Comment by kimixa
4 hours ago
But their goal would be to ensure it only runs on their chips, and not any competitors. I can't see how they could do that if the best models truely were "Open".
I can see one of Nvidia's biggest fears is the inference hardware becoming commoditised.
The existence of inference ASICs has been priced-in to Nvidia's stock since Google started designing TPUs. It's not that new, really.
What Nvidia has the market cornered on is flexible GPU architectures. Nobody else has stepped up to the plate on that, and it's how Nvidia will butter their bread with robotics and future model training efforts.