There are mixed reactions from Chinese companies and officials to US President Donald Trump's decision to let Nvidia sell its second-most advanced chip to Chinese customers. That's according to Chim Lee from the Economist Intelligence Unit. While tech companies desire the advanced chips for AI model training despite the 25% fee, officials worry about increased dependence on US technology. Lee also notes that switching to domestic alternatives like Huawei is difficult due to Nvidia's entrenched software ecosystem.
00:00Chim, now that U.S. President Donald Trump has signaled that advanced AI chips from NVIDIA, like the H200, will be allowed into China, how are companies and officials reacting there? Is this a victory for them?
00:13I think it's very mixed, to say the least. I think for some of the people who are thinking about faster AI model development and these sort of things, that is probably a win. There are advantages of using NVIDIA's chips, at least in today's development environment.
00:30But on the other hand, for a lot of the national security advocates, and including a lot of policymakers in China, that if the U.S. allows chips exports to China, it will increase the dependence on U.S. technology, which is not something that many of the national security policymakers in Beijing want to see.
00:50So I think I would say overall is a very mixed landscape.
00:53First, there was the H20, now the H200. The offer from Trump has been to allow these chips into China in exchange for a fee, this time for 25%. But what about pricing? If that 25% is passed on to Chinese customers, are these chips still competitive in China?
01:10If we look back into the H20 chips, well, they are clearly very expensive compared to, you know, in terms of these capabilities, but we still see very strong interest. And I think we can sort of see similar scenario playing out in the H200.
01:26And I think in this context, it's very important to note that H200, they are quite advanced, compared to what is currently available in terms of like domestically produced chips in China. So when it comes to model training, we have seen reports about how the top AI companies in China, they still really want like these advanced chips for their model training.
01:50Another possibility is what we saw in some of the recent reports saying that China's companies are actually training their models now overseas, where the access to chips are less restricted. So in this context, a lot of them are also waiting. That's why a lot of them are actually waiting for the new chips to arrive.
02:09All of what we're talking about here assumes these chips even make it into China. But there's been reporting that Beijing may limit the import of these US chips. What's your sense of how restrictive this may end up being?
02:20I think one trajectory that we might see is that we do see very strong advantages in terms of using more advanced chips that are outside of China for model training.
02:31So these are probably the areas that I think China's policymakers will allow access.
02:38At the same time, when we talk about deployment, and here where domestic chips are actually quite decent at this point,
02:45I think that trajectory is that they're probably not going to be allowing those chips, the foreign chips, essentially, to be used in deployment.
02:54Even if Chinese companies wanted to switch to 100% Huawei, for example, there's been reporting about how difficult that's been for some,
03:01including failed AI training runs that forced companies to switch back to NVIDIA.
03:06Why has it been so hard for Chinese companies to stop using NVIDIA systems?
03:09Yeah, there are quite a few reasons to that. But I think one of the major reasons is really the software ecosystem.
03:17So NVIDIA software ecosystem is something that has essentially become a global standard.
03:22So a lot of people, when they're developing their models in terms of the hardware side of things,
03:27like how they are all linked, they will be using NVIDIA standards.
03:31And Huawei, in particular, is using a different set of standards.
03:35Now, there are some Chinese AI chip companies that are actually trying to make their chips compatible with NVIDIA standards.
03:45But I think that overall, this kind of switching cost, it's almost like switching between an iPhone and an Android.
Be the first to comment