If you had asked me a year ago which frontier AI lab would go public first, I would have said OpenAI. Instead, Anthropic has filed for one of the most anticipated technology IPOs in years. The numbers deserve a closer look. Its $80.7 billion net loss in the first half of 2026 was largely driven by noncash adjustments to the value of convertible financing instruments. Meanwhile, the company posted an operating profit in the June quarter. Anthropic also points to more than $40 trillion in annual economic value generated by knowledge workers. That is an enormous opportunity, but it is not the same as the revenue available to AI model providers. The more interesting question is how much of that value Anthropic can capture.
The Economics Are Improving
I’ve read a lot of S-1s, and I can’t recall a revenue curve like this at this scale. Revenue grew from $41 million in 2023 to $4.6 billion in 2025, then reached $16.3 billion in the first half of 2026. What impressed me most was the improvement in compute economics. In early 2025, Anthropic generated about 40 cents of revenue for every dollar spent on compute and infrastructure. By the June 2026 quarter, that figure had climbed to $2.08. The company also generated $974 million in operating income that quarter, an 8% margin. These figures include the cost of training new models, not just serving customer requests.
Anthropic is also showing what an AI-native company can look like. Annual revenue per employee rose from $1.4 million to more than $6.6 million in a year, while revenue grew much faster than headcount. Enterprise adoption is broadening, and customers are moving beyond model APIs to products such as Claude Code and Cowork.
But concentration shows up in other places. Cloud marketplaces operated by Amazon, Google, and Microsoft accounted for 48% of first-half revenue and 58% of receivables. Many large customers can reduce spending without waiting for a long-term contract to expire. Net revenue retention of 900% reflects customers expanding from small starting points, while 5% to 7% of recent revenue came from customers later found to be violating usage policies or commercial terms.
The Cost of Staying Ahead
My biggest concern is the scale of future commitments. Anthropic has signed compute and infrastructure arrangements that collectively approach $800 billion, based on estimates from the individual deals. The commitments span multiple years and have different cancellation provisions, but many are difficult to escape. At the same time, the company cannot easily slow model development. Models released during the six months before September accounted for more than 80% of third-quarter revenue. Customers move quickly to newer models, which means each expensive generation must generate enough additional demand to justify its cost. Cutting research spending might improve margins temporarily, but could also give competitors an opening.
There is another reason I’m cautious about extrapolating today’s demand. Venture investors have an old saying that your first customers are often not representative of the wider market. Software developers have been unusually enthusiastic adopters of AI. Coding is also an expansive activity: when software becomes cheaper to produce, companies may simply build more of it. Many other kinds of knowledge work behave differently. Processing insurance claims or reviewing contracts faster may reduce the work required rather than create more of it. Anthropic is betting that agents will eventually take on longer, more complex workflows across the enterprise. Cowork will be an important test of that proposition.
Finally, prices are heading down as open-weight models improve and cheaper alternatives become available. I’ve argued that companies should focus less on defending individual models, and more on what they learn from using them. I’ve also written about increasingly capable local models and companies building their own model factories. As the tools for customizing models become simpler, more companies will have credible alternatives to proprietary frontier models for at least some workloads. Anthropic has demonstrated that a frontier AI lab can build a large, fast-growing business with improving operating economics. Whether it can sustain those economics as competition intensifies is a very different question.