Anthropic went public with nearly 12,000 customers spending at least $100,000 a year. It also disclosed $70.7 billion in computing commitments due in twelve months. Between those two figures lies the question confronting the entire industry.
On Thursday, investors who secured shares in Anthropic at the $83 offer price ended the day with a gain of 17 per cent. Anyone who bought at the opening price of $111.50 had a different experience. By Friday’s close, the shares were worth roughly $88, leaving that buyer down about a fifth in little more than a day. Both had bought the same company. The difference was the price they paid for its future. (Anthropic, Reuters
Across Wall Street, investors spent Friday, 20 November, reconsidering what they had paid for much the same future. The Nasdaq 100 lost 5.6 per cent on the day and 10.4 per cent over the week, its worst weekly performance since the pandemic. Measured from its 6 October intraday high, the index had fallen 20 per cent, the customary threshold for a bear market. The S&P 500 fell 3.7 per cent on Friday. Nvidia lost 14 per cent
A week of trading can tell us how sharply investors have changed their minds. It tells us little about whether a business will renew its Claude subscription next March, or whether a data centre now under construction will earn its cost over the next decade. For those questions, Anthropic’s flotation documents offer better evidence. They contain a credible case for the boom and an equally serious reason to doubt the prices it produced.
The customers are paying. The bills are coming due.
At the end of 2025, Anthropic had more than 1,500 customers from whom it had earned at least $100,000 over the preceding twelve months. By the end of September 2026, the number exceeded 11,900. That is a substantial body of customers willing to commit serious money to a young technology. Anyone arguing that the AI boom rests entirely on speculative capital has to account for those sales. (Reuters)
There is also evidence that Anthropic can serve a growing business more economically. In the first nine months of 2025, the company spent $4.4 billion on computing and infrastructure, equivalent to 172 per cent of revenue. A year later, the bill had risen to $18.8 billion, but its share of revenue had fallen to 57 per cent. That improvement is the strongest argument against a sweeping verdict that the economics of AI are broken. Revenue grew considerably faster than one of the company’s largest expenses. (Anthropic filing with the US Securities and Exchange Commission)
What shareholders must judge is how much of that improvement can endure. Anthropic earned an operating profit of $974 million in the June quarter, then recorded a $567 million operating loss in the September quarter. It identified two significant pressures. It had purchased $5 billion of computing capacity under short term agreements, which it said were considerably more expensive than its multiyear arrangements. It also recorded a $2 billion noncash charge associated with its donation matching programme. Attributing the quarterly loss entirely to the cost of running Claude would give readers a misleading picture. (Anthropic prospectus)
Here is the commercial dilemma. Buying computing capacity at short notice is expensive. Booking it years ahead can secure a better price, but it requires management to judge today what customers will buy several years from now. Anthropic says approximately $70.7 billion of its computing commitments fall due in the twelve months following 30 September 2026. If demand keeps growing rapidly, those advance bookings could prove astute. If customers spend less than expected, or a competitor offers comparable performance more cheaply, the commitments still fall due. (Anthropic IPO fact sheet and prospectus)
Some of the investors are also suppliers
The widely reported $120 billion fundraising deserves a closer look. The final public offering amounted to $93.5 billion before fees, including shares sold by existing holders. Anthropic itself received approximately $91.6 billion in gross proceeds from that offering. A further $26 billion came through a private placement with seven companies, among them Nvidia, Microsoft, Google, Broadcom and AMD. Together, the two transactions were worth $119.5 billion. (Anthropic’s closing announcement)
Consider what that means from Nvidia’s position. It has a commercial interest in a world where AI developers need more processors. Through the placement, it also owns an interest in one of the developers making those commitments. For Anthropic, the arrangement brings in capital from companies with the expertise and resources to help supply the computing power it needs. There are sound commercial reasons for both sides to participate. (Anthropic’s closing announcement and IPO fact sheet)
The danger for investors lies in counting the same relationship twice as evidence of demand. A chip company’s decision to invest in an AI laboratory demonstrates confidence in that laboratory. An order for computing capacity demonstrates that the laboratory intends to expand. The independent test comes later, when a bank, a software company or a law firm renews its contract at a price that covers the cost of serving it. Anthropic’s customer figures show that money is entering the industry from outside. Its spending commitments show how much more will have to follow. (Reuters, Anthropic IPO fact sheet and closing announcement)
What the Nvidia inquiry can explain
Nvidia had its own immediate reason to fall. Reuters reported that the US Securities and Exchange Commission was examining its accounting. The report establishes that an inquiry was under way. Its outcome remains open. For a shareholder, even that uncertainty matters because forecasts of future earnings depend on confidence in the figures from which those forecasts begin. (Reuters)
Yet Nvidia alone cannot explain the retreat. The Philadelphia Semiconductor Index was already far below its June peak and fell a further 10.7 per cent on Friday, taking its cumulative decline to 43 per cent. The chronology matters. Investors had been marking down chipmakers for months before Anthropic sold its first public share. Expectations of higher US interest rates also weighed on technology stocks. The more a company’s valuation depends on profits many years away, the more sensitive that valuation becomes to the return investors can earn elsewhere in the meantime. (Financial Times and Reuters)
The next meaningful evidence will come from purchasing departments and company accounts. Do large customers renew and increase their spending? Does the cost of serving them continue to fall as a share of revenue? Do cloud providers proceed with data centres on the promised timetable? A sustained deterioration across those measures would give the bears a stronger case than Friday’s share prices alone.
Anthropic has already demonstrated that thousands of businesses will pay substantial sums for AI. Its prospectus leaves a harder question for shareholders. A company can build a product its customers value and still pay too much for the capacity to deliver it. The gamble behind the AI boom is that paying customers will arrive fast enough, and stay long enough, to justify the bills the industry has already agreed to pay. (Reuters and Anthropic IPO fact sheet)