Behind Huang Renxun's China trip, there is a hidden battle between AI chips and Nvidia's survival game

Nvidia founder Huang Renxun has returned to China again, and this is not his first visit to China this year. As early as late January, Huang Renxun came to China for a 5-day 3-city trip. But compared to the past, Huang Renxun's visit to China this time is particularly special, as it is his first visit after the escalation of tariffs between China and the United States.
Nvidia unwilling to give up on the Chinese market
Unlike his previous high-profile visits to China, Huang Renxun's trip to China this time appears somewhat low-key. Just responding to the media, Nvidia will regularly meet with government leaders to discuss the company's products and technology. At the same time, a clear difference is that in the past, Huang Renxun would avoid meeting with officials even when he came to China, while this trip took the initiative to meet with officials.
The meaning is very clear. Huang Renxun stated in some talks that China is a very important market for Nvidia and hopes to continue to cooperate with China.
About a week ago, multiple media outlets including National Public Radio (NPR) reported that the Trump administration had changed its plans for export controls on the Nvidia H20 after Huang Renxun promised a new investment worth $500 billion in American AI data centers. It is widely believed in the market that Huang Renxun successfully convinced Trump.
But soon, this news was denied by the new chip export license requirements issued by the US Department of Commerce. It is required to display that all AI chips, including H20, that meet their memory bandwidth and interconnect bandwidth standards must apply for a license for future exports to countries and regions such as China.
This means that even the H20 specially customized for China will be restricted. Nvidia estimates that this will result in an impairment and procurement loss of up to $5.5 billion in the first quarter of its 2026 fiscal year, which is directly related to the H20 chip and equivalent to 32.16% of Nvidia's total revenue in China in 2024.
If there were no ban, analysts at Daimo expect H20 chips to account for 12% to 13% of Nvidia's data center revenue in April 2025. However, with the impact of the ban, revenue is likely to decline by 8% to 9% in the coming quarters.
Faced with such a huge loss, it is also one of the important reasons why Huang Renxun came to China again, because he needs to send a clear signal that Nvidia is unwilling to give up the Chinese market.
According to Nvidia's latest financial report, as of January 2024, Nvidia's revenue in China was $17.108 billion, a year-on-year increase of 66%. And in Nvidia's fiscal year 2025, 53% of its revenue will come from regions outside the United States.
At the same time, China has more than 1 billion Internet users, which is the world's largest market. Nvidia has established an ecosystem in China that includes 1.5 million CUDA developers and 3000 start-ups.
More importantly, China is not only an important sales market for Nvidia, but also the most active market for AI big model applications worldwide. Many domestic enterprises have launched their own large model products, such as ByteDance, Tencent, Ali, iFLYTEK, etc. It is estimated that the AI market size will reach trillions by 2025.
At the same time, the rise of DeepSeek in China, especially the extensive use of H20 chips in its training process, also proves that Nvidia is still deeply involved in the creation of large models in China. It is also because of DeepSeek that Chinese tech giants have increased their purchases of H20 in the past few months.
Or further design chips for China
Huang Renxun's visit to China is urgent, not only related to Nvidia's layout in the Chinese market, but also facing challengers from the Chinese market. For example, companies such as Cambrian and Boren Technology are accelerating their efforts to capture the GPU market, while Huawei's Ascend 910B has approached A100 performance and has been commercialized in the Pangu large model. The market share of domestic chips has also increased from 9% in 2023 to 18% in 2024. The former is related to Nvidia's revenue, while the latter is more related to Nvidia's future survival space.
Interestingly, during his visit to China, Huang Renxun specifically met with DeepSeek founder Liang Wenfeng. According to the Financial Times, Huang Renxun and Liang Wenfeng discussed how to design next-generation chips for China to meet customer needs and regulatory requirements from both China and the United States.
For Nvidia, DeepSeek can be considered a loyal "fan" of Nvidia. The early DeepSeek model was developed entirely based on the CUDA architecture, and the code optimization solutions contributed by Chinese developers in the past three years accounted for 23% of the total contribution to the CUDA ecosystem. The R1 model that made DeepSeek famous was achieved through a 128 block H20 cluster, which achieved a throughput close to A100.
However, with the increasing tension in Sino US relations, especially after the tariff escalation, there has been almost substantial decoupling. DeepSeek chose to collaborate with Huawei Ascend to successfully convert CUDA code into CANN instruction set, breaking the iron rule that CUDA code must run on NVIDIA hardware, and also improving the efficiency of model inference for Ascend 910B by 40%.
DeepSeek not only uses PTX language to bypass the CUDA framework and directly reach the GPU bottom layer, but also deeply integrates with the Ascend ecosystem through the MindSpore framework, becoming both an inheritor of the past and a pioneer of the future.
This also indicates that not only has hardware been replaced, but in the future, even at the software level, it will no longer rely on Nvidia. How else can Nvidia go? Is it possible for Huang Renxun and Liang Wenfeng to launch chips that are more in line with current policies during their meeting. On the one hand, the performance of the chip will once again decrease, which customers may not accept; On the other hand, there is policy uncertainty, which means that even if chips can be sold to China, the United States only needs to change its policy to restrict them again.
Therefore, in more cases, Nvidia may adopt new ways of cooperation with Chinese companies. In a situation where hardware is already difficult to implement, collaboration in software may become possible. For example, maintaining Nvidia's ecosystem in China through joint research and development, technology licensing, and providing toolchain support with Chinese AI companies. Of course, this may also be subject to technical scrutiny from the United States.
In addition, Nvidia can also open the CUDA source code to Chinese customers to retain Nvidia's market share in China. You should know that the performance of Huawei Ascend 910B has reached 80% of H20, while the price is 40% lower. With the addition of the MindSpore framework, it is expected to leverage Nvidia's foundation of 1.5 million CUDA developers in China.
As of the end of 2024, Nvidia's workforce in China has approached 4000, a significant increase from around 3000 at the beginning of 2024. Obviously, the development of China's new energy vehicles and AI markets has also driven Nvidia's business growth in China.
During this trip to China, Huang Renxun also admitted that "deep cooperation with Chinese enterprises has enabled us to grow into a more competitive international enterprise.
Write at the end
When the United States continues to restrict Nvidia's AI chips from entering China, attempting to prevent the development of China's AI industry, even the H20 specially designed for the Chinese market has been restricted in one fell swoop, it is Nvidia itself that cannot sit still. The emergence of DeepSeek proves that the development of AI big models is not just about simply stacking computing power, but can also achieve better performance through algorithms.
China has the world's largest market for new energy vehicles, with its production and sales scale ranking first for ten consecutive years. The scale of China's artificial intelligence industry has reached 747 billion yuan by 2024, a year-on-year increase of 41.0%. It is expected that the market size will exceed 1 trillion yuan by 2025, becoming the core engine of global growth.
These areas are Nvidia's core revenue areas, but as China and the United States gradually decouple, Nvidia's hardware products have been gradually banned, and future cooperation may move towards the path of software. But no matter what, the unwillingness to give up the Chinese market has become a consensus among more and more American companies, as if there are always things that become so precious only after they are lost.