AI is slowing down: the economics behind the industry's growth problem

A widely read essay by Ed Zitron, published on his newsletter Where's Your Ed At, argues that the AI industry's spending commitments now require revenue growth so steep that any slowdown becomes an existential problem for the companies involved.

The arithmetic of the buildout

Citing February data from Sightline Climate, the piece counts 190 gigawatts of planned data centre capacity. Applying Nvidia chief executive Jensen Huang's estimate that data centres cost $80 billion to $100 billion per gigawatt, Zitron puts the total at $9.5 trillion to $15 trillion - far above the "$3 trillion" figure often quoted. Charged at roughly $12.5 million per megawatt, 190GW would need about $1.75 trillion in annual revenue; even a half-built scenario requires $875 billion. The essay also notes that 54% of Nvidia's revenue comes from three unnamed clients, and that Anthropic has committed $330 billion to compute and chips across Google, Amazon and Microsoft, plus $30 billion with CoreWeave and $15 billion with SpaceX, against projected annual revenue of $174 billion by 2029. OpenAI has made more than $770 billion in compute commitments and is projected to burn at least $852 billion through 2030, Zitron writes.

Demand is concentrated

He argues Anthropic and OpenAI account for the vast majority of AI compute demand - at least 70%, possibly 80% to 90% - and for 89% of all AI startup revenue, according to The Information. Their combined projected 2026 revenues of roughly $60 billion would have to reach about $400 billion a year, meaning both must grow by hundreds of percent and raise hundreds of billions more in funding. Zitron reads recent equity sales by hyperscalers, including Alphabet's $85 billion raise, as a sign that debt is getting harder to obtain.

Signs of a slowdown

The clearest evidence, in his view, comes from billing. Anthropic and OpenAI moved customers to token-based pricing in the first quarter of 2026, and within months chief financial officers were publicly questioning the return on AI spending. A KPMG survey cited in the piece found only 26% of companies have a comprehensive view of their AI costs, while 22% have no visibility or learn the total only after billing. Companies have started to cap usage: Uber limited employee AI spending to $1,500 a month after burning its annual token budget in a single quarter, T-Mobile temporarily set a $2,000 monthly limit, and Brex gives engineers $500 a week in tokens and other staff $5. The essay concludes that the industry needs to roughly tenfold its revenue base to justify the infrastructure already under construction.