> There is already at least one company out there that showed its possible so others will follow IMO.

Are your referring to Taalas / chatjimmy ?

I definitely think we'll see an ASIC-like approach in the future, especially for embedded small models where it may require minimal silicon area and can result in near-realtime performance. But at the frontier, I don't think this is a solved problem and will continue to have model weight churn that will advantage more flexible general-purpose hardware.