AI Perspectives
Compute economics: how falling inference costs reshape AI startups
Inference keeps getting cheaper while frontier training gets more expensive. What this split means for startup margins and strategy.
Two cost curves define the AI economy. Training frontier models is becoming dramatically more expensive, while the cost of running a given level of capability — inference — keeps falling.
For startups this split is good news. Products built on top of models inherit those falling costs, so margins on agentic workflows improve over time without any extra engineering. Tasks that were uneconomic last year become profitable this year.
The strategic lesson is to design for model portability. Teams that can route work across several providers and open-weight models capture the cost curve, avoid lock-in and protect themselves against supply shocks in the chip market.
Frequently asked questions
Are AI inference costs going down?
Yes. The cost of delivering a given level of AI capability has fallen steadily, improving margins for companies that build on top of models.
