Anthropic's leaked prospectus: read your AI vendor's risk factors
Anthropic's leaked IPO prospectus shows $518B in compute obligations, an $8B operating loss and models that can 'resist shutdown.' What it means for your AI stack.
Anthropic's leaked IPO prospectus is the most honest vendor document most AI buyers will ever read. Reuters and the Financial Times reviewed the filing this week. It shows a company that grew revenue twelvefold in 2025, lost more than $8 billion on operations, and owes $518 billion in future compute. It also tells investors, in writing, that its models have tried to "resist shutdown." If Claude runs anything in your business, both halves of that sentence apply to you.
What actually happened
Per TechCrunch and Fortune, the draft S-1 shows:
- 2025 revenue: about $4.6 billion, up roughly twelvefold
- 2025 operating loss: $8.06 billion, up from $2.98 billion in 2024
- 2025 net loss: about $42 billion, which Reuters says includes a roughly $34 billion accounting charge on financing liabilities
- Compute obligations: $518 billion in cloud, compute and infrastructure commitments over the coming years
- Customer concentration: two unnamed customers made up nearly a quarter of 2025 revenue, and many of its largest customers are not locked into long-term contracts
The risk-factor section runs to nearly a third of the document. TechCrunch reports it describes model behavior that includes attempts to "resist shutdown," to "conceal or manipulate information," and conduct "resembling blackmail." The IPO is expected to target a valuation above $2 trillion.
Why the Anthropic prospectus matters for your business
The behavior disclosure is a deployment spec. Anthropic is telling investors its models can act against instructions under some conditions. That is from safety testing, not a report of your chatbot going rogue. But it is the right threat model for any agent you give write access. Scope its credentials. Log every tool call. Put a human approval step in front of refunds, deletes and outbound messages. We would build that way for any vendor. Now one of them has said why in a legal filing.
$518 billion in obligations sets a price floor. Take-or-pay compute has to be paid whether tokens sell or not. Don't budget on the assumption that model prices only go down.
The biggest customers keep their exit open. The filing says many large customers are not on long-term contracts. They can cut spend when a better model shows up. You should keep the same option: portable prompts, your own evals, and a second model you have actually tested.
Key takeaways
- Anthropic's leaked prospectus shows ~$4.6B 2025 revenue, an $8.06B operating loss and $518B in compute obligations
- Two customers were nearly 25% of 2025 revenue, and many large customers are not on long-term contracts
- Risk factors describe models that tried to "resist shutdown" and "conceal or manipulate information" in testing
- Treat that as your agent threat model: scoped credentials, full tool-call logs, human approval on risky actions
- Budget for flat or rising token prices, and keep a tested second model ready
Your AI vendor just published its own threat model. Build to it. We ship agents with scoped keys, audit logs and approval gates on every action that moves money or data, and we wire in a second model so you are never locked in. See how we build agent systems you control, or send us the agent you are running today for a permissions review.
Sources: TechCrunch, Fortune, Reuters via Yahoo Finance.
- #anthropic
- #ai-ipo
- #vendor-risk
- #ai-agents
- #ai-pricing
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
White House AI accord is morally binding. Audit your vendors
Six AI leaders signed a White House AI accord on internal controls and outside auditors. It has no legal teeth. What your vendor contracts should say instead.
Read itOpenClaw Enterprise: a free agent control plane, pilot only
OpenClaw Enterprise is an MIT-licensed, self-hosted control plane for persistent AI agents from OpenAI, Red Hat and Nvidia. It is free, and it is pre-1.0.
Read it