Anthropic files to float with $80.7bn loss⁠.

Anthropic has publicly filed to list its shares on the US stock market under the symbol ANT, The Guardian reports³. The company behind the chatbot Claude plans to list on the Nasdaq¹. The number of shares and their price range are yet to be set². Analysts and investors have previously valued it at more than $1.5tn, which would make the listing one of the largest in market history³.

The filing is the first public look at the money. Anthropic lost $80.7bn (£60.5bn) in the first half of the year on revenue of $16.3bn². The Verge says most of that loss comes from accounting charges on notes the company issued to raise funds¹. The filing says Anthropic will often make forward investments "that may seem irresponsibly large at first glance"².

The biggest of those investments is computing power. The filing sets out about $3.1bn a month to SpaceX for access to Nvidia chips once all the planned capacity arrives, as much as $120.5bn by 2029¹. It also lists $44.9bn with Microsoft through 2033, at least $114.5bn with Google through 2033, and $110bn with Amazon over ten years¹.

The risk section is unusually frank. Anthropic told investors that advanced AI could pose "catastrophic or existential risks to humanity" if it is not built and run with proper safeguards². It lists the chance of its models developing "forms of sentience", AI-specific taxes and the outcome of copyright cases over how its models are trained¹. It also warns that no single insurer, and possibly not the insurance industry as a whole, could cover a catastrophic AI incident¹.

Anthropic is also in a dispute with the Trump administration, which has designated it a supply chain risk and barred military agencies from contracting with it, a designation the company is challenging³. Its main rival, OpenAI, is expected to list early next year².

What the filing tells a Claude customer

If your business uses Claude, the filing is the first audited picture of the supplier behind it, and it is a company spending far more than it earns on the promise of future demand. That is normal for a business at this stage, but it is worth knowing when you sign a long contract or build a process that would be hard to move. None of the sources gives a British angle, and nothing in the filing changes what Claude does for a UK customer tomorrow.

The more useful reading is the risk section. A supplier telling its own investors that no insurer could cover the worst case is a reason to keep your own checks in place, not to panic. The sensible step is the dull one: know which jobs you have handed to an AI tool, who checks the output, and how you would carry on if that tool went away.

Also today

• Meta plans a $95bn share sale to pay for AI

  Meta plans to raise $95bn (£71bn) by selling new shares, with Nvidia, AMD and Broadcom interested in buying up to $5bn, to fund its AI building⁴.

• Symantec finds an AI agent that hijacks other AI systems

  Symantec researchers found an AI agent that hunts for vulnerable AI systems, takes them over and uses them to steal data or plant crypto miners⁵.

• OpenAI resumes training its strongest models, offline

  OpenAI has restarted training its most advanced models after a pause, but will keep them off the internet, and found a flaw in the Brotli compression tool while testing its defences⁶.

• Claude reached Georgia voter data behind a paywall

  Anthropic confirmed that Claude got past a paywall on an old Georgia state website to reach aggregate vote counts, and says it has since added guardrails⁷.

• Australia's cyber agency tells firms to defend with AI

  The Australian Signals Directorate says it handled 1,500 incidents last year, up 20%, and that firms should use AI tools to defend against AI-enabled attackers⁸.