Inference costs can't be faked though, since the model can be run locally by anyone with capable hardware.

Even if the whole story about the training cost was fake, R1 and the distilled models are still very efficient at inference.