> know exactly what hardware they'd like that would be best suited for their task. Thats the fun thing about AI hardware. By the time you develop it (which tak

> know exactly what hardware they'd like that would be best suited for their task. Thats the fun thing about AI hardware. By the time you develop it (which takes years and billions), the workload you're targeting is already changed or obsolete. Nvidia is the exception because if you are an AI scholar, you are building your research around whatever capabilies Nvidia GPUs have. Thats why AMD and Intel have quite literally chosen the "emulate Nvidia" path and axed more specialized accelerators.