AI feels like a licence to print money — sky-high valuations, breathless demand. But peer at the unit economics and a surprising picture emerges: for many products, every query you run might be sold at a loss. Welcome to AI’s margin problem. The uncomfortable math Running a large model isn’t like serving a web page. Each response burns real compute on expensive hardware. When a company charges a flat monthly fee for “unlimited” use, the heaviest users can cost far more to serve than they pay — and those costs scale with usage, not with a one-time build. Why inference is so expensive The bill is dominated by specialized chips, power, and cooling. Bigger models and longer answers cost more per request. Unlike traditional software, where serving one more customer is nearly free, AI carries a stubborn marginal cost that doesn’t vanish at scale. Who’s subsidizing it Right now, investors. Many AI products are priced to win market share, n...