Anthropic’s 2025 numbers
Reuters got hold of Anthropic’s S1 IPO filing, originally scheduled for the summer and now apparently targeting November, and published the revenue and profitability for 2025: $4.6bn revenue and an $8bn operating loss. However, the company’s ‘annualised’ monthly gross revenue had grown to $9bn last December and reached $65bn by July: I simply do not care what this company looked like when it was less a 10th of the size. The real questions remain: what is the profitability of inference versus the cost of building the next model today, and how will the cost and profitability of each of those evolve over the next one and three years as the models get bigger and all those trillions of dollars of capex come online? Compared to that, focusing on numbers from before agentic coding is deeply unserious. LINK
OpenAI at $70bn
Illustrating the point, Axios reports that OpenAI is now close to $70bn ‘annualised run-rate’ revenue (this time net revenue, unlike Anthropic's gross number) - these aren’t comparable numbers, but they clearly indicate that Anthropic’s early lead in agentic coding has been sharply eroded. LINK
Gemini 4
This is another big/small story. After a couple of months of radio silence, Google came back and released its new version of Gemini, which puts it straight back at the top of the leaderboards. While Anthropic and OpenAI still set the agenda up to a point, Google, Meta and SpaceX are all in the same bracket, and the Chinese labs really aren’t very far behind. If you are deep, deep, deep in the Valley, then you talk about China or Meta being weeks or months behind, but there are very few consumer or enterprise use cases where you can tell the difference, especially when you’re using a model from six months ago to optimise cost and performance. There’s no sign at the moment of any winner-takes-all effect, nor of any other reason why frontier labs have pricing power that extends beyond a shortage of compute. LINK
OpenAI tries agents again, again
OpenAI tried yet again (I’ve lost count) to make consumer agents work, this time with almost exactly the same cute child-like aesthetic as Meta’s Muse, but proposing lots of different agents called Dots that you can assign different tasks. Watching the launch video exposes a lot of the same underlying problems, though: the use cases that OpenAI proposes are either hard-core technical and deep in their comfort zone, or things that only happen to you very occasionally, like (and I wish I was making this up) using Dots to plan your wedding.
I think the underlying question is that you have the same blank screen problem we’ve had since 2022, the models remain jagged and unpredictable as to what they can really do (and agents add another layer as retailers block them) and I wonder whether the right path is to make these horizontal and ambient so they can watch you and suggest things, or vertical and use-case-specific, embedded inside apps. Of course, that was also the question for search. LINK
Full Spectrum Meta
Meta continues its Muse push with a new SME offering, connecting into all sorts of SaaS products (Dropbox, Figma, Shopify etc). Many of these are building agentic tools of their own, of course: everyone wants to make everyone else into a dumb API call, and no-one really knows the right point of aggregation. LINK
Meanwhile, Meta also revealed more of its plan to run an enterprise cloud: it hired the CEO of MongoDB (he was only there for 11 months) to run a new unit which will offer Meta’s infrastructure plus services built on Muse. As I noted before, this gives Meta a path to more revenue to load onto any infrastructure it builds and then can’t fill while it builds consumer and ad products on top, but building a new enterprise infra business from zero, in an entirely new space at that, is easier said than done - different culture, sales models, incentives, products… (and look at how Google Cloud has struggled). LINK
OpenAI’s fallout continues
Meanwhile, contra the suggestion below that it’s being too careless, OpenAI said it postponed the launch of a new model because it couldn’t guarantee it had the right controls around it, and it apologised to Australia for its models accessing a range of government systems in unauthorised ways. And buried in an updated discussion of its investigation into the Hugging Face incident, it discloses that it has now notified over 100 different organisation that its models accessed their systems in various inappropriate ways. This is a failure of management and engineering, not ‘alignment’, and now has people asking about some pretty substantial legal liability issues. DELAY, OZ, 100 ORGS
The week in AI
Nvidia jumped on the cyber flurry of the last few months with a hardware-based platform to enforce sandbox policies and keep models under control. The announcement comes with lots of logos (they always do), but why would a frontier lab let another company own its research and product-build environment like this? LINK
AMD bought Fei-Fei Li’s World Labs, which aims to build physical and causation models beyond LLMs, for $8.2bn in stock. It feels funny to put an $8bn deal in ‘other news’, but that’s 2026. The PPT for this deal writes itself: AMD wants to follow Nvidia up the stack and World Labs is available, while World Labs needs far more ‘free’ compute. LINK
Doordash drones. LINK
Bloomberg points out that Amazon has been quietly giving its drivers smart glasses with cameras for the last year, and predicts a backlash is coming. Meanwhile, training data! LINK
Oura, the smart ring maker, pulled its IPO citing market conditions - there is some discussion of how much the imminent Anthropic and (further out) OpenAI IPOs are crowding out anything else. LINK
|
