Instead of feeling insulted, you should just refute me. Why are ASICs a desirable alternative to GPU/NPU inference here?

The "baked-in" weight technology you are describing uses enormous amounts of memory. There is not enough SRAM or flash ROM in the world to put a "modern frontier model" on every iPhone, even if you quantized it to 1 bit. It is scarce and expensive, and we are in a global shortage of memory and flash storage.

If we lived in a world where Cerebras, Taalas and Groq could dethrone Nvidia, they would have done it years ago. Turns out, cutting architectural corners has consequences, and they're extremely offputting for the majority of inference providers.