What Anthropic’s S-1 shows about who pays for AI, what they report back, and how large the market could be

Anthropic, PBC filed a Form S-1 with the Securities and Exchange Commission on October 26, 2026, for a proposed sale of Class A common stock on the Nasdaq Global Select Market under the ticker ANT.¹ The filing has not become effective, so no shares can be sold yet.

Revenue was $16.3 billion for the six months ended June 30, 2026, up from $1.2 billion a year earlier. The company’s operating result moved from a $2.1 billion loss to a $693 million loss over the same periods, and the second quarter of 2026 was the only one of the six quarters filed with operating income, $974 million.¹

The filing has gaps. The share count, the price range, third quarter revenue, run-rate revenue at October 31, and the total of the company’s compute commitments are all blank, to be filled in by an amendment. Where this article needs a third quarter number, it uses an estimate and says so. Back-solving from the disclosed growth rate gives about $17 billion for the quarter and about $36 billion for the trailing twelve months. Neither figure comes from Anthropic.

The S-1 sorts customers by what they paid over the trailing twelve months. As of September 30, 2026, more than 11,900 paid $100,000 or more, up from more than 950 a year earlier. More than 1,900 paid $1 million or more, and about 250 paid $10 million or more. About 40 million customers paid less than $100,000 each.¹

Those $100,000 customers included every company in the Fortune 10, over 50% of the Fortune 500 and over 30% of the Global 2000. Within the Fortune 500, Anthropic says these included the 10 largest software and IT services companies, the 10 largest internet companies, the 10 largest banks, nine of the 10 largest pharmaceutical companies, nine of the 10 largest semiconductor companies and eight of the 10 largest healthcare companies.¹

The filing does not report revenue by industry. It says customers fall into two primary categories, enterprises and prosumers, and that the company does not formally track an enterprise category at all. It uses the $100,000 line as a stand-in for enterprise adoption.¹

Revenue from those customers was 90%, 57% and 76% of the total in 2023, 2024 and 2025. The share for the latest twelve months is blank in this filing. No customer accounted for more than 10% of revenue in the first nine months of 2026, and the top five together accounted for about 25%. The filing warns that many top customers have no long-term contract and can reduce or stop spending at any time.¹

Net revenue retention was approximately 900% at September 30, 2026. That compares what the same paying customers spent in the latest quarter with the same quarter a year before, excluding Pro and Max subscribers. The filing says its customer counts and retention metric have not been independently verified by third parties. The company says it expects its growth rate to fall and that recent growth “may reflect early-stage adoption dynamics.”¹

The S-1 includes profile pages for 21 customers, each chosen by Anthropic and each reporting its own results.² Sorted by the kind of evidence offered, 1 states a return on investment and 1 gives a dollar figure. 3 report a sales or volume result, 13 report speed or output, and 3 report only access or usage figures in their results tables.

The one stated return comes from a client that is not named, reported through Accenture, which sells this kind of work. Each entry below carries a note on what limits it. Apart from one study that Legora’s page calls independent and does not name, the pages do not say that anyone outside the companies checked the figures.

AIG’s page is the closest in the filing to the work of reviewing risk on a file. AIG, a customer since 2023, says its underwriting was slowed by the manual effort of taking in data from brokers, clients and third parties across many formats, and that it lacked the resources to review every applicable submission.²

It built a tool called Underwriting by AIG Assist, starting on the Claude 2.0 model. According to the page, underwriters used to spend weeks gathering data before starting a risk summary. They now begin each review with a curated risk summary and full audit trail generated by Claude. Where the tool is deployed, AIG reports that submission review time fell by more than 90% and that it reviews 100% of applicable submissions.²

These results cover one business line, reported by the company as of March 31, 2026. The page does not state a dollar return or a change in loss experience, so it does not show whether the additional bound business performs better or worse. AIG has extended the tool to eight lines of business, and in early claims pilots it says response time after a claim is filed fell from days to hours. In a multi-agent design it is exploring, the page says, underwriters would oversee every decision.²

“The way we think about AI and the work we do with partners like Anthropic is simple: it has to create value for clients and colleagues.”

Eric Andersen, President and CEO, AIG, in the S-1

Several customer pages in the filing describe connecting AI to company systems and data. Bristol Myers Squibb’s technology chief put it most directly:

“Most enterprise AI stops at the chatbot. The real prize is the untapped value still trapped behind decades of data silos... Anthropic’s Claude gives us the agentic capabilities, pace of innovation, and security necessary to connect our systems.”

Greg Meyers, EVP and Chief Digital and Technology Officer, Bristol Myers Squibb, in the S-1

The company reports that researchers can now query company data directly, and that time to answer a real-world evidence question across 30 to 40 external medical records and claims databases fell from up to two months to five minutes or less. Those are outside databases it subscribes to, so the figure does not measure a connection to internal systems.² DoorDash’s page ties usage to the same step:

“Use grew as employees connected Claude to DoorDash’s internal tools: weekly active users nearly doubled in the first month after rollout.”

DoorDash customer page, in the S-1

Travelers’ chief AI officer says the company pairs the models “with Travelers’ proprietary data and expertise.” SAP’s consultant tool is built on its knowledge base of more than 25 million documents. Salesforce’s chief executive refers to “the trusted data, workflows, and governance every enterprise runs on,” and Salesforce and SAP each disclose owning under 1% of Anthropic.²

What outside researchers say

The MIT NANDA report attributes stalled pilots to a “learning gap,” which it describes as “tools that don’t learn, integrate poorly, or match workflows.” It says the core barrier “is not infrastructure, regulation, or talent.”¹³ Its focus is on memory and adaptation, which overlaps with data access without being the same thing. PwC’s 2026 survey of 767 U.S. operations and supply chain leaders ranks integration complexity first among reasons technology investments fell short of expectations, followed by data issues and user adoption, according to summaries of the report.¹⁸

This research found no study that compares companies that block employees from connecting AI to internal data with companies that allow it. The evidence for a link consists of statements by customers selected by the vendor, and surveys asking executives what stands in their way. A firm that connects its systems may also differ from others in budget, staff or leadership, any of which could explain better results.

Why companies restrict access

The filing does not discuss why customers restrict access, but its risk factors describe hazards that bear on the question. They name prompt injection attacks that manipulate models. They warn that human reviewers “may experience alert fatigue, may lack sufficient context to evaluate agent decisions, or may not review agent actions in time to prevent harm,” particularly in high-volume deployments.¹ Banks and health care companies also work under federal privacy rules, including the Gramm-Leach-Bliley Act and HIPAA.

A WRITER survey with Workplace Intelligence, published in April 2026, reports that 35% of employees have entered proprietary information into public AI tools. WRITER sells AI software, and the figure is self-reported.¹⁷ The customer pages describe controlled versions of access: AIG cites an audit trail, DoorDash an internal platform running “security-approved” sessions, and Walmart says it keeps “human oversight and governance in place.”²

Microsoft’s annual reports give a five-year view of what a very large company looks like during the AI buildout. From fiscal 2021 to fiscal 2026, revenue rose from $168.1 billion to $331.8 billion, an increase of 97%.^5,6 Full-time employees rose 23% over the same period, from 181,000 to 223,000.

Headcount has fallen only recently. It peaked at 228,000 in fiscal 2024 and 2025, then dropped about 2% in fiscal 2026 while revenue rose about 18%. Revenue per employee went from about $929,000 to about $1.49 million by this article’s calculation.^20,21

Three things limit what this shows. Microsoft sells AI and cloud services, so part of its revenue growth comes from other companies’ AI spending, and its numbers do not measure what a buyer of AI gets back. Acquisitions also moved both lines: it completed Nuance in March 2022 and Activision Blizzard in October 2023.²⁰ And the gain came with a much larger capital bill. Additions to property and equipment rose from $20.6 billion to $115.9 billion, about 5.6 times, before counting finance leases.

Anthropic does not state a total addressable market. It says the work of knowledge workers represented more than $40 trillion in economic value in 2025, and it notes global technology spending of $4.3 trillion.¹ Neither figure says how much a company would pay for AI.

A different way to size it starts with what the filing says about customers’ budgets. Fortune 500 customers spent, on average, under 0.1% of their total expenses with Anthropic in the third quarter of 2026. Its top 10 Fortune 500 customers spent, on average, under 1%, more than eight times the overall average. The measure covers publicly traded companies only.¹

This article uses those two levels to ask a narrow question: how many companies could absorb each spending tier at 1% of revenue? A $100,000 bill is 1% of $10 million in revenue, $1 million is 1% of $100 million, and $10 million is 1% of $1 billion. The U.S. Census Bureau counts firms at those revenue levels.³ To estimate the world, the article multiplies the U.S. count by 3.2 to 3.8. The low end is the ratio of listed companies worldwide to U.S. listed companies at those revenue levels, and the high end reflects a U.S. share of world output of about 26%, on IMF figures.^4,22

The comparison in the last column is a rough one. The filing does not say how large these customers are, so a customer paying $1 million may have revenue far above or below $100 million. AI-native companies can spend well above 1% of revenue. The figure shows scale and does not measure market share.

Spending levels applied to the same revenue pool

U.S. firms with $10 million or more in receipts took in about $44.8 trillion in 2022, so the world pool on this method is about $143 trillion to $170 trillion. One percent of that is about $1.4 trillion to $1.7 trillion a year, and 0.1% is about $143B to $170B. For comparison, this article estimates Anthropic’s trailing twelve month revenue at about $36 billion. These are scenarios across every AI supplier, and they are not forecasts or Anthropic’s figures.

Several limits apply. Census data is from 2022 and includes nonprofits. Revenue counts sales between companies more than once as goods move through a supply chain. The multiplier assumes the rest of the world’s companies look like U.S. companies in size, and the S-1’s ratio uses total expenses where this method uses revenue. Commercial databases such as Capital IQ or Orbis would give a tighter world count.

Compute is the company’s largest cost, and the filing measures its efficiency as revenue per dollar of compute. Revenue per dollar of compute and infrastructure went from $0.40 in the first quarter of 2025 to $2.08 in the second quarter of 2026. Compute fell from 251% of revenue to 48%. Operating margin moved from negative 249% to positive 8%.¹

Two margins can be drawn from the first half of 2026. Using the cost of revenue line, which counts only compute serving customers plus some related costs, the gross margin is about 85%. Charging all compute and infrastructure, including research and internal use, the margin is about 45%. Both are this article’s calculations. The company says it manages compute as one shared resource, because customer serving and model development feed each other.¹

The company’s other efficiency measure is revenue per employee. Trailing twelve month revenue per employee rose from $1.4 million to over $6.6 million, with about 5,300 employees at September 30, 2026. It reports that over 90% of contributions to its codebase in September 2026 were written by Claude. At June 30, it held $131.7 billion in cash and short-term investments, having raised $141.1 billion in total. Operating cash flow was positive $3.9 billion in the first half of 2026.¹

The second quarter net loss was $66.3 billion. Most of that was a non-cash change in the fair value of convertible notes and related instruments, caused by the rise in the company’s valuation.¹

What the company has agreed to spend

The S-1 leaves the total of its compute commitments blank, but it describes several of its largest deals one by one and says it has others. Adding them gives about $782 billion to $807 billion. That sum is this article’s, and some of the deals are conditional or cancelable, so it should not be read as a firm obligation. It is about 6 times cash and short-term investments at June 30, and about 22 times estimated trailing revenue. The filing says these commitments “generally must be satisfied regardless of our revenue or utilization” and that it expects capital spending to rise significantly.¹

Surveys of U.S. workers show broad use and much less daily use. The Real-Time Population Survey, a nationally representative online survey of adults 18 to 64, feeds a tracker built by economists Alexander Bick, Adam Blandin and David Deming. It found that 39.2% of employed adults used generative AI for work in the past week in the second quarter of 2026, up from 28.2% in the third quarter of 2024.⁷ Gallup found in May 2026 that 52% of employed adults use AI in their role, 30% use it a few times a week or more, and 15% use it daily.⁹ A Federal Reserve note puts daily use at 12% in the Real-Time Population Survey for November 2025.⁸

Paid accounts are harder to count. Ramp, a corporate card company, reports that 55.9% of more than 85,000 U.S. businesses on its platform made a paid AI purchase in September 2026, and 73.7% in finance and insurance.¹¹ Those are Ramp’s customers, and the sample is not random. Microsoft said in July 2026 it had over 30 million paid Microsoft 365 Copilot seats, which counts licenses and not use.¹² This research found no representative count of U.S. workers with paid AI accounts.

Why the measures disagree

The Census Bureau’s business survey found 19.8% of firms using AI as of May 3, 2026, with 37% among firms of 250 or more employees.¹⁰ A worker survey counts people, and a firm survey counts companies, most of which are small. An employee can use a tool that the employer does not report. The Federal Reserve note recommends the Census survey for the share of businesses and the Real-Time Population Survey for the share of workers. It also notes that Census broadened its question in November 2025 from use “in producing goods or services” to use in any business function, and it reports a jump between the old and new series.⁸

On time saved, the Real-Time Population Survey reports 6.3% of work hours assisted by generative AI and 2.2% of work hours saved in its latest readings, up from 4.1% and 1.6%. The saved hours are self-reported estimates.⁷

Set beside the customer pages, surveys of companies report far less. Their questions differ, so the four figures below do not form a trend or a single measure.

The MIT NANDA report is the best known of these. It states that “95% of organizations are getting zero return” on $30 billion to $40 billion of enterprise investment. It describes itself as preliminary, covers January to June 2025, and draws on a review of more than 300 public initiatives, interviews at 52 organizations and responses from 153 leaders at four conferences.¹³ One critic, in the free portion of a paid post, lists the methodology as sound but with significant limitations and says the report is framed to support NANDA’s own agent infrastructure agenda.¹⁴ The report also found that workers at over 90% of the companies it surveyed regularly used personal AI tools, while 40% of companies had bought an official subscription.¹³

McKinsey’s 2026 survey, as reported by FM Magazine, found 37% of respondents attributing any EBIT impact to AI, about the same as a year earlier, with about 6% qualifying as high performers.¹⁶ KPMG found about 7% of senior leaders placing their organization in an established ROI phase, and those figures varied with how the work was managed: 14% against 4% where accountability was clear, and 15% against 3% with full visibility into AI operating costs.¹⁵

Individual perception runs higher than company results. Gallup reports that among workers using AI for seven or more purposes, 90% say it has had a positive effect on productivity, against 45% of those using it for one or two.⁹ That is self-reported and does not establish cause, since more productive users may simply try more uses. The customer pages and these surveys can both be accurate, because one presents selected successes and the others sample a wider group.

The S-1 has no customer page from a mortgage lender, an appraisal management company or an appraiser, and this research found no representative survey of AI use by appraisers. McKissock, an appraisal education company, has published trade survey results on where respondents find AI useful.²³

AIG’s underwriting tool takes in documents in many formats, produces a risk summary and keeps an audit trail. PayPal’s page says merchant underwriting and compliance reviews that once took hours now take a fraction of the time, without giving a figure. Bristol Myers Squibb’s researchers query records and claims databases directly. Of the three, only AIG gives measured results for a reviewer working from prepared material.²

Finance shows up near the top of the general usage figures. Gallup’s AI indicator page reports that 27% of workers in finance use AI daily as of May 2026, behind only technology.¹⁹ Census reports 33.9% of finance and insurance firms using AI, with no significant change since December 2025.¹⁰ Ramp reports 73.7% among its finance and insurance customers.¹¹ The gap between the last two reflects, in part, the different populations described in chapter 9.

The Signal’s earlier coverage follows the housing finance side of this: Fannie Mae’s work with Super Intelligence, the housing executive order, Fannie Mae’s third quarter and where appraisal risk sits, and a parallel between code review and appraisal review.

As of November 1, 2026, EDGAR shows no amendment to this S-1. The items below are blank in the October 26 filing. This article will be updated when they are filed, and its third quarter and trailing twelve month estimates will be replaced with filed numbers.

The filing’s own risk factors bear on everything above. The company expects its revenue growth rate to decline, and says it may do so significantly. It expects short-term compute purchases to compress margins in the near term. Customer spending fell about 20% in the two weeks ended January 2, 2026, compared with the prior two weeks, and the company expects seasonal effects to become more visible as growth slows.¹