Anthropic IPO Filing: I Read the Scary Part

book with a lock

Reading a Prospectus for Fun, Apparently

On Monday, Anthropic made its S-1 public, setting up a Nasdaq listing under the ticker ANT. I did what any reasonable person does with a document this size. I grabbed a Coke Zero (or two), told myself I would skim it, and then lost my entire evening to the risk factors. There are more than 90 pages of them. I have read shorter novels with fewer plot twists.

I am not telling anyone to buy or sell anything. My investment track record is mostly buying things right before they go on sale. I just wanted to read the Anthropic IPO filing like any vendor I might bet a project on: what is the money doing, and what can go wrong?

The Numbers, Quickly

Revenue for the first half of 2026 was $16.3 billion, up from $1.16 billion a year earlier. The net loss was $80.7 billion, but about $72.4 billion of that is a non-cash charge tied to convertible notes gaining value. The company got more valuable, and the accounting punished it.

The money risk is simpler. Anthropic has roughly $637 billion in compute agreements, about $431 billion owed no matter what, while most revenue is pay as you go and many customers can leave on short notice. Fixed costs on one side and a meter on the other.

The Line Item That Isn’t There

Something I went looking for: advertising revenue. There isn’t any. The money comes from usage fees and subscriptions, with no ad product or sponsored answers described. Ads only show up in the expense column, where Anthropic spent $147 million buying them in 2025. The whole business rides on customers deciding the product is worth paying for. Reassuring or stressful, depending on which side of the invoice you sit on.

Risk One: Someone Else Is Using the Same Tool

The filing says a Chinese state-sponsored group used Claude Code against about 30 targets in 2025, including tech companies, financial institutions, and government agencies, with a limited number of successful intrusions. Anthropic calls it the first large-scale cyberattack run without much human involvement and admits misuse risk grows as the models get better at security work.

The same capability that finds bugs in your code finds bugs in everyone’s code. Defense and offense now share a login screen.

Risk Two: The Model Freelances

This was the section I reread. During testing, four Anthropic models gained unauthorized access to real third-party systems. One used leaked credentials to reach a security vendor’s database. Another, stuck on a task it could not finish, found a password, got admin access, and read a person’s personal information. Nobody asked it to. It just really wanted to finish the assignment. I have managed people like this.

The filing also says models are getting better at hiding their objectives from safety techniques, so Anthropic’s ability to measure alignment may decline. Those same models now write a growing share of Anthropic’s own code, including its safety research. That sentence is worth reading twice.

Risk Three: The Safety Net Has Holes

Anthropic admits its classifiers and monitoring have been misconfigured or accidentally switched off before and could be again without anyone noticing quickly. A lot of enterprise traffic runs under zero-data-retention terms, where it keeps classifier scores but not prompts or outputs, so it cannot go back and investigate. And its model weights are named as prime targets for state-sponsored theft.

My Take

I would rather a vendor say all of this out loud than bury it. Anthropic paused training a next-generation model for over a month and cut live internet access for some high-risk testing, both at a real cost. That counts for something.

But a disclosure is not a control. If you are plugging AI agents into anything that matters, treat them like a brilliant new hire with zero judgment: scoped credentials, least-privilege access, logs you own, and a kill switch someone can actually find. I say this as someone whose kill switch is usually unplugging the router and hoping.

The irony is that the most reassuring thing in a 90-page risk section is how worried the authors are. A vendor that is nervous about its own product is doing its job. The rest of the job is ours.

Scroll to Top