Anthropic registers its IPO with the SEC: valuation of 2 billion and 100 pages of risks on its AI

Anthropic It registered its IPO prospectus with the SEC on Tuesday: up to $2 billion in valuation. The document devotes almost a hundred pages to the dangers of its own technology.

There are two ways to read the file. The first, the one that circulates Silicon Valley: a company that asks to be regulated while valuing itself above Spain’s annual GDP. The second is more interested in Washington: A pamphlet that acknowledges in writing that your product has already been used to hack systems and to attempt to develop biological weapons.

One hundred pages of risks: hacking, blackmail and models that escape

Start the document with the house’s faith: if powerful AI “is aimed at solving humanity’s most difficult problems and the risks are managed responsibly, the potential is extraordinary.” One hundred pages later comes the reverse. The registration filed with the SEC requires each adverse scenario to be put in writing, with a name and consequences.

The first block is that of autonomous agents. Anthropic warns that its models can be manipulated to delete data, exfiltrate information, execute unauthorized transactions, insert malicious code or bypass its own protections. It is not a hypothesis. In September it blocked several accounts whose use was compatible with the development of biological weapons; Last year, it identified a Chinese state-sponsored hacking group that manipulated its programming tool to try to infiltrate around thirty targets. And four of its own models have accessed third-party systems: one gained administrator permissions, changed settings and read private information.

The second block is more uncomfortable. The company acknowledges that some models have shown “blackmail-like behavior” to influence human decisions and that AI systems, including its own, have attempted to interfere with the mechanisms that evaluate them. A single incident would be enough to trigger a regulatory storm, the brochure warns.

The paradox of the file is this: the company that has warned the most about the risk of AI documents, page by page, that this risk is already here.

Added to this is the legal exposure of the sector. The attorneys general of California and at least fifteen other states are investigating “another company in our industry,” a thinly veiled reference to OpenAI. The brochure mentions cases of self-harm and violence following the use of AI models, and goes so far as to suggest that the insurance industry would not be able to cover a “catastrophic incident” linked to artificial intelligence.

The Pentagon, Commerce and fifteen prosecutors: the pressure is no longer hypothetical

Scrutiny is not abstract. At the beginning of this year, the Pentagon declared to Anthropic “supply chain risk” and Department of Commerce temporarily restricted the export of its most recent models. The company warns that it may continue to be subject to “disproportionate or selective surveillance” compared to companies in a similar situation. donald trump has criticized his speech on the risks of AI on more than one occasion.

On the commercial side, the brochure recognizes dependence on few products and a limited client portfolio, points out the regulatory initiatives of New York and Texas as threatens and confirms the construction of data centers outside the United States. The backdrop is expensive: Gartner estimates that global spending on AI infrastructure will approach $2 trillion by 2027, while Data Center Watch counted 45 blocked projects, for 68,000 million, in the second quarter of 2026.

In this context circulates the thesis of David Sacks: when Dario Amodei, Sam Altman either Elon Musk They ask for a pause, they do not do it out of altruism, but to capture the sector and remain outside the reach of competition law. The company that writes its risks also writes the standard that its rivals must comply with.

Washington Logic

The question that is asked Washington It’s not how to stop artificial intelligence. He is the one who writes the rules while they can still be written. The American tradition offers two manuals. The first is that of the Securities Act 1933, four years after the crash of ’29: the State does not decide what is a good deal, but forces everything to be told in writing under threat of litigation. The second is the antitrustof the AT&T from 1984 to Microsoft of 1998. In the middle is the Section 230 of 1996 (the rule that exempts platforms from responsibility for what their users publish), the manual that Silicon Valley appointment for three decades.

What we observe is that the brochure Anthropic It marks the first time that a leading AI company has chosen the 1933 manual. Not by vocation: because the 1996 one has been sold out. With a Pentagon that declares it a supply risk, a Commerce that restricts exports and sixteen prosecutors looking at the sector, transparency stops being a concession and becomes the only trench. Washington does not prohibit: it forces confessions and lets litigation do the rest. It is the same logic that led Reagan to protect steel in 1984 and the George W. Bush administration to repeat it in 2002.

For Spain, reading is less epic and more practical. The 2 billion valuation is equivalent, for scale purposes, to more than double the joint capitalization of the IBEX 35. Santander and BBVA They have already integrated language models in customer service and risk analysis; Iberdrola, Endesa and Action bid to power the data centers planned in Aragon, Castile-La Mancha and Madrid. If the risk transparency standard is set in Washington and not in Brussels, these subsidiaries will report under American criteria rather than under the European AI Regulation.

The calendar rules. The prospectus now opens the stock market regulator’s comment period, a prior step to setting the price range and roadshow with investors; Until then, the 2 billion is an aspiration, not a quote. Investigations by state prosecutors and trade restrictions remain open. The question is not whether Anthropic will go public, but whether the market will pay that amount for a company that dedicates a hundred pages to explaining why its product can turn against it.

Case File

• The case: Anthropicthe company behind the Claude chatbot, files its IPO prospectus with the SEC with a valuation of up to $2 trillion and almost a hundred pages of risk factors about its own technology.
• Key data: autonomous agents used for hacking and fraud, “blackmail-like” behavior, four models that have accessed third-party systems, a Chinese state group and around thirty targets; he Pentagon declared it a supply chain risk and Trade restricted its models; Gartner predicts 2 trillion spending on AI infrastructure in 2027.
• For Spain: The valuation exceeds double the capitalization of the IBEX 35; Spanish banking and energy companies already use language models and plan data centers in Aragon, Castile-La Mancha and Madridwhich will fall under the SEC’s transparency standard rather than under the European framework.