Another factor is that model capabilities are topping out, but inference costs of cheaper models are absolutely plummeting.

In other words, it’s a race to the bottom. And inference will become a dirt cheap commodity.

…And what do we need all those datacenters for if inference is dirt cheap?