Jensen Huang is anxious

Wallstreetcn
2025.07.17 09:00
portai
I'm PortAI, I can summarize articles.

NVIDIA CEO Jensen Huang emphasized the Chinese market at the third Chain Expo, announcing that the AI chip H20 has received U.S. export approval and will return to the Chinese market three months later. Although the H20 is a "special edition" chip with limited performance, it still faces challenges from local competitors in the Chinese market. According to IDC data, the market share of domestic computing power has increased from 14% to 34.6%. The launch of H20, NVIDIA's main sales driver in China, is due to the tightening of U.S. export policies to China, leading to a decline in its revenue share

On July 16, at the third Chain Expo, NVIDIA CEO Jensen Huang attended in a Tang suit, breaking away from his usual black jacket attire, and made efforts to deliver a speech in Chinese, which sufficiently demonstrated his emphasis on the Chinese market.

It was also during Huang's visit to China that NVIDIA announced that its AI chip H20 received export approval from the U.S. government for China, and this chip will return to the Chinese market in three months. This performance-limited "special edition" chip may still be in high demand, but after experiencing the "supply cut" turmoil, NVIDIA will inevitably face more local competitors in the Chinese market.

According to statistics from the third-party research firm IDC, from 2023 to 2024, the share of domestic computing power in China's data center accelerator card market has surged from 14% to 34.6%. Has the localization replacement made Huang anxious?

NVIDIA Reluctant to Abandon the Chinese Market

The "supply cut" of H20 began in April this year. On April 15, NVIDIA disclosed that it received a notification from the U.S. government on April 9 that exporting H20 to China requires U.S. government approval.

H20 was originally a "special edition" chip launched by NVIDIA for the Chinese market. Previously, NVIDIA had launched the A800 and H800, targeting the Chinese market based on the A100 and H100 chips, known as the "800 series chips." However, in the fourth quarter of 2023, the U.S. government tightened AI chip export policies to China again, banning the sale of the 800 series chips. As a result, NVIDIA launched H20 in the first quarter of 2024, with further limited performance, achieving only 15% of the peak computing power of H100.

H20 is based on the previous generation architecture Hopper, and there is a significant performance gap compared to NVIDIA's chips based on the latest architecture Blackwell. However, it has still become NVIDIA's main sales force in China, as another chip sold in China, L20, is two generations behind the latest architecture. IDC data shows that last year, NVIDIA accounted for 65.2% of the domestic data center computing power market.

Before the U.S. government restricted H20 exports to China in April, NVIDIA had hoped to exchange its commitment to invest in AI data centers in the U.S. for permission to export H20 to China. This is due to China's position as NVIDIA's second-largest market outside the U.S.

In the fourth quarter of 2023, after the ban on the 800 series chips, NVIDIA's revenue from mainland China has dropped from the previous 20%-25% to single digits. At the end of February this year, during a financial report meeting, Jensen Huang stated that NVIDIA's data center revenue in China had halved compared to before the export control measures were implemented.

To maintain this portion of revenue, H20 is crucial. In this regard, Yan Bo, head of the government services group at Chip Research, told China News Weekly that the frantic rush for H20 by major domestic manufacturers before the ban in April this year also indicates a huge demand for H20 domestically Due to export restrictions on H20 to China, NVIDIA incurred expenses of $4.5 billion from H20 inventory in the first fiscal quarter of this year, resulting in a gross margin of only 61% for the quarter, a decrease of 12.5 percentage points quarter-on-quarter and a decline of 16.9 percentage points year-on-year. NVIDIA stated that if this impact were excluded, the gross margin for the first fiscal quarter would have been 71.3%. Additionally, NVIDIA has $2.5 billion worth of H20 that could not be delivered, with a total impact on quarterly revenue of approximately $7 billion. As a chip "specially supplied" for the Chinese market, H20 is difficult to find in other markets.

AI Chip "Positioning Battle"

After obtaining export licenses from the U.S. government, Jensen Huang stated during a media interaction on July 16, "Since the announcement yesterday, I haven't had the chance to meet with any customers. It will take some time to enhance the capacity of the H20 supply chain, and NVIDIA will work to accelerate this process in the coming months. I hope to bring more advanced chips to China. In the coming years, as long as we are allowed to sell to China, we will do so."

H20's return to the Chinese market after three months clearly faces a more complex competitive landscape.

Domestic AI chip manufacturers undoubtedly hope to seize the opportunity for domestic substitution. For instance, Huawei launched the CloudMatrix 384 super node Ascend AI cloud service on June 20, which interconnects 384 of Huawei's most advanced Ascend AI chips and 192 Ascend CPU chips, claiming that the overall capability is comparable to NVIDIA's H100 launched in 2022.

Due to restrictions on obtaining advanced processes, domestic AI chip manufacturers are turning to cluster product routes. In June of this year, Huawei's founder Ren Zhengfei stated in an interview with the People's Daily that there is actually no need to worry about chip issues, as methods like stacking and clustering can yield results comparable to the most advanced levels.

In fact, since the ban on NVIDIA's 800 series chips in the fourth quarter of 2023, domestic manufacturers have begun to actively shift towards domestic GPU solutions. For example, the recently IPO-accepted companies Moore Threads and Muxi reported revenues of 121 million yuan and 51.41 million yuan respectively in 2023, while projections for 2024 are 432 million yuan and 742 million yuan, with revenue from AI computing undoubtedly being a major component.

The recent restrictions on H20 exports to China have further heightened the urgency among domestic manufacturers. For instance, during a performance exchange meeting in May of this year, Tencent stated that the company would look at both imported chips and those available in the domestic market to ensure that its AI strategy is not affected.

At the end of May this year, Jensen Huang stated that U.S. export controls would only enhance the global competitiveness of Chinese chip manufacturers. In explaining why the U.S. government allowed H20 exports, U.S. Secretary of Commerce Gina Raimondo stated, "We are not selling them the best stuff, not the second best, and not even the third best. I think the fourth best is what we consider acceptable. China is fully capable of developing chips independently, and what the U.S. needs to do is stay one step ahead of China's self-developed level so that China will continue to purchase U.S. chips." NVIDIA released its product roadmap for the next three years in March this year, including the next-generation chip Blackwell Ultra, which focuses on AI inference capabilities and has already been launched this year. In the second half of 2026, NVIDIA will introduce AI chips based on the next-generation architecture Rubin, with computing power metrics 3.3 times that of Blackwell Ultra. In the second half of 2027, it will also launch Rubin Ultra, with computing power metrics reaching 14 times that of Blackwell Ultra.

H20's return to the Chinese market will inevitably face more intense competition, but given that it still leads most domestic AI chips, this return is also of concern regarding its impact on the domestic AI chip replacement process.

Yan Bo believes that currently H20 will focus on clearing inventory, while the domestic AI application scenarios are rich and the demand for GPUs is large. In the future, if NVIDIA binds H20 sales with B30 or other "special edition" chips, many domestic manufacturers will still prefer to choose H20.

He stated that after experiencing the "ban-lift" event, the domestic demand side will adopt a "multi-pronged approach" to reduce the risks of a single supply chain. Manufacturers that have already adopted alternative chips such as Huawei Ascend and Cambricon are expected to continue maintaining a certain proportion of domestic chip usage. However, for end enterprises, H20 has advantages in the industrial ecosystem, which is also one of the main challenges faced by domestic GPU manufacturers when launching alternative products.

Author of this article: Chen Weishan, Source: China News Weekly, Original Title: "Jensen Huang is Anxious"

Risk Warning and Disclaimer

The market has risks, and investment should be cautious. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial status, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at one's own risk