A great deal of the discussion of the so-called benefits or problems with AI comes down to the theoretical jobs that are (or are not) lost as a result of things LLMs can (or cannot do), or the equally theoretical productivity benefits that’ll come from using LLMs in place of (or in conjunction with) humans.
Anthropic’s Economic Index and OpenAI’s Economic Research Exchange are marketing operations that exist to propagate the (wrongheaded) belief that LLMs are either leading or will soon lead to massive economic or productivity shifts, even though little or no actual evidence exists to show that this is the case, other than the occasional story about LLMs make people worse or slower at their jobs or single lines in studies that are used (incorrectly) to prove that “AI is making it harder to find a job for young people.” In fact, Anthropic’s Head of Economics recently said there was “no material increase in the unemployment rate to date.”
These conversations materially detract from the actual harms or effects of AI, and exist only to make you scared that AI will take your job. They do not have any vested interest in expressing the actual economic effects of AI, which are, at this point, a simmering cauldron of different speculative bets on whether or not LLMs — a definitively niche technology — will create or become general-purpose software (per Roger MacNamee) that scales into the next Google Search, iPhone, or Microsoft 365.
As I’ve argued again and again, the AI industry’s revenues are, outside of Anthropic and OpenAI, incredibly small. Even in Exponential View’s deliberately-pro-industry analysis, there’s only around $110 billion in trailing twelve-month revenues across the entire industry, including OpenAI and Anthropic’s cloud spend.
For those counting at home, that’s $12 billion less than the $122 billion OpenAI raised in March, and a full $145 billion less than all AI startups raised combined in the first quarter of 2026.
Anthropic and OpenAI want you to talk about the theoretical so that you don’t focus on the tangible — their hundreds of billions of dollars’ worth of commitments, said commitments effects on the remaining performance obligations of hyperscalers and chip manufacturers, and the sheer scale of venture capital’s investment in AI, which (as I’ve argued in the past) largely allows for massive on-paper gains with little or no hope of liquidity.
To put this bluntly, I believe the entire conversation around AI’s theoretical relationship to jobs to be masturbatory and a conscious attempt to avoid having a messy conversation about the scale of the actions taken based on the flimsily-founded promises of AI labs and hyperscalers.
Today’s piece will dig into the true scale of the money needed to make AI make sense, by which I mean how much OpenAI and Anthropic will need to meet their commitments, how much money hyperscalers will need to pay off their investments, venture capital’s true exposure to the AI bubble, and what will have to go right for the bubble not to be, well, a bubble.
I’ll also make the case that the longer the bubble continues to inflate, the harder the basic economic puzzle of AI becomes to solve, as creating and deploying infrastructure becomes vastly more expensive — meaning that in order to achieve profitability, hyperscalers and neoclouds need to charge significantly more for compute than before, and the only two real potential customers are ones that cannot pay for it.
This will be a more more-pointed newsletter than usual, focusing on hard numbers and harder truths.
Let’s get it on.