IPO

Anthropic Now Uses More Claude Internally Than It Pays Its Employees

© Brecht Corbeel / Unsplash

Trending Topics auf Google als bevorzugte Nachrichtenquelle festlegen.

At Anthropic, the most important employee is now its own product, at least on paper.

The maker of Claude has tucked a number into its IPO prospectus that says more about how the company works than any revenue table: Since April, the value of internal Claude usage, measured at the prices Anthropic charges its customers, has exceeded the company’s monthly employee compensation, excluding stock awards. In other words, employees consume more in tokens than they receive in pay. According to Business Insider, which first surfaced the passage in the amended S-1, internal usage grew more than 250-fold between January and September. And more than 90 percent of the code added to Anthropic’s software in September was written by Claude.

What the Number Means

Ahead of its IPO, Anthropic wants to prove two things with it. First, that the company uses its own product more intensively than any customer, which makes the growth curve more credible: Revenue has grown twelvefold within a year, to $17 billion in the third quarter, while headcount has grown far more slowly. Second, that Claude is accelerating its own development. A company that has 90 percent of its code written by A.I. needs fewer developers for the same result, or ships more with the same number. For investors, it is the template for the argument Anthropic and OpenAI use to justify their valuations: A.I. that builds A.I., widening the lead instead of merely holding it.

The flip side is in the prospectus too, among the risk factors. Anthropic notes that relying on Claude for its own development is a risk if models introduce errors or security flaws into code that humans no longer fully review. The industry got a taste of that over the weekend: Anthropic’s VS Code extension on the Cursor platform was backdoored for almost twelve hours before the attack was noticed.

What the Number Does Not Mean

As impressive as the comparison sounds, it should be read with care. Internal usage is valued at list prices, meaning what outsiders would pay, not what the compute actually costs Anthropic. Because the list price includes a margin that Anthropic does not charge itself, the company actually spends less cash on internal A.I. usage than on salaries. The statement „Claude costs more than the workforce“ would therefore be false. What is true: If the employees had bought their usage externally, the bill would exceed their pay.

How big the gap is can be roughly estimated. Anthropic is said to have a gross margin of around 85 percent, and in the API business, meaning the sale of tokens, it is likely in the range of 80 percent: Of one dollar at list price, about 20 cents go to compute and infrastructure, and the rest is margin. If internal usage at list prices just exceeded payroll in April, it actually cost Anthropic about one fifth of the payroll. Put the other way around: Usage would have to reach roughly five times the payroll at list prices before Claude really costs the company as much as its employees. After a 250-fold increase in nine months, that is no longer a far-fetched order of magnitude; the prospectus simply does not say where usage stood in September, only that it has been above payroll since April.

Add to that the fact that tokens measure consumption, not results. A 250-fold increase can mean Claude has become extremely productive in nine months, or that agents run around the clock and discard large parts of their work. Both produce the same number. And „written by Claude“ does not mean „created without humans“: Every line still has to be reviewed, tested and accounted for, and architecture and security decisions remain with people. The costs do not disappear; they migrate, from salaries to chips, electricity and data centers, for which Anthropic has committed to $895 billion in spending.

That a company picks the best-sounding number ahead of its IPO is normal. What is remarkable is that Anthropic publishes it at all: Until now, the thesis that A.I. labs multiply their own productivity rested mainly on anecdotes. Now there is a number in an SEC document, and from the IPO at the end of November onward, it can be checked every quarter.

Rank My Startup: Erobere die Liga der Top Founder!
Werbung
Werbung

Specials unserer Partner

Die besten Artikel in unserem Netzwerk

Deep Dives

Wasner + Steinschaden | Der KI Podcast

News, Modelle, Strategien
#glaubandich CHALLENGE: SVAN triumphiert in Klagenfurt

#glaubandich CHALLENGE 2027

Österreichs größter Startup-Wettbewerb - Top-Investoren mit an Bord
© Wiener Börse

IPO Spotlight

powered by Wiener Börse

RankMyStartup.com

Steig' in die Liga der Top Founder auf!

2 Minuten 2 Millionen | Staffel 13

Alle Startups | Alle Deals | Alle Hintergründe

Future{hacks}

Zwischen Hype und Realität

Trending Topics Tech Talk

Der Podcast mit smarten Köpfen für smarte Köpfe

Weiterlesen

Newsletter

Founders Dispatch

Zwei Mal pro Woche kostenlos in die Inbox: die wichtigsten Startups, Deals und Tech-Entwicklungen aus Europa, handgeschrieben von der Redaktion.

Jederzeit abbestellbar. Mehr über den Newsletter