I haven't been following anyone baking models into ASICs, is it not still necessary to pack just as many transistors onto a chip, whether it's an NPU or GPU, ASIC or not you still need to hold hundreds of gigabytes in memory, so how is it cheaper to bake it onto custom silicon than running it on commodity VRAM? (Asking because I don't know!)

Not my area either! But my understanding is that there are more efficient methods of representing static numbers when you can skip the vram lookup.

https://taalas.com/

Is an example startup in this area claiming 16k tok/s on an asic for llama 8b. Qwen has a 27b model at opus 4.5 quality.