The new economics of inference: why serving models now costs more than training them

For years the headline number in AI was the cost of training a frontier model. That framing is changing. As adoption scales, the recurring cost of inference — serving predictions to users around the clock — increasingly dwarfs the one-time cost of training. This premium analysis breaks down the unit economics: the drivers of per-query cost, the role of spec