The true profitability of frontier AI labs remains to be established. Right now we’re constantly seeing this cycle of new models being trained, released, and then scaled back with smaller context windows and hidden behind model routers to avoid paying the full costs of inference. But these tricks blunt the usefulness of the frontier models, leading to competitors stealing customers with their latest models. It’s an arms race where the costs of training dwarf everything else, but even the costs of inference may not be profitable.

At the same time, Chinese companies keep releasing new open weight models which are severely undermining the value proposition of the frontier AI labs. Customers of the frontier labs may be unwilling to subsidize model training if they can bypass that cost and run the open weight models on their own hardware.