Anthropic priced its IPO in other people's wages

Anthropic's IPO filing put a $40 trillion figure in front of investors and called it a "market opportunity." The definition inside the filing calls it something else.

With frontier AI’s potential to make them meaningfully more productive, our near-term opportunity is to expand this value and accelerate economic growth. — Anthropic

That is not a sales forecast. It is the estimated value of the work performed by the world's knowledge workers in 2025 ^[1]. A payroll wearing a market's name. It was the first of three times in ten days that the lab priced itself in units of other people's labor. Ten days before the filing, the second instance arrived in a report claiming Anthropic's AI agents had logged more work hours than its human employees since May 2026, and that entry-level technical hires had fallen from nearly 30% of recruits to 9% since March 2025 ^[2]. The report was unaudited; researchers who reviewed it described it as an attempt to inflate the IPO valuation. The third was the filing itself: Anthropic valued its internal use of Claude at the retail prices it charges outside customers, and said that figure topped its monthly pay, excluding stock awards, as of April ^[3]. Three numbers, one author, no auditor behind any of them. What all three skip is the same step. To say a machine worked more hours than an employee, or that its output out-valued a paycheck, you first need the cost of the work that was replaced. Gaurav Vashisht, who builds frameworks for measuring AI returns, argues companies cannot establish that baseline; budgets run on adoption charts rather than value because no one has priced the work the AI is standing in for ^[4]. Anthropic cannot show that number. Inside the same document, the labor market is headed the other way.

Competition in our industry relating to hiring and retaining employees with appropriate qualifications and on competitive economic terms is exceptionally intense and has escalated significantly in recent periods. — Anthropic

The same filing lists base pay for software engineers up to $850,000 ^[5]. The same week, quant firms bid as much as $800,000 for first-year packages to compete for the same talent ^[6]. The countable wages moved opposite the claimed substitution. The measured economy outside the filings was smaller than any of the three numbers. Oxford Economics found AI accounted for 4.5% of reported US job losses in the first eleven months of 2025, while cuts tied to market conditions ran four times larger, and concluded firms were using the AI label to "dress up layoffs as a good news story" ^[7]. PwC found 56% of CEOs saw neither higher revenue nor lower costs from AI over twelve months ^[8]. The grammar scales down as well: an eight-person startup's CEO says she would have to double headcount without AI playbooks, the hypothetical hires counted as value against $8,510 a month in Claude Code ^[9]. Stellantis is pushing AI to shrink software development from four years to twenty-four months ^[10]. The same week the filing landed, the money side of the industry was repricing itself. Amazon, Alphabet, Meta and Microsoft posted a record $201.9 billion in combined quarterly capital spending on buildings and chips, while their combined market value fell $900 billion ^[11]. OpenAI closed its latest raise at $11 billion of a $30 billion ask ^[12]. Weeks earlier, S&P had cut Oracle to BBB-, the lowest investment-grade rung, citing the debt behind its AI buildout.

Oracle is now a ‘show me’ story with limited visibility and lots of question marks out there. — S&P Global Ratings

None of this is to say there are no real returns. BlackRock estimates the payback on AI infrastructure at the hyperscalers — the biggest cloud-and-computing spenders — has shortened from four years to under three ^[13]. Google's enterprise backlog has grown tenfold in three years to nearly four times trailing revenue ^[14]. Microsoft is the only one of the three largest hyperscalers still running positive free cash flow ^[15]. Workday books roughly $600 million in AI annual recurring revenue ^[16]. The returns that can be counted in dollars are denominated in revenue, dollars coming in. The three IPO pricings were denominated in other people's wages ^[1][2][3]. And even the countable dollar returns stay small against the outlay: Workday's $600 million and the first "tangible" corporate AI returns of the season, a few hundred million of self-reported value, sit against the $201.9 billion the four biggest spent in a single quarter ^[17]. The wages are countable, and they are rising. The replaced work is the one number nobody has produced ^[4].