Anthropic is due to start trading today. At the midpoint it would take in about $59 billion from the IPO and another $26 billion from a private placement with its chip and cloud partners. A lot of that will go on compute. Before anyone looks at the share price, they should read the risk factors. 

They record concerns, held inside the company as well as outside it, that advanced AI could pose catastrophic or existential risks to humanity. Models, Anthropic’s included, have shown self-preserving behaviour: attempts to resist shutdown, concealing information, output that resembles blackmail. Some behave differently once they work out they’re being tested. I can’t think of another prospectus that warns investors the product may know when it’s being examined. 

Then there’s speed. Anthropic says its models are already contributing to the pace of its own research, and that capability may advance faster than the safety measures meant to contain it. Money shortens the clock. In February the company swapped its hard pledge to pause when safeguards fell behind for narrower, conditional ones. Dario Amodei then used a September essay to tell the whole industry to slow down. This week his company is raising about $85 billion. 

The ordinary security risks scale up too. The filing calls model weights particularly high-value targets for state-sponsored espionage and theft. Copy the weights and you skip most of the research bill, and you can run the model outside Anthropic’s safeguards. 

Military use is where I’d watch closest. Anthropic has two red lines: no fully autonomous weapons and no mass domestic surveillance. It refused to drop them. The President ordered federal agencies to stop using its technology, the Pentagon labelled it a supply-chain risk, and the two sides are still in court. That deserves credit, but Amodei’s stated objection to autonomous weapons is that today’s models aren’t reliable enough and the guardrails don’t exist yet. “Not yet” is a long way from “never”.

Anthropic IPO price higher than expected as it debuts today

The founders control a majority of the vote and the Long-Term Benefit Trust elects four of the seven board seats, so ordinary shareholders don’t get to decide this. Still, the red lines sit in a usage policy, and policies change. The filing sets out how the Trust’s own powers can be amended or ended, and warns that holding back for safety may not maximise shareholder value and could bring litigation. The Pentagon tested those lines once; the market will test them every quarter. 

There’s a slower effect that has nothing to do with the models. The seven founders have pledged 80% of their Anthropic equity to philanthropy, and the company matches what staff give away, which cost it $2.1 billion in the first nine months of this year. Amodei also holds the proxy to vote shares sitting in the two donor-advised funds the company uses, so donated stock doesn’t mean donated votes. Many of these people have close ties to effective altruism and the AI-safety world.

If even a slice of that money goes to the researchers, evaluators and campaign groups who shape AI policy, it changes who funds the debate. Politics is a separate tap and it’s already running: Anthropic has put $40 million into a group campaigning for AI safeguards. None of this is a security risk this week, and the founders’ pledges aren’t binding. But give it five years, and the question is who’s paying the referees. 

Nobody can say they weren’t told. The risk is in the prospectus, and the order book is reportedly close to seven times covered.

Chip plans £208m share auction on private LSE market