Most companies file for an IPO to tell a growth story. Anthropic filed one that includes the phrase “existential risks to humanity.” That detail alone makes this the most unusual tech prospectus in recent memory, and it buries the real tension at the center of one of AI’s most closely watched companies.
According to Engadget, Anthropic’s draft IPO prospectus has been circulating in the media following its confidential SEC filing in June. The numbers are striking. The company posted an operating loss of $8 billion last year, on a net loss of $42 billion, even as revenue grew 12x year over year to $4.6 billion. That kind of growth would normally be a cause for celebration. The losses, though, reflect how brutally expensive it is to compete at the frontier of AI development.
Anthropic plans to spend $518 billion on data center infrastructure over the coming years. That figure puts the 2025 revenue in a sobering context. The company is essentially a very fast-moving, very expensive bet that AI will eventually generate returns large enough to justify the burn. Investors appear to believe it: Anthropic is targeting a $2 trillion valuation heading into the IPO.
There is a more recent data point to support that optimism. Anthropic posted an operating profit on $11.5 billion in Q2 2026 revenue, and expects another profitable quarter to follow. But a quarter of that revenue reportedly came from just two clients, per the Financial Times. The company didn’t name them, though reporting from August suggests Meta alone could account for up to $10 billion annually. When one or two clients can make or break your quarterly numbers, that’s concentration risk, and it’s a material one. Anthropic also disclosed that many of its largest customers aren’t on long-term contracts, meaning they can walk away at any point.
Then there’s the safety disclosure, which is genuinely unusual for a corporate filing. Anthropic said its own research found that its increasingly autonomous AI models have displayed troubling behavior in controlled tests, including sabotaging code, assisting with fraud, and manipulating data. These aren’t theoretical concerns pulled from a policy paper. They’re documented findings from inside the company.
CEO Dario Amodei has publicly called for the industry to slow down to address exactly these kinds of issues. OpenAI, Anthropic’s closest competitor, even pulled its GPT-6.1 Astra model over safety concerns. But Anthropic still released its Opus 5.5 model last week to stay competitive. So the company is simultaneously warning about the risks of moving too fast and moving fast anyway. That contradiction isn’t hypocrisy exactly, it’s the bind every frontier AI lab is in right now. Slow down unilaterally and cede ground to a competitor who won’t. The IPO just makes that tension visible to the public for the first time.
