An interesting thread from Pierre Ferragu about why, even though Nvidia's GPUs are perpetually one generation behind on cost/performance, they win in the marketplace. He thinks that the flexibility of the GPUs are needed by most developers and therefore the custom approaches (Google TPU and presumably Dojo) don't and probably won't gain much share.
Not sure that I buy his argument, but it is interesting and seems worth considering.
Not sure that I buy his argument, but it is interesting and seems worth considering.