Cheap Inference Squeezes Frontier AI Economics
Frontier labs must keep funding expensive model advances even as older capabilities become cheaper and harder to distinguish from the latest release.
Developed from a conversation between Pete Winn, Anthony and Andy David
Frontier AI has an awkward economic problem. Andy questioned how long a new model remains commercially distinctive before its capabilities begin to commoditise. Unlike a semiconductor fabrication plant that can remain valuable for years, a frontier model may have only a short lead before competitors and cheaper alternatives catch up. That forces labs to keep spending heavily on the next training run without knowing how long each advantage will last.
The investment only makes sense if a new model enables genuinely new work. Andy used résumé screening as the test. If a frontier model can perform the task faster, but four older models could already deliver the same result, the lab has spent enormous sums improving the machinery without creating a new use case. Better benchmark performance alone doesn’t establish enough additional economic value.
Pete drew the crucial distinction between creating and running a model. Training the frontier model absorbs the capital, while inference can offer strong margins once that model already exists. Even if model development stopped, providers could keep driving down prices and run more existing models in parallel. Frontier labs therefore carry repeated creation costs while yesterday’s capabilities become progressively cheaper to supply.