From triggering a sudden plunge in Nvidia's stock price to driving the hot sales of its H20 chips, the relationship between Nvidia and DeepSeek, a top player in the technology circle, has become increasingly close. "DeepSeek is an outstanding innovation, and the computing demands brought by inference are huge." At NVIDIA's latest quarterly financial report, Huang Renxun praised DeepSeek.

This is indeed the case. Thanks to its data center business, Nvidia delivered an outstanding performance in the fourth fiscal quarter of 2025. The person in charge of Berlin Cloud, a computing resource information sharing platform, told a reporter from the Financial Associated Press that the demand for H20 has changed significantly recently, and the number of inquiries has tripled from a year ago. The main advantage is that the deployment of the DeepSeek full-blooded version is cost-effective. At the same time, the person in charge showed the reporter of the Financial Associated Press that he had established nearly 10 H20 demand docking WeChat groups in recent weeks.

(Photo provided by interviewee)

Reporters from Cailian News learned from multiple interviews that due to cost-effectiveness and compliance considerations, H20, a GPU that was once not favored by the market, has become "the favorite of major manufacturers." The current price of the 8-card H20 machine has increased by RMB 100,000 compared with a year ago, and some practitioners even predict that "the price will not come down."

TrendForce analyst Gong Mingde said in an interview with a reporter from Cailian News that from the perspective of the AI ​​server market, the current main GPU supply in the Chinese market is still dominated by NVIDIA (H20).

In fact, last week, a reporter from the Financial Associated Press conducted multiple interviews on the supply and demand situation of the AI ​​computing power market under the DeepSeek boom and learned that since the start of construction two weeks later, the number of inquiries for leasing and purchasing computing power equipment has increased significantly, and demand is still concentrated on NVIDIA H series computing cards. In addition, 4090 and A100 are also popular models.

Specific to the choice of the above-mentioned popular card counting cards, an NVIDIA card counting dealer told a reporter from China Finance News, "H20 is the mainstream, followed by A100, followed by cost performance. 4090 is a 'replacement card' (referring to consumer cards used for professional computing). Overall, it is not suitable for enterprises." Another dealer said bluntly, "After DeepSeek became popular, the demand for the H20 increased the most."

In terms of price, a reporter from Cailian News learned from a dealer that the H20 series has the largest increase in price. "The current unit price of an 8-card machine is about 1.1 million yuan, which is an increase of more than 100,000 yuan from a year ago."

"The price should not come down because major manufacturers need to do DeepSeek inference and place large orders for H20. Tencent alone costs 100,000 to 200,000 cards," the person said. He believes that due to cost-effectiveness and compliance considerations, H20 is already "the favorite of major manufacturers."

Just this week, some media quoted people familiar with the matter as saying that H20 orders from Tencent, Alibaba, and ByteDance have increased significantly. The above-mentioned dealer also confirmed to a reporter from Cailian News that the demand for H20 from major manufacturers is indeed strong due to the DeepSeek craze.

Although the growth rate of H200 is not as obvious as that of H20, it may still be Nvidia's current "leading star". The above-mentioned dealer said that the demand for H200 has not changed much in recent months. The current spot price is around 2.3 million yuan, and the futures time is about one month, which is similar to a year ago. A person from a server company admitted to reporters, "After DeepSeek came out, large models have evolved from training to inference, and the cost of computing power has dropped significantly. There is even news that many previous computing power projects have to be postponed because after equipped with DeepSeek, such strong computing power is no longer needed."

However, according to a customer demand form provided by the person in charge of Berlin Cloud to a reporter from the Financial Associated Press, the customer demands it came into contact with were mainly focused on Nvidia H200.

A source from a server company told reporters frankly: "After DeepSeek came out, large models have developed from training to inference, and the cost of computing power has dropped significantly. There is even news that many previous computing power projects have to be postponed because after equipped with DeepSeek, such strong computing power is no longer needed."

However, when the person in charge of Berlin Cloud was asked by a reporter from the Financial Associated Press whether the demand came from large manufacturers, he admitted that although the demand for H20 is strong, the demand parties he comes into contact with generally do not disclose specific information such as company names. "Intelligent computing projects are private, and there are too many middlemen. Demand parties also hope to find first-hand sources of supply more effectively, and most of them inquire first and then compare."

The person in charge of Berlin Cloud further stated that how to better match supply and demand is also key. For example: "Many demanders require machines to be deployed in designated computer rooms, but lessors already have stable computer rooms and operation and maintenance undertakings, and it is difficult to move machines to other people's designated computer rooms for deployment."

Compared with Nvidia's current popularity, one practitioner said that domestic cards have not yet reached the point of being "robbed".

However, Nvidia is not without pressure.

Wu Quan, founding partner of Huaxin Jintong Semiconductor Capital, told reporters from the Associated Press that DeepSeek has greatly enhanced local technological confidence, will upgrade the underlying elements of domestic economic and social development, and greatly promote the effectiveness of new productivity. Large models and AI-related applications in various industries are showing explosive momentum. "It also brings opportunities and uses to the entire computing industry chain, including domestic chips. As a key component and link, I believe that domestic cards have the confidence and ability to make new breakthroughs."

"In terms of AI chips, GPUs and ASICs will coexist in the future market. In addition to using NVIDIA GPUs, cloud companies are also actively developing their own ASIC chips. In addition to customizing their own applications, they can also reduce their dependence on NVIDIA chips and reduce expenditure costs." said Qiu Peiwen, an analyst at TrendForce Consulting. At present, many domestic computing power manufacturers such as Huawei Ascend, Kunlun Core, and Cambricon have adopted ASIC solutions.

Gong Mingde also observed that some large Internet companies and chip design companies are also actively promoting independently developed AI chip solutions. He predicts that in the next three to five years, under the continued influence of factors such as changes in the international situation and China's encouragement of domestically produced chips, the trend of AI chip autonomy will gradually increase, especially in the field of edge AI (EdgeAI) applications in various industries.