The story that stuck with me this week was a correction. For a few weeks everyone repeated that OpenAI was making $70bn a year. Then it turned out the real figure was closer to $50bn. Nobody lied, exactly. OpenAI counts only its own share of sales made through partners like Microsoft, while Anthropic counts the whole amount, and somewhere between the investors and the press the two got mixed up. I have been in this industry long enough to remember when people invented new metrics because the old ones looked bad. We are not quite there yet, but when a $20bn gap can appear and disappear depending on who is counting, you should hold every AI revenue number loosely.
The agent story keeps moving in two directions at once. Meta and Sierra announced a standard for how personal AI agents should identify themselves to businesses, with Walmart, Stripe and Shopify on the slide. The names missing from that slide, OpenAI and Anthropic, tell you as much as the ones on it. Meanwhile Amazon simply blocked Meta's agent from its shop. Part of that is caution about fraud. Most of it is that Amazon makes an enormous amount of money from ads, and an agent doing your shopping never sees an ad.
Two stories hit close to home for me. Mistral released a new flagship model that puts France back near the front of the open-weights pack, after months where most serious open models came from China. In the US, Reflection released its first model and Nvidia is already talking about buying the company. And in Australia, Firmus pulled its $5.5bn listing days after saying demand was strong. Part of that is nerves about the AI buildout being funded by borrowed money. Part of it is that Anthropic's giant IPO is sucking the air out of the room for everyone else.
A few smaller things worth knowing. Anthropic changed its terms so you cannot be deliberately cruel to its models, which some researchers find reasonable and others find absurd. OpenAI started watermarking ChatGPT text, but only in the EU, and was fairly open that it does not think the technology works well yet. Google launched a tool that builds playable games from a text description, which I think matters more than it looks, and a Mac app that takes meeting notes entirely on your machine with the memorable name "Google AI Edge Foresight".
Then there is SpaceX, which paid about $8bn for the kind of spectrum you need to run a normal mobile network, and the US carriers had their worst day in years. I am sceptical. Satellites are great for open sky and useless indoors, so if SpaceX really wants to be a phone company it will have to build towers like everyone else and will end up charging what everyone else charges. And Apple has started quietly blocking advertising companies at the operating system level, which is really a policy decision dressed up as an engineering one, made with no debate on devices half of America carries around.
Two pieces I would send to anyone. Cryptographer Matthew Green argues that AI will fix bugs so fast that police will lose the ability to break into phones, which would restart the whole encryption fight. And Satya Nadella wrote that companies should treat AI agents like a potential insider threat, with a human-controlled off switch the model cannot reach. Last, Gallup's global numbers show the usual pattern: the richer the country, the less its people like AI. The US sits near the bottom.