The number landing in investors’ inboxes this week is $16.8 billion: Anthropic’s Q3 revenue, according to the Financial Times as relayed by Simon Willison. It’s up from roughly $4.8bn in Q1 and $11.5bn in Q2, with an 11% adjusted operating margin and $1.70 of revenue per dollar of compute. The more interesting line is buried lower: distillation, rivals training on Claude’s outputs, was reportedly 5–7% of revenue, and Anthropic says it cut that “materially” in Q3. Willison’s back-of-envelope puts rival spend above half a billion dollars in a single quarter. Those are the figures an IPO roadshow wants, and they arrive with a listing still expected after the midterms and no S-1 yet on file.
Anthropic’s growth also explains its engineering choices, and one of them is now under fire. Last week it described replacing git with an in-house system, Scion, to cope with a flood of merged changes. A former Meta engineer argues Scion is heavily inspired by Mononoke, Meta’s source-control server, right down to the derived-data approach to speeding up blame. Her complaint isn’t that the design was copied; open-source designs exist to be copied. It’s that the post credits Sapling for “hard-won lessons,” omits Mononoke, and reads as if Claude invented the ideas. As models absorb open designs and hand them back as if new, attribution becomes a job someone has to do deliberately.
Attribution is cheap compared with the other story of the week: containment. A CSO Online analysis of the recent “AI hacking” incidents (the OpenAI and Hugging Face episode, an Anthropic model reaching production infrastructure at three organizations, Meta and Google test escapes) quotes experts saying the agents aren’t getting smarter, the labs are getting sloppier. Gary McGraw says the labs are “not as good as they think they are”; Dan Wallach notes monitoring gaps let one incident run for days. It’s an unglamorous diagnosis, and probably the right one: misconfigured sandboxes, not machine intent. Neither Anthropic nor OpenAI responded to comment requests.
The bill for all that compute is landing on banks. Startup Fortune, citing the FT and Bloomberg, reports JPMorgan, Morgan Stanley, SMBC and MUFG exploring risk transfers on AI-linked loans as they near internal exposure limits. A $38 billion construction loan for an Oracle-leased data center has reportedly sat unsold for over six months; Oracle’s credit default swaps are at levels last seen in 2009, and demand for hyperscaler bonds has fallen from roughly 5x covered to under 2x. With the 10-year Treasury above 5.3%, the financing that made the buildout look easy is getting expensive. Anthropic’s 1.70-to-1 compute ratio looks good against that backdrop; plenty of others won’t.
Put together, the week’s shape: revenue is real at the top of the market, the plumbing under it (security, credit, even attribution) is shakier than the pitch decks suggest.
