A successful case would presumably reduce Nvidia’s markup on GPUs. That by itself would make GPUs more affordable and more widely available to people working on AI.

The problem is if development of those GPUs has such a high fixed cost that in a more competitive environment revenue does not pay for it. But that seems unlikely to me at a first glance. AMD, Apple, Google, and Facebook all have their own chips. They just don’t have CUDA to go with it, or don’t sell their chips to customers.