His thesis is that owning a good software platform is a prerequisite for a successful AI hardware company. He is (or was) using AMD as a stepping stone to creat

His thesis is that owning a good software platform is a prerequisite for a successful AI hardware company. He is (or was) using AMD as a stepping stone to create and flesh out that software platform (tinygrad) before starting custom ASIC development, instead of going the other way around with a costly ASIC first and trying to develop a software platform after, or trying to push support for an ASIC into someone else's platform such as PyTorch. Not sure I agree but it's an interesting idea. It's also worth noting that tinygrad is bypassing as many layers of AMD's software platform as they can get away with (MIOpen, ROCm, and even the userspace GPU driver).