This is true for the next 6 months if that.

The future is models baked into directly into hardware, onto bare metal, serving 10k+ tokens/sec.

We can have our little models, they'll be serving a different customer.