the simplest way to present the counter argument is:
- suppose you could train the best model with a single H100 for an hour. would that hurt or harm nvidia?
- suppose you could serve 1000x users with a 1/1000 the amount of gpus. would that hurt or harm nvidia?
the question is how big you think the market size is, and how fast you get to saturation. once things are saturated efficiency just results in less demand.
Supposedly DeepSeek trained on Nvidia hardware that is not current generation. This suggests that you don't need the current generation to make the best model, which a) makes it harder for Nvidia to sell each generation if it's more like traditional compute (how's Intel's share price today?), and b) opens the door to more competition, because if you can get an AMD chip that's 80% as good for 70% of the price, that's worth it.
I'm skipping over some details of course, but the current Nvidia valuation, or rather the valuation a few days ago, was based on them being the only company capable of producing chips that can train the best models. That wasn't true for those in the know before, but is now very much more clearly not true.
I can't see how DeepSeek hurts Nvidia, if Nvidia is what enables DeepSeek.