A curve with no unit

On 26 October 2026 Anthropic made public the prospectus with which it asks to list on the Nasdaq under the symbol ANT ^[1]. The headline that travelled first was the loss. The company billed 16,286 million dollars in the first half of the year and lost 80,748 million, almost five times what it sold ^[1]. It is the most repeated figure and the one that explains least.

As soon as the stock starts trading, the public will be able to buy the closest thing there is to a system that improves itself, and nobody yet has a unit to measure that curve. The prospectus says so in its own words: each version of Claude helps build the next, and since April the work those models do in-house has been worth more each month than the payroll ^[1]. The company writes that it is approaching what has been termed «recursive self-improvement» ^[1]. What the document does not carry is a measure of the rate at which that machine improves the next one, because no financial statement yet knows how to record it.

A loss that rises with the price

Of the 80,748 million lost in the half, 72,388 million correspond to the change in fair value of convertible notes, derivatives and simple agreements for future equity, the so-called SAFEs ^[1]. Another 8,757 million come from valuing the assets and liabilities tied to some note purchase agreements ^[1]. Neither item is money that left the till. Without them the half would have closed with a net profit of about 400 million. The prospectus attributes the jump to the increase in the company's valuation between 2025 and 2026 ^[1], so that loss grows precisely because Anthropic is worth more.

The operations tell another story, because the operating loss for the half was 693 million and in the second quarter the company earned 974 million before the financial items ^[1]. It is the first quarter with an operating profit of all those the document publishes. Revenue went from 41 million in 2023 to 386 million in 2024 and to 4,585 million in 2025, and the second quarter of 2026 alone brought 11,555 million ^[1]. Four of every five dollars arrive through consumption, that is, through what customers spend on tokens and not through subscriptions ^[1].

The line on page 184

It is a single sentence: «The value of our own internal Claude usage at list prices has grown by over 250 times from January 2026 to September 2026, and in April 2026, this monthly value first exceeded our monthly personnel expense, excluding stock-based compensation» ^[1]. The company presents it as proof of leverage. Each employee works with a team of agents, and that lets a comparatively small team take on work that once required a far larger one ^[1].

The numbers around the passage point the same way. Headcount grew more than fivefold between December 2024 and September 2026, to about 5,300 people, while quarterly revenue grew 86 times ^[1]. In September Claude wrote over 90 % of the contributions to the company's codebase, and the daily volume of code per engineer exceeded the 2024 average more than 45 times ^[1]. The engineers, the text says, increasingly direct and review the machine's work rather than write code line by line ^[1].

Two caveats are worth keeping. The first is that internal usage is valued at what a customer would pay and not at what it costs Anthropic to produce. Business Insider, which reported the finding on 1 November, put it bluntly: «Anthropic isn't actually spending more cash on AI than salaries. It's valuing its internal AI consumption at retail prices» ^[2]. The second is that the comparison leaves out stock, which weighs heavily in a lab's pay. Of the 1,857 million that personnel cost in the second quarter, 1,300 million was stock-based compensation ^[1]. What remains is around 186 million dollars a month, and that is the order of magnitude of the bar the machines' work cleared.

Closing the loop

The figure would matter little if the models only wrote the code of the corporate website. The prospectus says something stronger: the research, training and testing of each new generation of models now involve Claude ^[1]. Claude runs experiments, generates and processes training data, automates evaluations and safety audits, and writes and reviews code ^[1]. The company calls that recursive dynamic «closing the loop» and describes it as one of the most important structural advantages at the frontier ^[1]. Claude's success rate on open-ended engineering problems, the ones that come without a clear specification, reached about 90 % in September according to its internal benchmarks, some 50 points more than nine months earlier ^[1].

The same document puts the risk in writing: using AI to help build more capable AI «could also pose a considerable safety risk», particularly if it sharply accelerates progress ^[1]. The company says its responsible scaling policy sets out how to watch for this as it approaches what has been termed «recursive self-improvement» ^[1]. One of its risk factors warns that the danger grows if the systems come to automate much of the research and speed progress beyond its ability to evaluate and protect the new models ^[1].

A curve with no unit

The prospectus offers signs of that curve, but none of them is a measure. Internal usage multiplied by 250 is a value at list price, which rises just the same whether the machine does more work or each piece of work consumes more tokens ^[1]. The 90 % of code measures how much Claude writes and not how much better the writing gets, and the success rate that rose 50 points comes from tests the company designs itself ^[1]. Of compute multipliers, the improvements that let it train and serve more on the same hardware, the text says Claude helped write them and that their pace is accelerating, without giving a figure ^[1].

The accounts have nowhere to record it either. The only line in the financial statements that tries to put a price on it is the half's loss of 72,388 million, which measures how much more the shares promised to investors are worth ^[1]. The market's bet on the curve enters the books with a minus sign.

Until now, US exchanges offered that bet only wrapped inside other businesses. Alphabet carries DeepMind alongside search and advertising, and xAI has traded since June inside SpaceX, which absorbed it in February and went public with its rockets and its social network ^[3]. Anthropic arrives alone, with no other business to cushion the bet.

A twenty-year calendar

The models doing that work run on compute committed in advance. As of June 2026 Anthropic had minimum guaranteed commitments of about 265,600 million dollars, with terms of one to twenty years ^[1]. Compute is already its largest expense. It cost 8,945 million in the half, almost three times what personnel cost, and the company manages it as a single pool it shares among research, customers and internal use ^[1].

Nº 007 described in August the calendar bet of those who finance the chips: that demand arrives before the loans fall due. The prospectus shows the same bet from the lab's side. On one side there is a curve nobody knows how to measure, and on the other, fixed payments over twenty years.

A stock market knows how to price what grows fast as long as it can count it, and whoever buys ANT will be buying a rate of improvement that the company describes but does not measure, and that the accounts record only as a loss ^[1]. The signs in the prospectus admit very different readings, and the share price will depend on which one prevails.

What the listing does change is how often the reading happens. A company listed in the United States must file annual, quarterly and current reports, and the prospectus itself says so ^[1]. Until now the curve was read in interviews and private rounds, and from the listing on it will be read every quarter before the regulator. The figures that come closest to a meter are already in the prospectus: the value of the machines' work against payroll, revenue against headcount, and committed compute against what the company bills ^[1]. None of them measures the curve, but together they will say whether it bends upward or flattens.

The last implication is for those who use the curve without owning it. The public offering will not reach Latin America, where the shares can only be placed privately among professional investors ^[1], but in use the curve does have a unit, and that unit is the token. Ramp's index found in June that the 1 % of companies that spend most on AI devote 7,449 dollars a month per employee to it, «less than half the typical monthly salary for an engineer» ^[4]. That same figure is almost twice the median compensation of a software engineer in Mexico, which is 72,455 pesos a month, about 3,900 dollars, while Brazil's is around 2,800 and Argentina, Colombia and Chile fall in between ^{[5] [6]}. The token costs the same in Monterrey as in San Francisco, and the salary does not. If the engineer's craft becomes directing and reviewing what the machine does ^[1], the region offers the cheapest judgment on the market next to an identical machine, the reviewed result that replaces the billable hour of Nº 008. It pays for it, though, in dollars and to a supplier with 265,600 million in commitments to cover ^[1].

Counterpoint. A system improves itself, by a reasonable definition, when each version does a growing share of the work of building the next. By that yardstick Anthropic is the closest thing there is and not the thing itself, because it is still hiring and its engineers still direct and review ^[1]. The April crossing is measured at list price ^[2], the code and success figures come from internal tests ^[1], and a prospectus is a sales document. If the loop does not close at the promised pace, the 265,600 million are still there for twenty years.

THE PLAY

1. Measure your own crossing. Each month, divide what you pay for AI by your number of employees and compare it with your payroll per person, in dollars. It is the indicator Anthropic chose to present itself to its investors, and it will tell you when the nature of your costs changed.
2. Sell the reviewed result. Quote per unit delivered and signed off by a person, not per hour, because the region's advantage no longer lies in writing the code but in directing it and answering for it.
3. Don't give your token budget a single name. Split the spend between two or more providers, and try an open model for routine work before the token weighs more than your payroll.