Anthropic's S-1 lays out an $850 billion compute bet ahead of Nasdaq listing

Anthropic has filed an S-1 registration statement to go public on the Nasdaq, and trading could begin as soon as next month at a valuation between $1.5 trillion and $2 trillion, according to The Verge. The hundreds-of-pages document is the first detailed public look at the finances of a frontier AI lab, and it describes a company with remarkable revenue — and spending commitments to match.

What the numbers show

According to The Verge, Anthropic's revenue grew fourteen-fold to $16.3 billion in the first half of 2026 versus the first half of 2025. The filing shows an $80 billion loss, but The Verge reports that figure is largely an accounting artifact; excluding it and other charges, including employee compensation, leaves $1.62 billion in adjusted income for the half.

Joey Brookhart, an analyst at SemiAnalysis who co-leads the firm's Tokenomics research, told The Verge's Decoder podcast that the company scaled efficiently to more than $70 billion in annualized recurring revenue, with gross margins, compute margins and free cash flow all looking strong. He linked the filing to a renewed rally in semiconductor, neocloud and compute stocks.

The revenue is enterprise-heavy and concentrated. About three-quarters of customers in the first half of 2026 spent at least $100,000 a year, and The Verge cites net revenue retention of 900 percent. Brookhart said spending skews heavily toward coding — he called Anthropic almost a single-vertical company — with the top 1 percent of customers spending $50,000 to $100,000 per employee. The two leading labs together account for roughly $120 billion of ARR, by his estimate.

The compute bill underneath

The filing's most striking figures are its forward obligations. Anthropic discloses $265.6 billion in fixed, non-cancelable purchase commitments with terms of one to 20 years, plus $321.4 billion in not-yet-commenced leases primarily for compute and infrastructure capacity as of unaudited June 30 figures, and roughly $245 billion in new compute and infrastructure agreements. Additional transactions that have not been finalized are estimated at $125 billion to $150 billion, and The Verge notes the commitments run through 2045.

The Verge's own tally puts the minimum at around $850 billion over 20 years, rising toward $1 trillion if the pending deals close. Harrison Rolfes of PitchBook models the payment schedule at about $90 billion a year from 2027 to 2030, implying the company needs roughly $188 billion in revenue to keep its current ratio of cost to compute stable. "Growth is not optional for this company," he writes. "The contracts assume it."

During the first half of 2026, compute and infrastructure tied to customers cost $2.3 billion, while model development cost $9.7 billion — and Rolfes says Anthropic became more efficient over the period, spending less on compute for each dollar of revenue.

Competition and pricing pressure

The S-1 names only OpenAI and Google as competitors, but the field is wider. Claude Opus 5.5 currently sits atop the Epoch Capabilities Index, though The Verge notes such leaderboards change frequently, and open-weight models serve buyers who can tolerate lagging the frontier by a few months.

Brookhart's estimates show how quickly the landscape shifts: Anthropic captured 90 percent of the industry's net new ARR in the first half of the year, but OpenAI took the majority in the third quarter, while open-source inference ARR grew from about $1 billion in the first quarter to more than $10 billion in the third. He also pointed to Google's DeepMind and Meta's superintelligence lab as rivals building comparable training compute.

Pricing is another pressure point. Claude Opus 5.5 cut customer costs by 40 percent versus Opus 5, and Anthropic argues that cheaper tokens drive enough additional usage to preserve margins. Brendan Burke of Futurum Group cautions that the S-1 freezes a favorable moment when usage outstripped available capacity, and that frontier labs need high-margin products to avoid a race to the bottom on token prices. Gil Luria of DA Davidson offers a counterexample: cloud providers such as AWS cut prices for years while margins kept rising.

The AGI assumption underneath

To reach the trillion-dollar revenue that CEO Dario Amodei has said he can envision, Brookhart argues the industry must expand beyond coding into white-collar work, shifting from AI augmentation toward actual labor replacement. The Verge's analysis of the filing draws a similar conclusion from the commitments themselves: the plan leans on models improving rapidly enough — including through recursive self-improvement — to generate the growth the contracts already assume.

Why it matters

This is a landmark filing for both the AI industry and public markets. Investors are effectively being asked to underwrite hundreds of billions of dollars in compute spending through 2045 on the premise that AI continues to absorb white-collar work. The unit economics look solid today, but they rest on a narrow, coding-heavy customer base, intensifying competition from OpenAI, Google and open-source models, and a bet that capability gains keep compounding. If they do not, the fixed costs remain either way.