Nvidia wants China's market share to secure the future of CUDA in the region — America's trade war threatens Huang's influence, and could bolster competition
Even as the head of the world's first five trillion dollar company, this week hasn't been great for Nvidia CEO Jensen Huang. Despite all his efforts, President Trump did not discuss the sale of advanced Nvidia GPUs with Chinese Premier Xi Jinping this week, nor did he give the official go-ahead for Nvidia to develop and sell a cut-down version of its flagship Blackwell GPUs.
The ongoing trade war between the two world's largest economies has Nvidia stuck in the middle, with the U.S. looking to maintain its technological lead, while Nvidia seeks to maintain its strong software and hardware position in Southeast Asia to fight off a growing domestic chip industry in China.
From diffusion to dissilution
Nvidia has been selling GPUs to China for a long time, but while it was previously mostly related to gaming and cryptocurrency mining, the growth of the AI industry has made these sales more crucial and the fallout more dramatic. Export controls put in place in 2022 curtailed its sales of its latest GPUs to China, but Nvidia has made bespoke models like the H20 and 5090D that offer reduced performance compared to the flagship models it sells in Western markets.
That continued in 2025 with the introduction of the AI Diffusion rule, which was designed to restrict Nvidia's sale of its top chips to China and other blacklisted countries around the world. This was ultimately pulled by the Trump administration mere days before it was set to become enforceable, but since then, it has swung back and forth, allowing sales, then disallowing them. Right now, Blackwell GPUs remain off the table for China, at least for now.
Even with Nvidia praising the administration for its stance on AI and global sales, the U.S. has maintained an antagonistic stance towards China on trade with various tariffs and policies, often brandishing access to Nvidia GPUs as a bargaining chip in negotiations.
China's response has been strong, cutting international companies' access to its domestic manufacturing facilities, and pushing for its own firms and government offices to utilize as many domestically produced chips as possible. Huawei has developed a rack-scale CloudMatrix 384, which is being pushed by the Chinese government, though it is not as efficient as Nvidia's latest offerings.
Both China and the U.S. are driving toward chipmaking self-sufficiency, or at least more self-controlled supply chains. For Nvidia, this has led to a waning of the company's influence in China. Access to its hardware is simply unreliable, so countries and companies are looking for alternatives, and the Chinese government is pushing companies toward homegrown chips.
The CUDA stronghold is under siege
CUDA has Nvidia's proverbial moat, shielding it from attack, and welcoming users into its castle gates. The software stack has become so ubiquitous in various industries that Nvidia GPUs are all but required for everyday use. Although competitors do exist, and AMD's hardware is by many measures, almost as capable as Nvidia's. But, without CUDA support, it's a major headache for companies to move to a new standard.
That's the business approach that Nvidia has taken globally, too. As hyperscalers deploy huge AI data centers to AI workloads, it wants their GPUs at the heart, supported by other crucial technologies like NVLink Fusion and ConnectX, just to name a few. This is to ensure Nvidia's long-term dominance in the data center segment.
With around half of the world's leading AI developers living and working in China, Nvidia wants those users working with its GPUs, harnessing the CUDA platform. But if those developers can't get Nvidia GPUs, they can't use CUDA, meaning that they'll use something else. We saw the first examples of how quickly this can happen at the end of September, when the latest DeepSeek model included support for China-native chips and the CANN platform, developed by Huawei.
CANN is also open source, which will aid in its adoption in China and elsewhere, and industry alliances of Chinese AI firms may help to consolidate their efforts in pushing away from Nvidia's near-monopoly.
Though DeepSeek still supports CUDA, its developers are encouraging users to leverage the TileLang kernel for prototyping instead. Although these models and the associated domestic hardware may not be as capable as Nvidia's platforms yet, China has a number of levers it can pull to make them more competitive.
Long-term investments from China make it clear that it wants to control its own destiny when it comes to the latest hardware. Its domestic chip suppliers are ramping up, and the investment is there to help that continue. But it will take time, and until that inflection point happens, Nvidia wants to have enough of a foot in the door that its relevance will never evaporate completely. Unfortunately for Nvidia, the U.S. government is loosening the company's control.
China is still behind
China purchased most of the million-plus Nvidia HGX H20 GPUs that were produced last year, and GPUs were also smuggled into the country. DeepSeek R2 was trained using Nvidia chips, despite a push by the authorities to use domestic chips instead. China's chips cannot compete when it comes to training workloads quite yet.
If Nvidia has one saving grace, it is that the chip industry is enormous, expensive, and takes time to change. China may want its own domestic chip supply, and it may have a decent AI development framework, and its hardware may even be pretty good for inference, even if it's not quite as efficient. But there's no denying that Nvidia and other Western companies are years ahead of even China's best, likely due to the lack of advanced EUV lithography machines, which are required to develop the latest cutting-edge nodes.
There are plenty of reasons that China views Taiwan through such hungry eyes, and access to tools which can produce the most cutting-edge process technologies is undoubtedly one of them. Even if China can convert a huge portion of its AI industry to using domestic chips and push for the development of its own domestic supply chains, those are going to take years to build up, and without ASML's advanced tools, they may be on the back foot. But, how long that will hold up for remains to be seen.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.
