alt Hacker NewsSo this is something I kinda agree with Jensen Huang on. Let me explain. The US export bans to China in this space take two forms:
1. ASML (a Dutch company) is not allowed to sell EUV lithography machines to China; and
2. NVidia is not allowed to sell their best GPUs (made with those machines) to China.
This is all done under the auspices of "national security" and the vague hand-waving notion of "military applications". China is rapidly developing EUV on their own. I kinda expect this to be commercially producing chips by ~2030. Other people believe it will take a little longer.
So what's the problem? By banning the sale of chips to China, the US has inadvertently created captive domestic demand for Chinese made chips because Huawei (etc) cannot buy Western made chips. This is really kickstarting China's lithography efforts because anything they produce has billions of induced demand with no competition from foreign imports.
That is I believe Huang's argument: we should be selling the chips but not the lithography. I agree with that assessment. But Us foreign policy really is just blinded and can be summed up as "China bad".
Replies
There are a lot of Game of Thrones type of stuff going on when it comes to chips and AI.
OpenAI, Anthropic, Microsoft, etc do not want Nvidia chips to sell to China because it makes the GOU supply constraint even deeper to American companies and it increases model training/inference competition.
It wouldn't surprise me if some of these players worked behind the scenes to try to oust Huang.
The idea behind the export controls is that by 2030 we’ll have some sort of superhuman intelligence and it’ll be impossible to catch up. I spoke to Biden’s ai advisor and that’s what he told me. They know China will catch up but are betting it’ll be too late for them by then.
A counterpoint I’d like to suggest.
Frontier labs (and Nvidia) have started using AI in chip development.
AI that requires frontier chips to train (or you steal chain of thought and try to distill…which has mixed results).
That in turn becomes a flywheel for developing domestic supply chains for chips.
See the latest reports for super intelligence takeoff. They (again assuming the timelines are semi correct as it’s all guesswork) show takeoff significantly before 2030.
The worst thing you could do is allow China to accelerate their domestic chip development via giving them access to chips needed to accelerate the AI flywheel.