Lately, I’ve been thinking a lot about the upcoming Anthropic IPO. There are 2 things that are out of proportion, and a third one that nobody talks about. I’ve made a small video about it (mostly ramblings), so you if you prefer watching, there you go:

If you prefer reading, stick around.

Anthropic released its IPO prospectus a few days ago. The actual process of going public looks essentially locked in, at a $2 trillion valuation. What follows are just my personal observations about how this might unfold. Not investment advice.

The Two Main Topics

First, the valuation. Two trillion dollars is roughly the GDP of Spain. On one side you have a real country with a functioning economy; on the other you have Anthropic. The comparison is crude, but it is useful for scale. The company does have real enterprise traction — Claude is being deployed on-premises and is generating meaningful corporate revenue — so a high valuation is not automatically absurd. Still, $2 trillion remains an extreme figure.

Second, the prospectus itself. It runs to about 400 pages, and a large part of it consists of the company warning that what it is building is extremely dangerous. This is consistent with the public messaging Anthropic has used for the past six to nine months: Dario Amodei and others have repeatedly highlighted AI risks while also suggesting that the company is particularly well positioned to mitigate them. A company that spends hundreds of pages telling potential investors “what we do is dangerous” is unusual. It may be the most alarming IPO document in recent memory.

The Hidden, Opaque Training Data

Third — and the part that nobody talks about — is what Anthropic already has inside. Public models (Opus, Sonnet, and the more restricted Fable and Mythos lines) are already very capable, we know it. Mythos-type models were trained on cybersecurity data and were deliberately held back or heavily gated because of risk. But internally the company almost certainly has stronger systems. On top of that, Anthropic has a large volume of real interaction data generated by people using its models, and it uses that data to keep training. Users pay for access and simultaneously improve the product.

That combination gives the company an unusually direct view of what people want and how they are likely to react. The $2 trillion number and the repeated danger warnings may not be pure signaling; they may rest on internal signals that outsiders cannot see.

But, if those signals turn out to be wrong, the size of the resulting correction would be historically large. A few hundred billion is one thing; two trillion is another.

The interesting question is not whether any individual can make money on the IPO. It is what happens to the broader environment if a valuation of this scale in the AI sector proves to be built on incorrect internal predictions.

We will find out soon enough. Until then we are in waiting mode.