Per-model ASICs become economically viable
The case was made that the massive capital investment in training frontier models (billions of dollars) now justifies building custom ASICs per model to optimize inference efficiency.
Keep reading this one
You've read the thesis and who argued it. A free account opens the argument, what validates it, the risks the show raised, and the moment in the episode where it was said — 3 ideas in full a day, no card.