Inference is super cheap, and getting progressively cheaper.

Greenlighting training a new foundation model is very expensive, but is also a human decision that can be postponed based on available capital.