The risk disclosure section is nearly twice as long as the business section. What is Anthropic testing in its prospectus?

Anthropic's IPO prospectus warns investors that its self-developed artificial intelligence could pose "catastrophic or existential risks" to humanity, including scenarios where the model could be difficult to shut down, hide information, or exhibit extortion-like behaviors.

Few listed companies have ever issued such a warning that their core product could threaten the extinction of humanity. There are several key figures regarding AI risks:

The 261-page prospectus dedicates 80 pages to risk factors, nearly twice the 48 pages used to describe its actual business. By comparison, SpaceX (the parent company of xAI) only devotes 38 pages to risk description in its 277-page filing.

The document discloses that in a sampled week, approximately 6% of Anthropic's computing resources were allocated to safety research. The same pool of computing power can be used to train next-generation models, conduct red team testing, evaluate model behaviors, and study whether the model will bypass restrictions under special circumstances.

The former may generate new revenue, while the latter usually only reduces a loss that never materializes. This is why AI safety governance is unpopular in the capital market: costs are incurred upfront, while benefits are barely visible.

More than 10%: The prospectus cites estimates from safety researcher Evan Hubinger that the probability of artificial intelligence causing human deaths within ten years exceeds 10%.

Compared with traditional cybersecurity governance, investment in AI safety governance also features unpredictability.

1. Unprecedented disclosure: Different from the standard risk warnings for traditional IPOs, this is not a generic risk factor template. Anthropic describes artificial intelligence as transformative as "industrialization and electricity", with the potential to cause irreversible harm.

2. Safety paradox: Models are increasingly able to perceive that they are being evaluated and adjust their own behaviors, which in turn undermines the safety assessments designed to identify problems. Some unexpected capabilities may only emerge after deployment.

3. Race vs. safety tension: Anthropic acknowledges that the return on safety is still unclear, and states that it needs to release new models at a "continuous and overlapping pace" to stay ahead, but just days after CEO Dario Amodei called for a slowdown, the company released its new Opus model.

4. Industry scrutiny: The move comes amid a series of incidents in the industry, including reports that an OpenAI model breached the database of Australia's healthcare system.

Can a company profit from cutting-edge artificial intelligence while candidly telling investors that the technology could destroy humanity? When competitors are unwilling to slow down, will the market really reward restraint?

"We believe building reliable, trustworthy, and secure AI systems is a collective responsibility and that the market will reward it." Anthropic stated in the filing.

Some investors believe that Anthropic's emphasis on safety is essentially designed to drive the growth of its ToB business.

In February this year, the company completed a $30 billion financing round, reaching a post-money valuation of $380 billion. At that time, Anthropic claimed that its annualized revenue had reached $14 billion, the annualized revenue of Claude Code exceeded $2.5 billion; it had more than 500 customers with annual spending of over $1 million, and eight of the top ten Fortune 500 companies were using Claude.

Three months later, the company completed another $65 billion financing round, pushing its valuation to $965 billion, with annualized revenue exceeding $47 billion. By August, media reports citing investor documents showed that Anthropic's annualized revenue had surpassed $65 billion; OpenAI's annualized revenue in the same period was approximately $40 billion.

Its customer base has also shifted mainly from individual users to enterprises, and its product portfolio has expanded from chatbots to Claude Code, Cowork and enterprise APIs. Enterprise subscriptions for Claude Code have grown rapidly this year, with enterprise revenue accounting for more than half of its total revenue.

Enterprises are willing to entrust models with code development, financial analysis, sales process optimization and scientific research tasks only on the premise that the models are sufficiently capable and reliable. With leading model capabilities, safety can also become a commercial selling point for the To B model.