The U.S. government has officially approved NVIDIA to sell its mid-range H200 AI computing chips to China and other "approved customers," but it must pay a 25% sales commission to the U.S. government. The latest Blackwell architecture chips from NVIDIA (including B200 and B300) as well as new AI computing chips from the future Rubin architecture series remain banned from being exported to China.

This is the result of CEO Huang Renxun's months-long lobbying efforts. In October 2022, the U.S. government first implemented significant export restrictions on advanced computing chips for China, requiring licenses to export specific high-performance chips to China; in January of this year, the Biden administration released stricter export control measures related to AI chips at the end of its term, further restricting the export of advanced AI computing chips to countries such as China, and establishing a tiered licensing system. This ban caused NVIDIA's market share in AI computing chips in China to drop sharply from 95%.

(Reuters reports that Chinese tech companies want to purchase NVIDIA H200 chips)

According to recent reports by Reuters, after the White House approved the export of NVIDIA's high-end H200 chip, demand from Chinese tech companies for this "specially approved" AI chip suddenly surged. Over the past few days, many Chinese tech companies have contacted NVIDIA to explore the possibility of large-scale purchases of H200 chips. Some sources said the demand from China was strong enough that the order quantity had already exceeded NVIDIA's current production capacity, and NVIDIA is considering expanding its production capacity. Industry estimates suggest that even with a 25% sales commission, re-entering the $200-$300 billion AI computing chip market in China remains significant for NVIDIA.

However, NVIDIA has not publicly commented on these reports. At the same time, Reuters reported that Beijing is urgently convening meetings to assess whether to approve this potential large transaction. Unlike some rumors from self-media, Chinese tech companies have not formally signed procurement contracts with NVIDIA.

(Performance comparison of AMD MI300X, NVIDIA H200, and B200 chips running a large model)

The main reasons mentioned by NVIDIA CEO Huang Renxun during his months of lobbying to the U.S. government were that China's AI large model development level is very close to that of the United States, and China's domestic AI computing chips (especially Huawei's products) are only a few nanoseconds away from the U.S. top level. Chinese industry professionals clearly know that this is flattery with an agenda. The U.S. government's approval of this sale mainly has three reasons:

First, as Huang Renxun said, the U.S. export ban on AI computing chips to China has not achieved the purpose of preventing the development and application of China's AI large models. Currently, open-source AI large models in China, such as DeepSeek, Alibaba Qwen, and Kimi from Moonshot, are among the most popular open-source AI large models in the world, with performance comparable to closed-source AI large models from U.S. companies, with little overall difference.

Second, lifting the export restrictions on NVIDIA's lower-end AI computing chips to China could potentially save NVIDIA's declining stock price. It is well known that major U.S. AI-related companies, including NVIDIA, OpenAI, Amazon, Microsoft, and Oracle, reached a series of circular investment agreements in October, causing their stock prices to soar. On October 29, NVIDIA's stock price reached a peak of $207, making its total market value briefly exceed $5 trillion, surpassing Germany's GDP.

However, the good times did not last long. Due to the poor performance of OpenAI's new large model, its revenue and financing goals remained distant, and it was also surpassed by Google's AI large model, which launched its own AI computing chip, leading to NVIDIA's stock price dropping back to over $170, and possibly continuing to fall. There is a risk of the AI bubble in the U.S. stock market bursting. If NVIDIA can sell AI computing chips to China, it could provide a "strong stimulus" to its stock price, maintaining the AI bubble in the U.S. stock market from bursting.

(Huawei's AI computing chip development plan)

Finally, allowing the export of NVIDIA's lower-end AI computing chips to China could strike down the domestic AI chip industry that China is developing. The Chinese government is actively promoting the application of domestic AI computing chips. Although the domestic AI computing chips from companies like Huawei are limited by domestic semiconductor manufacturing processes (recently reaching the 5nm level), their performance is not as good as NVIDIA's latest AI computing chips. However, due to China's power advantage compared to the U.S., they can use energy consumption to exchange computing power, using more chips in parallel to achieve the computing power that American chips can reach;

Additionally, open-source AI large models in China generally optimize the training and inference process, and when achieving similar performance to U.S. AI large models, their computing power requirements are only a fraction or even less than those of U.S. AI large models. This means that China can rely on domestic AI computing chips to complete its own AI industry iteration and development.

Aside from Huawei having already launched products, domestic GPU company Moer Xianzhi has gone public, while Bi Ren Technology and Mu Xi Technology are in the pre-listing phase. Future products from these companies will certainly replace NVIDIA chips. If NVIDIA regains access to the Chinese market, it will hinder the development of China's domestic AI chip industry, which is already a "strategy" openly stated by some U.S. politicians.

(Performance parameter comparison between H200 and NVIDIA's latest B300 chips)

So why are some Chinese tech companies still interested in purchasing NVIDIA H200 chips? The main reasons are two: First, although H200 is not NVIDIA's strongest chip now, its memory bandwidth and AI inference capabilities are significantly improved compared to the previously banned H800, which is sufficient to support the training and deployment of current mainstream large models. For Chinese enterprises struggling with computing power bottlenecks, this is almost like "a drought meeting rain." Second, the most powerful "moat" of NVIDIA's AI computing chips is its CUDA programming environment, which has been in place since 2006.

The CUDA platform can utilize NVIDIA GPU's graphics processing units for scientific computing. After nearly 20 years of development, the CUDA platform has accumulated countless scientific computing applications, which can save a lot of manpower and resources for the development and training of AI large models. This is something that most domestic AI computing chips lack (some companies claim their products can be compatible with the CUDA platform), which is very attractive to Chinese tech companies.

At present, local chip companies in China are booming. Allowing NVIDIA chips to flood the market would inevitably impact domestic chips, which is indeed a problem worth considering. If domestic chip companies do not develop, once the U.S. turns around again in the future, we may not find domestic companies to fill the gap. In fact, a U.S. Senate anti-China legislator has proposed a so-called SAFE Act, which aims to completely prohibit the export of high-end AI chips to China within the next 30 months, but the bill is still in the legislative stage.

Regarding this, the Chinese Economic Daily recently published an article stating that technological blockades will eventually become the "whetstone" for China's independent innovation. Whether China buys or how much it buys has long ceased to be decided by the U.S. The relaxation of H200 chips may become the "trout" in China's AI chip market, but it will not slow down the pace of domestic chip R&D.

Because Chinese enterprises that have experienced the "choke point" pain have already established the consensus that core technologies cannot rely on imports. David Sachs, an artificial intelligence and cryptocurrency advisor to former U.S. President Trump, recently told Bloomberg that "the Chinese government has come up with a plan to challenge the U.S. in the AI market and refuses to buy H200 chips for domestic companies because they want semiconductor independence."

Whether China will purchase H200 chips and how many it will purchase will become clear in the near future.

Original: toutiao.com/article/7584336030359224841/

Statement: The article represents the views of the author.