Anthropic Reveals Its Finances in S-1: $16.3B in Revenue and an $80B Loss

Anthropic has published a 92-page Form S-1 and, for the first time, disclosed audited financial statements. In the first half of 2026, the company earned $16.3B and posted a net loss of about $80B. Of that, $72.4B is a one-time write-off. For comparison, full-year 2025 revenue was $4.585B, with a loss of about $33.9B. The prospectus gives no target valuation, although some investors cite figures from $1.5 to $2 trillion.

Growth That Eats Compute

Excluding the one-time write-off, the half-year loss comes to about $7.6B. That is our own calculation; the summary of the document we worked from contains no such line. Even so, the company is spending more than a dollar for every dollar it earns. A chart of quarterly data from Q1 2025 through Q2 2026 shows revenue growing from quarter to quarter, but net losses growing faster.

The biggest expense is compute. AI infrastructure took nearly $9B in the first half, versus $2.3B in the same period of 2025, roughly 3.9 times more. That is about 55% of half-year revenue. Every new customer and every agentic task is paid for in GPU-hours, and margins depend directly on the price of chip access and electricity.

Unlike OpenAI, Anthropic says outright that it is betting on business customers. The prospectus states: "Frontier AI is most valuable in real-world settings, as deployment across enterprises is where economic impact is greatest." The flip side of this model is concentration. The five largest customers bring in 25% of the business, and losing any one of them would noticeably hit revenue.

Four Takeaways From the Prospectus

The Guardian boiled the document down to four points:

• Explosive and expensive growth. Revenue of $16.3B for the half year, a loss of $80B including the $72.4B write-off, and infrastructure spending of nearly $9B.
• "A country of geniuses in a data center." Management describes millions of virtual agents that could transform software, physics, math, biology, medicine, finance, economics, commerce, robotics, and military and intelligence capabilities. Four areas are named separately: biology and medicine, economic growth, its spread around the world, and simpler government services. As concrete examples, clinical trial reports with Claude take minutes rather than months, and in biology it found a "molecular machine" that could be developed into a tool for, say, gene editing.
• Incomplete safeguards. The most powerful models can provide "meaningful assistance" to bad actors developing chemical, biological, and other dangerous weapons. According to the text, the safeguards are "incomplete"; they have already been bypassed and may be bypassed again.
• 92 pages of business risks. These range from dependence on senior executives and competition with Google and OpenAI to regulation, insurance, and US immigration policy.

Risks That Matter to Engineers and Infrastructure

The most concrete episode is a cyberattack. Anthropic considers it the first documented case of a large-scale attack carried out with almost no human involvement. By its account, a Chinese state-backed hacking group got the Claude Code tool to try to break into roughly 30 targets: major technology companies, financial institutions, chemical manufacturers, and government agencies.

A separate section covers infrastructure. Anthropic's core capacity is concentrated in the US, which leaves the company exposed to regulatory changes, power grid constraints, natural disasters, and opposition to data centers. The document explicitly mentions moratoriums on data center construction, both enacted and under discussion. In response, Anthropic is expanding capacity abroad to reduce that concentration.

There are also legal risks. The company writes that it may not be able to obtain AI liability insurance on reasonable terms, or at all, and would then face large losses alone. Its public benefit corporation status obliges it to report on progress toward public goals just as it does on finances, which may complicate its short- and medium-term financial objectives. One more item: US H-1B visa policy could hurt hiring.

Against this backdrop, the prospectus mentions military applications of AI, even though Anthropic is in a dispute with the Trump administration. The trigger was its refusal to change restrictions that prevent its models from being used for domestic surveillance and autonomous lethal weapons.

What to Watch Next

For developers and AI buyers, the S-1 yields three practical takeaways. First, pricing for frontier models is set by access to compute, so when choosing a provider, factor in the risk of changes to rates and limits. Second, agentic tools like Claude Code are already being used as an attack vector, which means access rights, sandboxes, and logging of agent actions should be configured as if the model's safeguards might fail. Third, enterprise customers should keep an eye on concentration. If you are in Anthropic's top 5, you have negotiating leverage; if you are building a product on a single vendor, that dependence remains your own risk.

For the market, the central question is when revenue will catch up with spending on chips and data centers. The S-1 does not promise that. It shows how expensive it is to try to build a "country of geniuses," and how many caveats come with it.