[dogma1138]

Why? The compute requirements would still continue to grow the more efficient and more capable the models become.

If it’s cheaper to inference you end up using the model for more task, it it’s cheaper to train you train more than models. And if you now need only 1000’s of GPUs instead of 10’s or 100’s of thousands you’ve just unlocked a massive client base of those who can afford to invest high six to low seven figures instead of 100’s of millions or billions into to try their luck.

[locallost]

It could be, but maybe the feeling is the investments now are already massive and everyone has jumped on the AI train. If you are suddenly 10x efficient, and everyone gets 10x more efficient, there's less room to grow than before. What you're saying makes a lot of sense, but it's one thing to write it on a message board and another to use it to back up your decision that affects billions of dollars you have in your fund.

The proof is in the pudding, you're welcome to prove "everyone" wrong.

[margorczynski]

Doesn't this situation also imply to some degree that China is focused on beating the US on AI and probably they will develop a competitor to NVIDIA that will cause margins to drop significantly?

They have a lot of very smart people and the will to do it, seems like a matter of time before they succeed.

[redwood]

It might take 5 yrs to find the use cases. That's what happened with the dark fiber from the .com boom. Go look at Cisco 2001 for parallels