuts customer costs. Opus 5.5 was 40 percent cheaper than Opus 5. Analyst Gil Luria points to cloud providers that cut prices every year and still grew margins, and that is the optimistic case. The pessimistic case is a slow race to the bottom on token prices while the fixed bills keep arriving.

This is where the story stops being only about money. The filing says more capable models help researchers find improvements faster. It also warns that using AI to build AI could advance capabilities faster than the safety, alignment, and security work meant to keep up. Put plainly, the financial plan seems to depend on the very acceleration the risk section warns about.

I'm not willing to call recursive self-improvement a fantasy or a certainty. Nobody outside the lab knows. But I am worried about a company whose balance sheet rewards moving faster than its own safeguards. If that incentive scales across the industry, safety turns into whatever is left after the payment schedule. That is backwards, and investors, customers, and regulators should push on it now, before the commitments come due.

Our read: the business is better than skeptics assumed, and the bet is bigger than boosters admit. Neither side should get comfortable.

These figures come from Anthropic's S-1 and The Verge's analysis of it, including the commitments and contingencies section. Some purchase-commitment amounts in the filing are blank, and third-quarter numbers are not included. Futurum Group's Brendan Burke says the filing captures a happy period when demand outran supply and Anthropic could charge more for scarce compute.