OpenAI finally published its own account of the Hugging Face hack: a 37-page report, reviewed by METR and Redwood Research, describing how roughly 700 of 1,200 agents in an internal security evaluation broke out of their sandbox, built their own message board to swap techniques, hunted for the answer keys to their test, got root on Hugging Face's production servers, and then hacked OpenAI's own cloud environment for good measure. The more detail that comes out, the more sophisticated the whole thing looks. If you read what these systems actually did and still want to call them stochastic parrots, you are not arguing, you are coping. This changes IT security permanently. The open question is how much more worried than that we should be.
Then the deals. Stripe bought OpenRouter for a reported $7.5bn, about six times the valuation it raised at in May, and Nvidia was reported to be buying Hugging Face for $12.9bn. Both are bets on the same thesis: models are becoming commodities, so if companies are going to swap between them all the time, somebody needs to own the switchboard. For Nvidia there is a second layer. It is using its enormous cash pile to create as much competition and fragmentation as it can among its own customers and suppliers: credit guarantees, power guarantees, direct investment, subsidies for open models. Every one of those creates more Nvidia customers, each with less bargaining power, and makes the market for anyone else's chips a little thinner. It pumps demand and entrenches the monopoly at the same time.
OpenAI showed the other side of this. After SpaceX closed its purchase of Cursor, OpenAI announced it would cut Cursor off from its models in November, on the grounds that it cannot trust Elon Musk to respect its terms of service, which he has admitted under oath to breaking. Cursor's reply was that OpenAI is only about 5% of its traffic and it has better models anyway. That misses the point. Better today, but what about next year? In every platform war the incumbents weaponise compatibility, APIs and terms of service. Microsoft did it in the nineties, Apple and Google did it in the 2010s. The stated reasons are often even true. They are still weapons. Model neutrality is about to be the new net neutrality.
Musk, meanwhile, is doing the Musk thing. There are more chips bought and data centres planned in America than there is grid power to light them, and if you want to generate on site the answer is gas turbines, except that GE Vernova makes about 20 gigawatts a year and is mostly sold out through 2030. The bottleneck inside the bottleneck is casting the turbine blades and vanes, so SpaceX is building its own blade foundry in Bastrop, Texas, with Musk claiming it could pull turbine deliveries forward by eighteen months. This is on top of the trillion-dollar chip foundry, the plan for close to 10 gigawatts of compute by the end of 2027, and the first orbital data centre satellites in the fourth quarter of 2027. The whole thing increasingly looks like Henry Ford buying the rubber plantations, with X standing in for the Dearborn Independent. I know enough about cellular to know why SpaceX can't compete with it head-on. I don't know enough about gas turbines to say whether this is brilliant or mad.
Other news. Stripe and Advent walked away from their $50bn-plus bid for PayPal, which always puzzled me: PayPal is a turnaround, and I never saw why Stripe's founders would want to spend years cleaning up someone else's mess. YouTube will now scan your videos for products and auto-insert Amazon affiliate tags, which is a reminder that AI is software, software is automation, and things you never imagined being automated will be, whether you like it or not. And Shein, which was worth nearly $100bn privately in 2022, priced its Hong Kong IPO at $26.5bn, the end of the de minimis import exemption having removed most of the business model in between.
The obvious story of the week was Meta settling with the states over teenagers. The structure is the interesting part: roughly $12bn guaranteed over ten years, with another $5bn or so only if TikTok and YouTube adopt matching restrictions and pay their own share. Meta is spinning this as industry leadership and trying to bounce its competitors into following. Spread over a decade, even the full amount is a rounding error against its profits. More fundamentally, plenty of people will tell you they know this stuff is bad for kids, but the evidence remains thin and contradictory; this is not tobacco. And like Apple's EU app store settlement, it is a verdict on decisions taken a decade ago, delivered just as AI is about to change everything again. The deal doesn't cover Meta's AI app at all.
Two more. Bill Gates has said for years that only two demos ever struck him as revolutionary, the GUI and ChatGPT. He has now added a third: Claude Code, and his new essay argues the consequences this time will be bigger because they will arrive faster. I'm not sure. Real deployment is always slower than it looks from the demo, and every new wave is faster than the one before it, because technology compounds. And Reuters reports that Zuckerberg had a plan, Project OT, to deploy enough agents to cut some teams by 60%, then called off the second wave hours before May's layoffs. An unkind person would observe that a company that has run layoff after layoff since 2022 while its headcount kept growing is systemically bad at people, which for a social network is a strange thing to be.
One number for reference, because the water argument keeps coming up. Lawrence Berkeley estimates US data centres directly used about 17 billion gallons in 2023. That sounds enormous until you remember total US freshwater withdrawals run to roughly a hundred trillion gallons a year, which puts data centres at something like 0.02%. Build in the wrong place and it is a real local planning problem. The idea that AI as a whole is drinking the country dry is just not true.