This seems similir to https://www.cerebras.ai business model. I believe they build their own hardware aimed at running LLMs extrenely fast. Not quite the same as burning AI to a chip, which I am really fascinated by.
Frank Wiles (npub122k…wx8u) has a few demos of using it and I think it's 2k to 3k or more tokens per second.
By the end of 2026, I bet most, if not al lof the frontier models will run like Cerebras does.
