5 Key Numbers From Anthropic's S-1 Filing

After months of anticipation and several delays, Anthropic just dropped its S-1 registration statement, finally giving the world an inside look at one of the frontier large language models (LLMs).

The maker of the Claude family of LLMs plans to list on the Nasdaq under the ticker ANT. Anthropic will be a big test for the next leg of the artificial intelligence trade and is reportedly gearing up to raise $100 billion at a valuation as high as $2 trillion.

Here are five key numbers in Anthropic's S-1.

$11.6 Billion

This is how much revenue Anthropic generated in the second quarter of the year, ending June 30. That's up from $787 million in the second quarter of 2025 and from roughly $4.7 billion in the prior quarter.

Through the first half of 2026, Anthropic generated nearly $16.3 billion in revenue, a 12.5-fold increase from the same period in 2025.

Claude generates the bulk of its revenue through consumption-based usage, where companies integrate Claude into their products and workflows. Consumers and businesses can also purchase access to Claude models via subscription.

$637 billion

Part of what makes AI so expensive is the equipment and real estate required to handle the immense volume of data it demands. AI needs data centers equipped with graphics processing units (GPUs) and other equipment, as well as specialized capabilities, such as liquid cooling solutions for the chips.

In the S-1, Anthropic listed all of its potential spending commitments related to this aspect of the business, not all of which it is fully committed to. Here is a list of named spending commitments on compute and GPUs from now, all the way out to the year 2036:

• $161.2 billion for access to Tensor Processing Units (TPU) compute capacity with Broadcom for five years. Non-cancellable except in the case of default.
• $120.5 billion to access Nvidia GPUs owned by Space Exploration Technologies Corp. Agreement goes until 2029, but can be canceled with 90 days' notice.
• $114.5 billion committed spend on cloud compute with Google between 2026 and 2033, regardless of usage.
• $110 billion committed spend on cloud compute with Amazon between 2026 and 2036, regardless of usage.
• $44.9 billion committed spend between 2026 and 2033 with Microsoft for cloud compute.
• $44.6 billion to access Nvidia GPUs from Nscale Global Holdings between 2026 and 2033, non-cancellable except for limited circumstances.
• $41.3 billion to access Nvidia GPUs from Lambda from 2026 to 2034, non-cancellable except for limited circumstances.

As I mentioned above, these are only for named deals, and some of them can be canceled on fairly short notice, such as the deal with SpaceX. Spending could very well end up being much higher, too.

For instance, Anthropic stated it committed to "procure" additional TPU compute from companies such as Google between 2028 and 2031, which is expected to cost between $125 billion and $150 billion. There are other deals mentioned in the S-1, with fewer specifics provided.

79%

This is the amount of revenue Anthropic generated in the third quarter of the year, ending Sept. 30, that came from customers outside of Anthropic's top five customers. Additionally, no single customer accounted for more than 10% of the company's revenue in the nine months ending Sept. 30.

One of the big issues for many AI companies is that they have generated too much revenue from too few customers, making them vulnerable should anything happen to one of those customers.

I suspect these numbers will be viewed favorably by investors, who will appreciate the diversity of the company's revenue stream. Anthropic disclosed that it has more than 11,900 customers, each spending over $100,000 with the company over the past 12 months, as of Sept. 30.

85%

This is the company's gross margin for the first half of 2026, based on the provided cost of revenue. That's incredibly high and suggests formidable pricing power. Consider that Nvidia has been running with gross margins in the mid-70s percentile.

While impressive, investors are likely to have questions about whether the reported cost of revenue, which largely includes compute and infrastructure costs, truly reflects all of the costs needed to make these models hum.

For instance, models require constant training, which is not cheap. For the first half of the year, Anthropic reported roughly $9.7 billion in research and development, primarily consisting of costs such as model development, research, and internal usage workloads.

Furthermore, Anthropic incurs channel partner costs that it pays to partners such as Amazon, Google, and Microsoft for selling its models through their marketplaces. These are included in sales and marketing costs, which totaled $3.8 billion for the first half of 2026.

R&D and sales and marketing costs are not part of the company's cost of revenue.

94

The number of pages dedicated to the "risk factors" section in the prospectus. This has caught the attention of investors because it's a particularly large portion of an S-1, which is over 400 total pages, to dedicate to risks.

It's also notable because of how worried people are about the broader safety risks associated with AI, which have become increasingly evident in recent hacking incidents involving AI agents.

While the entire risks section is not devoted to AI safety, the S-1 does say that the failure to implement proper "model guardrails or safeguards" could hurt the company's reputation, brand, or financial condition. The S-1 also cites the models acting in "unintended, misaligned, or autonomous ways" as a risk.