The Economics of LLMs: Training, Inference, and Who Pays
Training a frontier model costs nine figures. Serving it costs fractions of a cent per token. The gap between those two numbers defines who can compete in AI — and who gets left out.
Current stance (June 2026): The AI industry’s economics are defined by a structural tension: training costs are rising exponentially while inference prices are collapsing. This creates a narrow window where only well-capitalized labs can afford to build frontier models, but the resulting models become cheap enough for anyone to use. The strategic implication is clear — if you’re building, you need a moat beyond model quality. If you’re buying, the market is your friend.
The training cost problem
Training a frontier AI model is one of the most expensive single endeavors in the history of technology. The numbers are staggering and growing: