Don’t expect supersize profits from Anthropic soon
Anthropic publicly filed its IPO paperwork on 26 October, with a listing due within weeks that would be close to the largest ever. Reported figures: revenue rose 30-fold from $376m in Q1 2025 (when Claude Code launched) to $11.6bn in Q2 2026, more than McDonald’s or 3M; Q2 brought an operating profit of just under $1bn (an 8% margin) after a $1.7bn loss the quarter before. Compute spending fell from 251% of sales to 48% over the same period and personnel costs from 58% to 16% of revenue, which Anthropic credits to more efficient models. The Economist argues caution is warranted. Growth is slowing: Anthropic told shareholders Q3 revenue was up about 45% quarter on quarter, a third of Q2’s rate, and the prospectus warns growth could slow “more rapidly than we anticipate”. Big customers such as Meta and Microsoft are pushing staff to in-house tools, OpenAI’s Codex daily use is up tenfold since early August, and firms are policing AI budgets after the “tokenmaxxing” craze. Almost half of revenue now comes through the clouds of Alphabet, Amazon and Microsoft, and sales, marketing and distribution took 18% of revenue in Q2 2026, up from 6%. Compute is scarce and pricier: the prospectus warns more short-term compute could hit margins, and Anthropic has agreed to pay SpaceX about $3.1bn a month for Nvidia chips into 2029 (with an exit from early next year on 90 days’ notice). Its new Anthropic Direct programme will buy chips (e.g. from Broadcom) for rented data centres, which should control operating costs but makes the business more capital-intensive and could require more fundraising. The paper’s conclusion is that investors hoping for a valuation as high as $2trn may be disappointed; the founders counter that spending which “may seem irresponsibly large” is justified by AI’s exponential trajectory, and promise to be “bold but sane”.
Why it matters
Anthropic’s IPO will set the public-market benchmark for every frontier-lab and AI-application valuation Dealroom tracks, and the prospectus gives rare audited-style numbers on AI unit economics: compute and staff cost ratios, cloud-channel dependence and the swing to profit. Slowing growth and rising capital intensity would ripple into late-stage AI rounds, secondaries and the OpenAI listing timetable.
Executive takeaways
- Reported: revenue $376m (Q1 2025) to $11.6bn (Q2 2026), 30x; Q2 2026 operating profit just under $1bn (8% margin) after a $1.7bn Q1 loss.
- Reported: compute fell from 251% to 48% of sales and personnel from 58% to 16% of revenue, so scale and model efficiency are real.
- Reported: Q3 growth about 45% q/q, a third of Q2’s pace; the prospectus itself warns growth could slow faster than expected.
- Reported: nearly half of revenue comes via Alphabet, Amazon and Microsoft clouds, and sales/marketing/distribution rose to 18% of revenue, so distribution is getting costlier and more concentrated.
- Reported: about $3.1bn a month to SpaceX for Nvidia compute into 2029 (cancellable on 90 days’ notice from early next year) and a move to own chips via Anthropic Direct.
- Economist view: more capex means less free cash flow and possibly more fundraising, so a $2trn valuation hope looks stretched; the founders’ answer is ‘bold but sane’.
What The Economist may be missing
Thin on valuation itself: the piece doesn’t give the expected IPO valuation range, share structure or how the price compares with the last private round, so the ‘$2trn may disappoint’ point is hard to size. It looks at one quarter’s 8% margin without asking how much is accounting timing, and doesn’t compare Anthropic’s numbers with OpenAI’s or the clouds’ own model businesses. The 45% q/q Q3 figure came from a private shareholder meeting rather than the prospectus. There’s also little on what slowing coding-agent demand means for the wider application layer (Cursor, Cognition and others built on Claude), or on European or sovereign demand for Claude outside the US clouds.
