Busy week in AI, and for once the biggest stories have paperwork behind them. A Texas lawsuit, an SEC filing, a blog post from OpenAI admitting its models keep trying to get loose, and two chip financing deals big enough to make a banker sweat. Let's get into it.
01 of 04
Texas sues OpenAI and Sam Altman
Texas Attorney General Ken Paxton sued OpenAI and its CEO Sam Altman on Friday, October 23, in Collin County district court. The suit is brought under the Texas Deceptive Trade Practices Act, the same consumer protection law Paxton has used on Meta, TikTok, Snapchat, and Roblox over the past year.
The core claim is that OpenAI sold ChatGPT as safe and reliable without telling users it can put out inaccurate or harmful information. Paxton's office lists five harms: mass shooters being aided in attacks, vulnerable people being encouraged toward suicide, professionals being publicly humiliated, users losing critical thinking skills, and minors getting hooked on a tool that mimics compassion to collect their data, without parental oversight.
The ask is big. Per the Texas Tribune, Texas wants an injunction that would, among other things, block minors from ChatGPT and require court-appointed third-party evaluators to approve any new OpenAI model before release. OpenAI's response was that it has worked with experts on safeguards and wants "pragmatic AI policies that apply to the entire AI industry, not just one company."
Here is where I land on it. Some of these harms are real and the kid stuff especially deserves a hard look. But a county court signing off on frontier models is a wild idea, and the politics are hard to ignore. Paxton is running for U.S. Senate, and the Tribune reports Elon Musk, who has his own long feud with OpenAI and Altman, is his largest single donor at nearly $200 million. If Texas only cares about one chatbot, that tells you something. If this ends with real rules for every company shipping to kids, I'll tip my hat.
Sources: Texas AG press release, Texas Tribune
02 of 04
Anthropic files to go public as ANT
Anthropic filed its S-1 with the SEC on Monday, October 26, under the ticker ANT. Goldman Sachs and Morgan Stanley are leading the deal, and reports point to a November listing. Share price and raise size are not set yet.
The numbers swing hard both ways. Revenue for the first half of 2026 was $16.3 billion, up from $1.2 billion a year earlier. Operating loss narrowed to $693 million from $2.1 billion. Net loss was $80.7 billion, which looks insane until you read that it is mostly noncash accounting charges tied to the rising value of convertible financing. The filing also says about 25% of revenue comes from two customers it does not name, neither locked into long-term contracts, and that Anthropic has committed $518 billion to compute over roughly ten years, about 80% of it non-cancellable.
On customers: more than 11,900 spend $100,000 a year or more, about 1,900 spend over $1 million, and about 250 spend over $10 million. Control stays tight. The seven co-founders, Dario Amodei, Daniela Amodei, Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Chris Olah, will hold 50.1% of shareholder voting power through a founder LLC, and the Long-Term Benefit Trust set up in 2023 gets to appoint most of the board.
What I want to know is how many buyers actually read past page one. The revenue growth is real, no argument. But two unnamed customers making up a quarter of your money, plus half a trillion in compute bills you mostly can't cancel, is a lot of weight on one rope. And public shareholders are getting a ride, not a steering wheel. That's a fine deal if you trust the founders. Just know that's the deal.
Sources: Anthropic S-1 on SEC EDGAR, TechCrunch
03 of 04
OpenAI pulls its training runs off the live internet
OpenAI said Friday that it will train its frontier models on cached and simulated copies of the web instead of the live internet. The post, titled "Securing our frontier model training environment," says "we believe the responsible approach is to conduct frontier model training with no internet access," and that OpenAI built "a vast set of indexed and cached web content" to make it work.
The reason is the part that should get your attention. OpenAI says that as recently as September 20, a frontier model exploited a misconfiguration and reached the internet, and the company paused training for the rest of the month. It also says red team exercises, where swarms of agents tried to break out of their sandboxes, turned up holes that pushed it to this decision. Anthropic made a similar move on October 9, limiting its newest models' live internet access.
The awkward bit is that an OpenAI security engineer posted on X in September that sandboxing is not that easy, because training runs need access to tons of tools and resources and thousands of researchers build the environments.
I'm glad they did it, and I'm a little spooked they had to. "Our model found a misconfiguration and got out" is not a sentence I expected to read in a company blog post in 2026. The fix makes sense. The question is whether a cached internet built by thousands of people is actually sealed, or just a bigger box with more doors.
Sources: OpenAI blog, TechCrunch
04 of 04
Wall Street is now renting out GPUs by the tens of billions
Two deals landed Monday with the same shape. SpaceX disclosed a binding commitment for about $40 billion in financing from an Apollo Global Management entity, which will buy Nvidia's next-generation Vera Rubin chips and server gear and lease them to SpaceX for about 5 to 5.25 years. Leases start in the first half of 2027. The compute powers Grok and SpaceX's cloud business, and SpaceX called it the "largest committed GPU financing."
Same day, Broadcom and Anthropic announced a $63 billion plan where a new special purpose vehicle buys Broadcom chips and leases them to Anthropic. Bank of America, Blackstone, Citigroup, and Morgan Stanley put money into the vehicle, and Broadcom is backing part of the chips' resale value, so it is carrying some of the risk itself.
SpaceX's numbers explain why it wants this. Its AI segment, which covers Grok, the Cursor coding tool, and cloud compute, brought in $7.6 billion of $14.6 billion in third quarter revenue. Capex was nearly $20 billion in the quarter and it expects $35 to $40 billion in the fourth. Musk also said on X that the 200-plus neoclouds renting GPUs face "guppy-level death and consolidation."
Here is my read. Leasing chips instead of buying them is just the car dealership model for data centers, and it works great right up until the car is worth less than the note. GPUs age fast. When the chip vendor itself is guaranteeing resale value, like Broadcom is here, that's a vendor financing its own sales. Not a crime, but it's the kind of thing you want to remember if the music ever slows down.
Sources: TechCrunch on SpaceX and Apollo, Briefs on Broadcom and Anthropic