"CPUs (central processing units) are now one price a day." On April 18, in a computer city in Guangzhou, a dealer muttered while staring at the scrolling information on the mobile phone screen to give customers a quote, and the shopping mall radio repeatedly broadcast "Welcome to the Canton Fair merchants to come to ......

On the same day, Nvidia (NVDA.O) CEO Jensen Huang had just arrived in Shanghai to meet with Shanghai Mayor Gong Zheng. A day earlier, he was invited to visit Beijing for talks with Ren Hongbin, president of the China Council for the Promotion of International Trade. It wasn't until April 19 that Huang's trip to China ended.

On this trip, Huang made a rare appearance in a suit and removed his leather jacket. In the past half a month, the U.S. tariff policy has impacted the global semiconductor industry chain, and even Nvidia, the overlord of AI chips, has not been spared.

On April 15, local time, Nvidia said that the company recently received a notice from the U.S. government that H20 chips and chips that reach H20 memory bandwidth and interconnection bandwidth need to be licensed for export to China and other countries and regions. The Company expects first-quarter results to be approximately $5.5 billion in inventory, purchase commitments and related reserve charges related to H20 products.

What impact will the "tariff stick" of the United States have on the entire semiconductor industry chain? No one can give a definite answer at this point. An industry insider told Times Finance that there are still variables in this question, "It is very likely that it will be answered today, and new changes will appear in a few days." ”

However, the trend that can be seen is that the localization process in the semiconductor field is accelerating, and under the catalysis of DeepSeek, many local chip and large model manufacturers have carried out closer cooperation to jointly build a software ecosystem suitable for local chips.

Source: Picture Worm Creative

Navigating uncertainty

The tariff shock has added a lot of uncertainty to the semiconductor industry in the near future.

"Recently, the market will fluctuate, but [the price] is not too high." A stall owner in Guangzhou Computer City told Times Finance that the supply of chips in his hand is still sufficient, but he still advises customers who are willing to buy popular categories of chips to place orders as soon as possible, "Who knows if it will continue to rise in the future?" ”

When quoting, he looked through his mobile phone over and over again, repeatedly emphasizing that this price is only the price of the day, and it may change later, but it will not be too high.

A similar situation occurred at another shop that deals in computer assembly. "[U.S.] tariffs will be a little cheaper before they come." The owner of the store said that taking the assembly of the whole machine as an example, the current price is probably a few hundred yuan more expensive than before, which is not too much, but the price of the graphics card has been rising recently.

"The Gigabyte 4060 8G graphics card was only more than 2,100 yuan in the first year, and it has risen to more than 2,500 yuan now." Said the boss of the above-mentioned computer assembly.

On the second-hand platform, many people have listed Nvidia H20 modules/complete machines for sale. Times Finance learned that the price of the 41G version of the H20 machine is 1.2 million yuan, and the price of the server equipped with 8 H20 is 1.48 million yuan. "We've always had this price, and there has been no price increase." One of the sellers said. In addition, many sellers say that they have spot in their hands and can pick them up at any time.

Source: Screenshot from second-hand platform

Previously, some media reported that many popular chips such as CPUs and GPUs in the market have been suspended from quotation and trading, and there has also been a phenomenon of channel hoarding.

He Hui, research director of Omdia Semiconductor, also observed that there has been a wave of speculation in the market recently, and some people believe that AMD, Intel and other CPU products come from the United States, and there will be a risk of shortage or pressure from tariff increases in the future.

However, on April 11, the China Semiconductor Industry Association issued a notice, suggesting that the origin of the "integrated circuit" at the time of import declaration shall be subject to the location of the "wafer tape-out factory", regardless of whether it has been packaged or unpackaged.

"The semiconductor chip industry is highly globalized." He Hui said that although well-known semiconductor companies such as Qualcomm, Intel, AMD, and Nvidia are American companies, these companies mainly control the design of high-end chips. Once the design is complete, they entrust the chip manufacturing to the foundry, followed by packaging and testing, and finally delivery to the customer.

He Hui believes that over the past few years, the Asia-Pacific region has developed into a core region for semiconductor production. Most of the wafer production of these U.S. companies mainly relies on fabs in China (mainly mainland China and Taiwan), South Korea or Japan, while chip packaging and testing are mainly concentrated in Southeast Asia, Chinese mainland and Taiwan. Therefore, when the country of origin of chips is defined by the origin of the wafer, the vast majority of companies in this round can avoid tariffs.

"It's a 'seesaw effect.'" Wu Quan, president of Huaxin Jintong Semiconductor Industry Research Institute, explained to Times Finance that the semiconductor supply chain was not smooth enough, and as long as the policy was not loosened in the United States, relevant companies would face many restrictions and tests, which would affect the stability of the supply chain, and there was uncertainty in both the supply time and the supply of goods.

In Wu's view, the stockpiling is not entirely about the fear of supply disruptions, but more likely because the supply cycle has changed, and goods that could have been delivered in 10 days or a month may now need to be extended to two or even three months.

Huang looks forward to continuing to work together

Perhaps most chips can be used to avoid tariff risk through the origin of the wafer, but Nvidia's chips don't seem to have such good luck, and the United States has been tightening export controls on its AI chips.

Since 2022, NVIDIA's A100 and H100, A800 and H800 advanced AI chip GPU products have been included in the export control list by the United States. H20 is a chip developed by NVIDIA to reduce performance configuration in order to meet the performance red line of U.S. export controls.

In February, during Nvidia's fourth-quarter fiscal 2025 earnings call, Huang said he expected the revenue share of China's business in the first quarter of fiscal 2026 to be about the same as in the fourth quarter of fiscal 2025 and before, and about half of what it was before export controls compared to the previous quarter, but still at a certain level.

At that time, due to the popularity of DeepSeek, the market demand for low-cost artificial intelligence models surged, and the Financial Associated Press quoted Reuters as reporting that Chinese companies were increasing orders for Nvidia's H20 artificial intelligence chips.

On April 17, one day after Nvidia disclosed that H20 chips were notified of purchase restrictions, Huang came to China at the invitation of the China Council for the Promotion of International Trade.

Source: CCTV News

According to CCTV News, Huang said in a meeting with Ren Hongbin that China is a very important market for Nvidia and hopes to continue to cooperate with China.

According to CCTV Finance, in response to the U.S. government's decision to export H20 chips to Nvidia to China, Huang Jenxun said that the U.S. government's strengthening of chip export controls has had a significant impact on Nvidia's business, as a company that has been deeply involved in the Chinese market for 30 years, Nvidia and the Chinese market grow together and achieve each other, China is not only one of the world's largest consumer markets, but also an important driving force for their continuous innovation with its booming industrial ecology and leading software strength. NVIDIA will continue to spare no effort to optimize the product system that meets regulatory requirements and unswervingly serve the Chinese market.

Shanghai is Huang's second stop, where Nvidia has set up its R&D center. In a meeting with Gong, Huang said the Chinese market is important to Nvidia, not only because of its size, but also because of its high-quality customer base and good industrial ecosystem. He also said that NVIDIA complies with the laws of various countries, is committed to the Chinese market, and is committed to serving the Chinese market in the best possible way.

After 48 hours, Huang ended his trip to China. No one knows what impact this trip will have on Nvidia's subsequent development, but Nvidia is the focus of global AI chips, and Huang has to make many preparations.

On the U.S. side, Nvidia announced on April 14 that it is working with a number of supplier manufacturing plants to realize the localization of AI supercomputers in the United States in the next few years. Nvidia expects to produce "up to $500 billion" of AI infrastructure in the U.S. over the next four years, which is expected to create "hundreds of thousands" of jobs.

At present, the agency still believes that Nvidia may be greatly affected by the new export regulations of the US government. On April 17, Morgan Stanley analysts lowered Nvidia's revenue forecast, citing concerns that the impact of the latest U.S. chip export restrictions on China on Nvidia will be "more disruptive" than expected. Analysts said that as a result, they expect Nvidia's data center revenue to decline by 8% to 9% in the coming quarters.

The rise of domestic chips

Nvidia's H20 chip purchase is restricted, how big is the impact on domestic AI companies?

A person from a major model manufacturer told Times Finance that the inference speed of H20 is not fast and does not achieve the ideal effect similar to H100. Therefore, for large companies in the field of large models, the models can also run with the complete use of domestic chips.

"At the beginning, when the performance of domestic chips was low, we carried out simple adaptation and operation, and later through in-depth operator level adaptation, the performance can be doubled." The above-mentioned large model manufacturer said. However, he also admits that this kind of adaptive development work is indeed a big investment for small enterprises.

A large model manufacturer added that his company has already built a computing power cluster for large models. Since the CPU of the older model may fail, they will continue to purchase various components in small quantities to ensure the stable operation of the cluster. At the same time, they are also actively using domestic computing chips, such as Huawei Ascend and Cambrian, and building large-scale clusters around these chips. Even in the case of tariffs, the company was still able to maintain low computing costs.

Wu Quan said that the computing architecture is composed of software and hardware. From a purely hardware point of view, the capabilities of the H20 have been greatly "emasculated". The comprehensive performance of Nvidia H20 combined with CUDA platform still has certain advantages compared with similar domestic products. Domestic chip companies also need to pay equal attention to the development of software as much as hardware.

Source: Picture Worm Creative

In recent years, with the continuous progress of domestic technology and the strong support of policies, the domestic chip industry is also rising rapidly.

"DeepSeek reduces the demand for users' AI computing power through algorithm optimization, but at the same time, it also greatly lowers the threshold for users to use large models, creating more new market opportunities." IDC, a data research agency, said that by adapting to DeepSeek, China's local chips have achieved breakthroughs in the field of software ecology and gradually improved the software ecology. This provides strong support for the competitiveness of local chips in the market. At the same time, it also promoted the technical exchange and resource sharing of local chip manufacturers, breaking the deadlock in the ecological construction of domestic chips. Many local chip manufacturers have begun to cooperate around DeepSeek to jointly build software stacks, tool chains and other ecological components suitable for local chips.

Recently, HUAWEI CLOUD launched CloudMatrix 384 SuperNode and announced that it has been launched in the Wuhu data center. HUAWEI CLOUD has also continuously optimized the Ascend AI cloud service, and has been fully adapted to more than 160 third-party large models, including DeepSeek.

At the same time, the Cambrian (688256. SH) recently released its first quarter 2025 report, which also attracted market attention. Cambrian achieved operating income of 1.111 billion yuan in the first quarter, a year-on-year increase of 4230.22%, and the net profit attributable to shareholders of listed companies was 355 million yuan, compared with a loss of 227 million yuan in the same period last year.

"The time of CUDA ecological operation is relatively long, and the domestic ecology is also constantly advancing, but it needs a process, a process that promotes each other." The above-mentioned large model manufacturers said that there is no need to over-deify Nvidia, nor do they need to be overly anxious.

"At present, some of the customers we have contacted, especially the head customers, are willing to try to adapt and use a variety of domestic chips, accompanying us and the entire domestic chip ecosystem to grow together." He said.

Editor in charge: Zhu Gabei