> When the models stop improving

Soo, never? Granted Cerebras is a thing, if the process can be commoditized.

At the moment the area of edge inference at speed seems pretty bleak though.