> For Nvidia: all it will take is for one smart grad student to find a better training algorithm to destroy 75% of their value.

I don't think so. We've already seen several generations of better training algorithms, but they all rely on CUDA, hence on NVidia.

For inference they already face competition: that deal OpenAI announced with AMD .. at a certain point it will be worth broadening away from CUDA.

Yep, and don't forget about Mojo, which has the potential to make such a transition less painful.

Possibly, yes, but so far, it's barely noticeable, hardly a threat.

Computation on human brain is on a totally different substrate than silicon. It's in memory and highly error prone.

It's questionable a mere algorithm would get us there without a fundamental change in computer architecture. (in terms of Intelligence / W)

Is that true? I was under the impression that one reason NVIDIA is so entrenched is that their GPUs are highly flexible. Versus the ASIC offerings of other specialized players are more focused on transformers remaining the dominant architecture.

Maybe the better algorithm will be able to run on GPUs.

the 1 million is debatable, but you can't argue with the power consumption of 20W.

that said, if all the current infrastructure could run brain-equivalent models at 20W each, I'd wager we would have much more demand than currently.

There's speed, efficiency, and determinism. Pick 2

But the flip side is excess capacity will be used for other fields which are not getting a chance.

On the other hand, GPUs die down fairly quickly, so this excess capacity had better prove useful rather quickly.