Foundation Model Margins, Three Years In
Training is the capital expense everyone discusses. Serving is the recurring one that decides who survives.
The short version
Inference cost is falling faster than most forecasts assumed. Each step down the curve unlocks a class of product that was uneconomic the month before, which is why the application layer keeps surprising people.
Margins at the model layer are thinner than the headlines suggest. Training is a capital expense; serving is a recurring one. The business that wins is rarely the one with the largest model.
Every software company now describes itself as an AI company. Most of that is positioning. The ones that mean it have rebuilt a core workflow, not bolted a chat box onto the dashboard.