Anthropic's S-1: existing customers spend ~9x more than a year ago, and why the first AI pilot is just the beginning

By Nate (Nate's Substack) · October 28, 2026.

The analysis starts from the initial S-1 that Anthropic filed on October 26. According to the author, more than half of Fortune 500 companies are already Anthropic customers generating at least $100,000 in annual revenue. Yet among the publicly traded companies in that group, spending with Anthropic still averages less than one dollar per thousand dollars of total spending. Anthropic reports $16.3 billion in revenue for the first half of 2026: a huge business that is nonetheless still a very small purchase within the world's largest companies.

The author's thesis is that Anthropic could become one of the most successful businesses in capitalism by selling intelligence that makes the customer's existing people and capital more productive, and that also makes viable new projects that are currently too expensive, too difficult or never made it into a budget. Customers learn to use this intelligence by using it: they discover new uses, build more capable systems and revisit projects they had shelved. Anthropic's revenue grows with that activity, and so do its research and compute costs.

The first order says little about what comes next. Today's AI budgets reflect what companies have learned to do so far; as employees gain experience and models improve, more work becomes practical. A hundred thousand dollars is a real relationship for Anthropic, but for a Fortune 500 company it may still be a small deployment in a narrow part of the business.

S-1 figures cited in the email: as of September 30, more than 11,900 customers exceeded $100,000 in trailing twelve-month revenue, up from more than 950 a year earlier. More than 1,900 had reached $1 million and about 250 had reached $10 million. Fortune 500 customers include the ten largest banks and nine of the ten largest pharmaceutical companies.

There is already a wide gap in purchasing intensity. Anthropic's ten largest Fortune 500 customers spend on average more than eight times the average share of the qualifying group, although even the average for that top group remains below 1% of total spending, according to the S-1's comparison with publicly traded customers. For the author, the opportunity is for intelligence to help the customer's existing spending achieve more, so that greater spending on intelligence can pay for itself by improving the use of resources the company already has; Anthropic only needs to capture part of that additional value.

Anthropic reports net revenue retention of roughly 900%: existing customers collectively generated about nine times more revenue in the September quarter than a year earlier. Accounts can grow because more employees use Claude, more departments adopt it and teams assign it more demanding work. A customer that integrates Claude into a product also buys more intelligence as that product gains users.

Growth is spreading across the customer base: customers outside the top five contributed roughly 79% of revenue in the September quarter, up from about 64% in the March quarter. Each of the five largest customers also hit record revenue in that quarter.

The bill follows the work. Total customer spending fell roughly 20% in the two weeks ended January 2, 2026 compared with the two weeks ended December 19, 2025; Anthropic attributes this to lower holiday activity, including code freezes. About 80% of first-half revenue was consumption-based, so an account grows when the customer finds more work and shrinks when that work stops. For spending to keep growing, customers must find enough useful work to justify their bills.

The Novo Nordisk case: in 2023 the company launched the NovoScribe project with clinical study reports. Statisticians produced results, tables and figures; medical writers assembled them into templates, followed by extensive quality checks. The system Novo built retrieves approved text and relevant statistical results, uses models to draft sections and preserves the link to sources so authors can verify the output. Its leads describe an architecture using Claude 3, Amazon Titan and a private deployment of ChatGPT. Novo reports a 90% reduction in the time needed to write clinical study reports, with people reviewing and approving the documents. The team has extended NovoScribe to device protocols and patient materials, and now uses Claude Code so domain experts can prototype ideas within an eleven-person team.

According to the author, Novo supplied the trial results, the approved language, the authors who understood the reports and the people responsible for signing off on them; the application connected those resources and sped up their work, and the team then extended what it had built to other documents.

Across the customer base, usage increasingly spans both Anthropic's platform and its applications: in the September quarter, more than half of customers with at least $100,000 in annual revenue used both Claude Platform (to build with the models) and at least one of Anthropic's own applications, up from about 30% in the last quarter of 2025. Counting employees misses part of what they buy: a company can use Claude to assist its people, run a recurring process or provide intelligence inside a product it sells, so an account can grow without adding employees.

Research can change what customers attempt. In its September 16 announcement, Novo said it would test Claude Science in selected R&D workflows, work with Anthropic on discovery challenges chosen by its scientific and computational teams, and use frontier models for software development. The author describes this as moving from using AI to describe finished research to using it to help carry research out. Anthropic has also used Claude to make scientific tools cheaper: in its September work on biomolecular modeling, it says it optimized more than thirty specialized biological models, with average speedups of nearly 2x on the tasks studied.