OpenAI's $1.25T valuation, Anthropic's IPO risk warning, Democrats eye AI

The most telling detail in today's news is not the $1.25 trillion valuation. It is that the same week OpenAI closes roughly $11 billion of a round at that price, Anthropic is preparing an IPO filing that tells prospective investors advanced AI could bring "catastrophic or existential risks to humanity." Money is flooding in and the warnings are getting louder, and the two are now arriving together.

Start with Anthropic. Companies going public usually bury risk language in boilerplate nobody reads. Calling rogue AI "a global security threat of the first order" in a document meant to attract capital is a different kind of statement. I think it cuts two ways. It may be honest, since securities law punishes omissions and Anthropic has long argued that frontier AI is dangerous. It is also positioning: a company that says the technology is this powerful and this risky is implicitly telling investors it is among the few trustworthy enough to build it. I find that dual role worth watching, and I would not assume only one reading is true.

Meanwhile, Washington is changing. The Democratic midterm sweep sends five AI safety advocates to Congress, among them Scott Wiener, who wrote California's advanced-AI legislation as a state senator and now takes Pelosi's seat. By the way, the timing matters. Pritzker, Emanuel, Booker and Khanna have already been meeting safety advocates in San Francisco, which tells me 2028 hopefuls see AI policy as a live issue rather than a niche one. For anyone building on frontier models, the practical question is whether California-style obligations around testing and incident reporting go federal. Europe is trying to get ahead of that, with a proposal that it lead a global "common floor" for sharing information about AI failures. An international incident-reporting regime sounds dull, but aviation became safe largely through exactly that kind of boring, shared record-keeping.

The least discussed item is NVIDIA's reported GR00T 2, a robot foundation model that predicts future video and actions together. This comes from leaked slides of a Jim Fan talk, so I would hold it loosely until NVIDIA says more. If it holds up, joint prediction of what the world will look like and what the robot should do about it is a meaningful step beyond models that only map images to actions. It is the kind of capability that moves AI risk from text on a screen to machines acting in physical space, which is precisely where the incident-reporting argument gets harder to dismiss.

Put these together and the pattern is uncomfortable. Capital is pricing AI as if it will transform the economy, labs are warning in legal documents that it might go badly wrong, and legislators are arriving with the votes to act. Do those three forces pull against each other, or does the funding pressure quietly win? I suspect the answer depends on whether the first serious AI incident happens before the rules are written or after.