Text & Image | Tang Jie

This year, Jen-Hsun Huang, CEO of NVIDIA, has repeatedly praised Chinese AI entrepreneurs and capabilities in front of global media, unapologetically criticizing the US decision to restrict AI chip exports to China as a catalyst for Huawei's rapid rise. However, public information suggests Huang's rhetoric has had little impact, with no consensus on AI chips between China and the US.

It was only during his third visit to China this year—making him one of the most frequent American business leaders to visit—that he announced the resumption of H20 chip sales and the introduction of a new, fully compatible RTX PRO chip tailored for the Chinese market. Coupled with his occasional humorous Mandarin and continued praise of China's technological prowess in AI chips, large models, and applications, his stance remains consistent with the past six months.

Known for his quiet achievements rather than eloquent words, Huang's intense "performances" over the past six months demonstrate a mature understanding of how to establish a friendly and confident image as a business leader in China. These statements also reveal that the resumption of H20 sales and customization of new GPUs for China fall short of his psychological expectations.

Compared to NVIDIA's latest chips, the H20 is more suitable for inference, offering only about 15% of the performance of the same-architecture H100 chip. The RTX PRO is also a GPU tailored for specific applications. With China's rapid progress in chip design and manufacturing, and sufficient data center capabilities, it's likely that domestic brands will soon surpass NVIDIA's market share.

NVIDIA's market share once peaked at over 95%, nearing a monopoly. However, since August 2022, when the US government issued regulations effectively banning the sale of NVIDIA AI chips, NVIDIA's market share in China's AI chip business dropped to 50% by May 23 this year, as admitted by Huang in an interview. This over 40% decline was almost entirely taken up by domestic chips like Huawei's Ascend and Baidu's Kunlun.

If Huang does not intensify his efforts, completely losing the Chinese market is imminent within the next one or two years. Even in the broader international market, NVIDIA will face fierce competition from Chinese manufacturers. Clearing the inventory of billions of dollars worth of H20 chips is merely a short-term objective. With NVIDIA's current market value of 4 trillion dollars, Huang's motivations extend beyond mere inventory management.

The impact on domestic enterprises must be viewed from two perspectives:

For companies like Tencent, Alibaba, and ByteDance, this is good news. The H20 is currently the best-performing chip in China utilizing the CUDA ecosystem, crucial for large model evolution and with no global substitute.

However, it's premature to conclude that H20's entry into China will stifle domestic chip enterprises. The unpredictability of the Trump administration and the US government in recent years has made many enterprises cautious, and cultivating domestic AI infrastructure ecosystems takes time. Therefore, abandoning innovative exploration in AI large models and their applications is not advisable.

In summary, the unbanning of H20 and the supply of new chips are good news for the tech industries of both China and the US. NVIDIA can leverage this opportunity to solidify its 4 trillion-dollar market value, while Chinese tech enterprises can continue innovating in large models and applications, feeding back into chip research and development while significantly expanding on the application side, striving to establish their own AI software and hardware ecosystem sooner.

01 Smooth Attitude Change

In the second half of last year, Jen-Hsun Huang faced criticism for making sensitive remarks. Turning around in the new year, he began praising China vigorously, making three visits in half a year, driven by two major events:

First, the emergence of Deepseek highlighted China's AI innovation capabilities and technologies on par with the US, signaling that "cost reduction in computing power" was on the horizon. The market realized that the future of the AI industry might not necessarily rely on expensive NVIDIA graphics cards. As a result, NVIDIA experienced its worst crash in history after the Deepseek release.

Second, from Trump's election victory until he officially took office in January this year, Huang maintained an ambiguous relationship with Trump, especially around the presidential inauguration, when Huang seemingly "avoided" the ceremony, choosing instead to visit various places in Asia. Only later, when trade disputes intensified, especially after the US used NVIDIA's products as a weapon against China, did Huang need to communicate more with Trump.

Fierce competition from China and being "weaponized" by the US government have become significant crises for NVIDIA, affecting both its business operations and reputation in the capital market. More critically, China's potential AI scale is sufficient to nurture a new company that can compete with NVIDIA and birth AI applications impacting the world. Missing out on these opportunities would be unbearable for NVIDIA.

Huang has publicly mentioned multiple times that China's current AI market is worth 50 billion dollars, a modest figure for a company with a 4 trillion-dollar market value. However, China is the second-largest country in the world in terms of AI-related infrastructure investment, committed to making China a world leader in AI. Morgan Stanley predicts that by 2030, the value of China's AI industry and related fields may grow to 1.4 trillion dollars.

If NVIDIA is excluded from this market, how much growth potential does the company have in the future?

Facts prove that long-term ceding of China's market share not only prevents the company from sharing booming market opportunities but also nurtures powerful competitors. In some scenarios, especially running large models like DeepSeek, the domestic ecosystem can already compete with NVIDIA, making the continued H20 ban meaningless.

If this market is given more time to develop, how fierce will the competition NVIDIA face in the future?

Moreover, China is the most active innovation center for AI applications and large models. NVIDIA's core competitive barrier is not the chip but the CUDA ecosystem. If the most active innovation hubs no longer use NVIDIA's ecosystem, the long-term consequences will be severe, meaning NVIDIA will no longer share the dividends of AI innovation in China.

It's not an exaggeration to emphasize how crucial the Chinese market is to NVIDIA, especially with up to 5.5 billion dollars worth of H20-related inventories still in hand this year. Just selling the inventory makes the past six months' efforts worthwhile. Additionally, rumors suggest that the next-generation H20 product is on the horizon.

It's rumored that NVIDIA is developing a castrated version of the AI chip called "B30" for the Chinese market. This chip will support multi-GPU expansion for the first time, allowing users to create higher-performance computing clusters by connecting multiple sets of chips. The B30 chip is expected to adopt the latest Blackwell architecture and use GDDR7 memory instead of High-Bandwidth Memory (HBM) and will not utilize TSMC's advanced packaging technology.

If HBM memory cannot be used, it's safe to assume this chip cannot handle inference and training of models with larger parameters. The final product parameters await NVIDIA's announcement, but with affordable prices and the CUDA ecosystem, it can still meet the needs of most small and medium-sized enterprise customers and some large enterprise requirements. The availability of higher-performance NVIDIA chips depends on the progress of domestic products.

02 AI Has No End

On July 18, Beijing time, OpenAI released its agent ChatGPT Agent. According to founder Sam Altman, it's a unified agent platform combining Operator's remote browser execution capabilities, Deep Research's network information integration technology, and ChatGPT's dialogue advantages.

"It can think and act, proactively select from the agent skill toolbox, and use its own computer to complete tasks for you," Altman described at the launch event. "This agent represents a new height in AI system capabilities."

Whether it's a new height remains to be seen through specific usage tests, but from the video demonstration, when users asked the Agent to help formulate an early retirement plan, ChatGPT Agent provided a PPT within 20 minutes. Neither the response speed nor content quality constituted a dimensionality reduction attack on Manus and Genspark.

Additionally, the globally popular Kimi K2 open-source model released a few days ago faced a similar situation. Due to its sudden rise to the top five globally, surpassing Deepseek-R1 to take first place in open-source large models and being specifically optimized for AI agents, the number of visits surged, causing a sharp decline in response speed. The company issued an announcement advising everyone to use third-party platforms.

Coupled with the emphasis by Huang and Alibaba Cloud founder Wang Jian in their dialogue that computation is the foundation of all intelligence, and artificial intelligence is a product of the explosion of computing power, everything leads to the same conclusion:

The computing power required for artificial intelligence is "infinite".

This is the reason and significance behind Huang's relentless efforts to maintain a "presence" in Chinese business, a "first principle" he's emphasized multiple times. Of course, the so-called infinity isn't just about scale; efficiency is an even more crucial aspect.

In fact, many are aware that the core barrier that truly propelled NVIDIA to the peak of global market value and near-monopoly of AI chips isn't the chip itself (where the truly powerful enterprise is TSMC) but the software-level CUDA ecosystem. And what software excels at is improving chip utilization efficiency, not the scale parameters of the chips themselves. This means that through software optimization, training and inference speeds can be faster with the same scale of parameters.

Compared to silicon-based chips, which have almost reached their development limits, improvements at the software level will be faster, larger, and more stimulating for demand expansion. With unchanged parameters, the less computing power required for training and inference models, the more enterprises/individuals will use AI, leading to greater total demand for computing power. For nascent AI large models, the demand for computing power brought about by this improvement is "infinite".

It's evident that AI will remain one of the core topics in the capital market this year, next year, and even the next five years. Extending from the "infinite" demand for computing power, NVIDIA, TSMC, Alibaba, Tencent, and the power sectors in the respective capital markets of China and the US will all be direct beneficiaries.

Additionally, for the domestic semiconductor industry, the unbanning of NVIDIA's H20 isn't entirely bad news. First, neither the H20 nor the subsequent B30 chip will provide much support for training. Domestic brands like Moxi Technology, which recently submitted its prospectus, already have products comparable to the A100 chip, with room for iteration and growth.

Secondly, the indecisiveness of the US government has already taught Chinese enterprises many lessons. Today, a ban is issued, and tomorrow it's lifted. All enterprises must consider a safer alternative plan. Therefore, the domestic ecosystem will continue developing at a "small-scale, fast-iteration" pace, which may be affected in the short term but remains optimistic in the long run.

03 Conclusion

Huang's ultimate goal is to lift all bans and ensure his products sell well in China, rather than engaging in step-by-step negotiations, consuming energy in politics, and gaining only minor progress.

Especially after Musk and Tim Cook gradually lost Trump's trust, Huang is the only American business leader with the weight and ability to play a "mediator" role between China and the US. But this role isn't easy, and any slight misstep may displease both sides. After Musk joined Trump's core team, the progress of FSD's landing in China has lagged, with certain risk factors behind it.

Therefore, if Huang wants to achieve commercial success that "blossoms in both China and the US," he must inevitably bear risks similar to those faced by Musk. For the founder of this 4 trillion-dollar market value company, this will be a brand new test.

Declaration: This article is for learning and exchange purposes only and does not constitute investment advice.