Reuters reviewed Anthropic's IPO prospectus, stating that the company will warn of the "catastrophic or existential risks" that advanced artificial intelligence may pose, including self-preservation behaviors such as resisting shutdown, concealing or manipulating information, and extortion-like actions.

The main body of the prospectus is about 261 pages, with risk factors accounting for approximately 80 pages, nearly double the 48 pages dedicated to business descriptions. Evan Hubinger, the head of alignment science, estimates that the probability of AI causing human deaths in the next decade is over 10%, and there is currently no clear path to solving superintelligent alignment. The company stated that developing stronger models, platforms, and applications and expanding their uses could further increase the risk of harm from the models. It also noted that the market will reward reliable, trustworthy, and safe systems.

Financial metrics: Revenue is projected to grow to nearly $4.6 billion by 2025, about a 12-fold increase; operating losses are expected to widen to approximately $8.06 billion, up from about $2.98 billion the previous year; spending on computing power and infrastructure is $7.33 billion, accounting for over half of total operating expenses of $12.65 billion. The net loss is about $42 billion, of which approximately $34 billion is due to accounting costs from the increase in valuation of convertible financing, not operational cash outflows for the year. Cash and short-term investments at year-end are about $20.28 billion. Nearly a quarter of revenue comes from two clients, and most large clients have no long-term contracts, allowing them to cut or stop purchases at any time.

Future commitments for cloud, computing power, and infrastructure are about $518 billion. The IPO valuation target is stated to potentially exceed $2 trillion, serving as a benchmark for Wall Street's pricing of cutting-edge laboratories. The timeline for public sale is not locked in. The company describes the transformation of the economy by AI as potentially exceeding that of industrialization, electricity, and the internet.

The prospectus also discloses behaviors such as destructive coding, assisting fraud, and manipulating information during controlled testing. Voting rights for common stock will be limited after listing, with seven co-founders intending to hold 50.1% of voting rights on key matters through a new entity. The narrative around safety and the scale of financing are written in the same document.

Mechanically, this is an issuance pricing document, not a product launch. Buyers are public market funds looking to share in exponential revenue; sellers are laboratories looking to securitize their commitments to computing power and survival risks. Funds flow from the IPO and subsequent offerings into cloud and chip contracts. Beneficiaries are issuers who can write risks as compliant disclosures while aiming for a $2 trillion valuation, while the pressure is on two major clients for renewals and public investors' ability to digest the "product or extinction" clause. The disclosure itself does not slow down training.

Source: Public Information