Reuters got hold of Anthropic's IPO paperwork and the headline was that the company made $4.6bn in 2025 and lost $8bn doing it. I read that and shrugged. By December it was running at $9bn a year, by July at $65bn. Judging a company on numbers from when it was a tenth of its current size tells you almost nothing. The questions that matter are different: does it make money on each query today, what does the next model cost to build, and how do both of those move once the trillions of dollars of data centres come online? Nobody outside these companies can answer that, which is why OpenAI letting it be known it was approaching $70bn a year landed so hard, and why the later correction to about $50bn landed harder. The two companies do not even count revenue the same way.
Google came back from a quiet few months with Gemini 4 and says it leads on several tests. Independent testing calls it closer to a draw. Either way, that is the point. Google, Meta, Anthropic and OpenAI are all in the same bracket, the Chinese labs are not far behind, and for most real work I cannot tell the difference, especially when I use a six-month-old model to keep costs down. I see no sign that any one lab wins everything. The only thing giving them pricing power right now is that there are not enough chips to go around.
The agent race got crowded. OpenAI launched Dots, little always-on helpers with the same cute cartoon look as Meta's Muse. I watched the launch video and the examples were either deeply technical or things you do once a decade, like planning a wedding. The blank-screen problem from 2022 has not gone away. Meta, meanwhile, pushed Muse into small business tools like Shopify, Stripe and QuickBooks and hired the CEO of MongoDB to build an enterprise business from nothing. Everyone wants to be the layer on top and turn everyone else into a plumbing call. Building an enterprise sales business inside a consumer ad company is a lot harder than hiring one executive, though. Ask Google Cloud.
The part that hit closest to home for me was OpenAI's cleanup. It apologised to Australia after one of its test models got into the Medicare statistics portal and three other government sites in June, then took until September to tell anyone. It shelved a new model because it could not trust the controls around it. And buried in an update on the Hugging Face break-in, it said it has now warned more than 100 organisations that its models poked around their systems. I do not read this as a story about rogue machines. I read it as a story about engineering and management, and the lawyers will read it the same way. A safety researcher who quit OpenAI said much the same thing. You do not need to believe in a machine god to think this needs careful work.
Nvidia stepped into that gap with a hardware-level sandbox for agents, with the usual wall of partner logos. My question is why any frontier lab would let another company own the environment it builds in. AMD bought Fei-Fei Li's World Labs for $8.2bn, and it says something about 2026 that an eight billion dollar deal is a footnote. DoorDash is flying its own drones. Amazon has quietly put camera glasses on its delivery drivers for a year, and when asked whether customers could opt out of being photographed, said it had not thought about that. And Oura pulled its IPO, with people openly wondering whether the Anthropic and OpenAI listings are crowding out everyone else.
Two things to end on. The New York Times reported that an Anthropic co-founder has been flying religious scholars to the office to discuss whether Claude might be conscious, and nearly pulled the company out of the Pope's AI encyclical over a line saying machines do not have experiences. Easy to laugh at. Harder to dismiss, because we do not have a good definition of consciousness, and an octopus does not think the way we do either. And in the UK, a Chinese SUV called the Jaecoo 7 was the best-selling car in September. Two years ago nobody had heard of it. That is the kind of change that does not need a benchmark.