NVIDIA's computing chips are centered around their performance advantage as their core competitive edge, while Chinese chip products, such as Huawei's Ascend, are accelerating their development, attempting to close the gap with more advanced technology. According to reports, due to lackluster demand for the AI chip H20 series specifically designed for the Chinese market, NVIDIA has lowered its prices. Initially, when the H20 series chips were open for pre-order, the price range was from $12,000 to $15,000, with some distributor quotes even exceeding 110,000 yuan. However, the price of a single H20 chip has now dropped to about 100,000 yuan, and the price of an eight-card server is between 1.1 million to 1.3 million yuan. In certain cases, the price of the H20 is over 10% lower than the Huawei Ascend 910B.

Several individuals within the supply chain have confirmed the sluggish demand for H20 and the price reduction. The H20 is currently the most advanced NVIDIA chip legally available on the Chinese market, primarily used for training large models. Over the past two years, the United States has continuously tightened export controls over Chinese chip exports. After two rounds of export controls in 2022 and 2023, not only was the H/A800 series banned from export, but all advanced GPUs followed, including the H100.

After the second round of control, NVIDIA introduced three new chips with reduced performance for the Chinese market: H20, L20, and L2. Of these, the H20's performance is only about 15% of that of the H100, and in some cases is even worse than the Ascend 910B. Since the H20 began accepting pre-orders in the first quarter of this year, only large customers have been eligible to pre-order, including major Internet companies such as ByteDance and Alibaba, which have also become the main purchasers.

On March 30, 2023, the United States further escalated restrictions on chip exports to China, adopting a "case-by-case review" policy, strictly limiting the sale of more advanced AI chips and semiconductor equipment by companies like NVIDIA and AMD to China. Although this ban did not involve new special products like H20, it is known that this policy escalation led to decreased interest in this series of products by Chinese companies.

Due to the latest ban adjustments, Chinese companies have turned their attention back to previously restricted products. The H20's computing power has been weakened, and meanwhile, there is still inventory of NVIDIA products on the market. Compared to these products, the H20 lacks both performance and cost-effectiveness. Rumors about Alibaba's pre-order of over 30,000 H20 chips have not been officially confirmed. Alibaba Cloud's procurement plans often involve using these products to build their own cloud servers to provide powerful computing power to the public. A professional from Alibaba Cloud indicated that the market still prefers older NVIDIA products. The newly launched H20 series has been met with a cool reception due to weakened computing power. Industry insiders confirmed that in the first quarter of this year, a large number of previously hoarded NVIDIA computing powers were released into the market, with prices dropping. The price of a server equipped with eight H800 cards dropped from a peak of 3.2 million yuan last year to about 2.8 to 2.9 million yuan. While the prices are still high, the supply situation is relatively loose. In comparison, there is virtually no customer inquiry for H20 products through unofficial channels.

Meanwhile, NVIDIA is facing intense competition from domestic chips in the Chinese market. Typically, Huawei represents the top strength of domestic chips. Companies like Baidu and iFlytek have publicly announced their purchase of a large number of Huawei's Ascend chips. It is understood that Huawei's Ascend chips still lag behind NVIDIA in terms of performance and ecosystem development. To enhance the competitiveness of their chips, Huawei is not only actively increasing shipments but also providing ample technical support to assist partners in adapting to the computing platform. Individuals close to Huawei have revealed that the Ascend 910B is not necessarily more cost-effective compared to the H20 and faces many technical limitations. Huawei's current strategy is to enhance the attractiveness of their products in the market by improving the cost-effectiveness of their overall solutions. The market demand for the Ascend 910B has not diminished, but Huawei is more focused on the actual performance of their products and seeks further optimization.

The head of a company's computing center mentioned that the market price for a server with eight H20 cards is about 1 million RMB, while a server with the same configuration of Ascend 910B chips could cost between 1.7 to 1.8 million RMB or even more, with the final price varying according to the scale of the order and how urgent it is. It has been reported that "Huawei's 910B is priced at about 120,000 RMB, and an eight-card server ranges between 1.3 to 1.5 million RMB," but this price has fallen behind the current market situation. After a new round of regulations, the price of Huawei's chips has significantly increased.

A technical leader at a large model startup company revealed that the market demand for the Ascend 910B is greater than its supply. Even though they are able to procure NVIDIA chips, they aim to also obtain Huawei's computational chips. This is mainly out of consideration for supply continuity, as solely relying on NVIDIA's computational power could cause investor concern.

Additionally, NVIDIA recently announced the launch of a new revolutionary GPU product series based on the Blackwell architecture, including the B200 and GB200, which greatly increase the computational power for training machine learning models. These new products are expected to officially hit the market in the second quarter of the year.

Every time NVIDIA introduces new products, it affects the market pricing of existing products. Industry insiders point out that although the new products may not be sold in the Chinese market, in comparison, Chinese enterprises will feel that the technical advantage of NVIDIA's current products becomes vague and that the cost-effectiveness of the H20 is reduced. In this context, NVIDIA is seeking to change.

NVIDIA's chips are produced by TSMC. Senior industry insiders close to TSMC revealed that due to limited production capacity, NVIDIA often prioritizes the needs of large customers in Europe and America. For example, even though NVIDIA's special edition H800 and A800 were ordered by Chinese enterprises last year, with the A800 launching in the fourth quarter of 2022 and the H800 in early 2023, Chinese customers usually don't receive the A800 until the middle of the year; even when the second round of chip bans was implemented in November 2023, many of the H800 chips ordered through legal channels had not been delivered yet. By contrast, the shipping speed of the H20 chips is evidently faster.

Market news says, H20 began taking pre-orders in the first quarter of this year, with delivery times as short as one month from order placement. Due to its larger memory capacity, the production cost of the H20 is higher than that of the H100.

Analysts had previously predicted that even if the demand for new China-specific versions of products is not high among Chinese companies, NVIDIA might not reduce their prices out of habit. Since NVIDIA products are in short supply, if the H20 is priced with limited profit margins after a reduction, NVIDIA would allocate its precious production capacity to other products with higher profit margins. Currently, AI companies from around the world are vying to purchase NVIDIA products.

According to UBS reports, NVIDIA has significantly shortened its delivery times, cutting the previous 8 to 11 months at the end of 2023 in half. However, NVIDIA CEO Jensen Huang recently mentioned that the company is still facing immense market demand pressure, with approximately 15,000 to 20,000 AI startups waiting to become NVIDIA customers. At the same time, technology giants are also increasing their purchases.

Data from Counterpoint Research shows that global cloud service providers' capital expenditures grew by 4% in 2023 and are expected to surge to 42% in 2024, mainly due to massive investments in artificial intelligence infrastructure by American tech giants. This means that the situation has changed differently from what industry insiders had previously anticipated.

In order to maintain its position in the Chinese market, NVIDIA is making unprecedented concessions. As the Chinese market gradually shrinks, the ecological advantage NVIDIA previously established in China is also diminishing. If the ecological advantage no longer exists, NVIDIA's possibility of returning to the Chinese market will be greatly affected.

SemiAnalysis, a semiconductor research institution, believes that NVIDIA can only secure enough orders in the Chinese market if it competes on price with local competitors like Huawei. IG Market analyst Hebe Chen points out that NVIDIA is walking a fine line—striving to find a balance between maintaining the Chinese market and dealing with the tense situation with the U.S. government.

When considering long-term strategies, NVIDIA must be prepared to deal with extremely adverse conditions. Recent financial reports reveal that in the third quarter of 2023, NVIDIA's revenue in mainland China reached an astonishing $4.03 billion, growing 47% quarter over quarter. However, with the implementation of new export control rules, revenue plummeted to $1.95 billion in the fourth quarter, a decrease of 52%. During this period, mainland China's contribution to NVIDIA's total revenue was only 8.8%, marking a historic low.

In the past, AI-related data center business accounted for a steady 20%-25% of revenue in China. Since October 2023, however, this revenue has significantly decreased. Although senior executives previously stated that the proportion of data center revenue from mainland China had dropped to single digits, this trend does not seem to have changed.

On May 23, 2023, NVIDIA released its latest quarterly financial report. According to the report, revenue reached an impressive $26.044 billion, a 262% increase year-over-year; net profit amounted to $14.88 billion, a staggering 628% surge compared to the previous year. Notably, the performance of the data center business broke historical records, with revenue for just one quarter hitting $22.6 billion.

Despite these results, information revealed by executives during the earnings call indicated that the revenue from data center business in mainland China continued to decline. Jensen Huang, speaking about the Chinese market, said: "Our business in China has seen a significant reduction compared to the past. The competition in China has become more intense due to technical challenges encountered. We will continue to do everything we can to serve our customers and the market there."

Regarding the current challenges NVIDIA faces in the computing chip market, we welcome readers to share their views and opinions in the comment section.