Anthropic files S-1 for public listing at $1.5 trillion to $2 trillion valuation

The path to a public listing

Anthropic recently filed its S-1 document, which is the formal registration required to go public on the Nasdaq. This filing provides the first detailed look at the finances of a leading frontier AI lab. It suggests the company could begin trading as soon as next month at a valuation of $1.5 trillion to $2 trillion.

The document reveals that Anthropic has efficiently scaled its operations in a very short amount of time. It reports over $70 billion in annualized recurring revenue (ARR). This growth has occurred despite the fact that the broader enterprise market has not yet fully adopted these tools at a massive scale.

Analysts note that the company's current spending is heavily concentrated on coding-specific tasks. While this has fueled rapid growth, the company faces the challenge of expanding into more diverse verticals. To reach extreme valuations, it must move beyond niche tasks into broader white-collar work.

Improving unit economics

The S-1 highlights a significant shift in how Anthropic manages its costs. Two years ago, the company reported negative 94 percent gross margins on inference, which is the process of running a model to generate a response. Today, those margins have turned positive to 85 percent.

This improvement is driven by better compute efficiencies and a diverse chip fleet. Anthropic uses a variety of hardware, including Trainiums, TPUs, and GPUs, to achieve higher token throughput. Token throughput refers to the number of tokens a model can produce per second, which has seen gains of 5x to 10x year-over-year on certain chips.

The company is also seeing significant gains in revenue per megawatt. Currently, Anthropic generates roughly $65 million of revenue per inference megawatt. Experts suggest that as they move to newer hardware like Nvidia's Blackwell, that figure could reach between $100 million and $120 million.

Research vs. revenue spending

A key takeaway from the filing is the distinction between revenue-generating compute and research spending. Only about 30 percent of Anthropic's compute is actually used for inference-serving revenue. The rest of the compute is dedicated to experimentation, training new models, and post-training research.

This dynamic is central to Anthropic's goal of reaching AGI, or artificial general intelligence. The company intends to reinvest the profits from its high-margin inference business back into the R&D side. This allows them to build the next generation of models while maintaining a viable business model.

The filing shows that Anthropic is managing its training costs against its inference margins quite well. However, the company has messaged to investors that it may want to increase the percentage of compute allocated to training. They are willing to sacrifice some margin leverage to continue their pursuit of more capable models.

Market competition and growth

The AI landscape is becoming increasingly competitive as more players enter the space. While Anthropic led the industry in net new ARR in the first half of the year, OpenAI took the majority of that growth in the third quarter. Additionally, open-weight models have seen a surge in inference costs, growing from $1 billion in the first quarter to over $10 billion in the third.

Enterprises are also showing a preference for diverse options, including on-premise AI. Many companies are nervous about sharing all their data with a single provider like Anthropic or OpenAI. This has led to a rise in law firms and other organizations running models on their own internal hardware.

To reach a $10 trillion valuation, Anthropic will need to move into sectors like biology or cybersecurity. These areas offer eight-figure contracts that can scale quickly. The company's success will depend on its ability to transition from AI augmentation to full labor replacement in these high-value white-collar industries.