The US chip sanctions against China have escalated again
About 50% of the revenue of American chip manufacturers such as NVIDIA, Intel, and AMD comes from the data center chip market, and China is one of the largest markets for data center chips in the world
Text | Reporter Wu Junyu Gu Lingyu
Edit | Xie Lirong
On October 17, the U.S. Department of Commerce's Bureau of Industry and Security (BIS) updated the "Export Control Rules for Advanced Computing Chips and Semiconductor Manufacturing Equipment." This is a modification and enhancement of the October 7, 2022 rule. The rule is still in the public notice period and will take effect in 30 days. The U.S. Department of Commerce's Bureau of Industry and Security has also added two Chinese GPU (graphics processing chip) companies, Moore Thread, WallChat Technology and their subsidiaries to the Entity List.
The purpose of the move by the Commerce Department's Bureau of Industry and Security is to strengthen chip sanctions against China, and the reason given in the updated rules is to improve the effectiveness of sanctions, cut off the way to evade restrictions, and ensure U.S. national security.
The October 17 update of the "Export Control Rules for Advanced Chip and Semiconductor Manufacturing" mainly adds three new contents:
First, take performance density as an export control standard. Prohibit the sale of data center chips to China that run at speeds of more than 300 teraflops, or 300 trillion operations per second. Chips running at 150-300 teraflops will also be banned from sale if the performance density exceeds 370 gigaflops per square millimeter (billion calculations).
Second, the scope of advanced chip export licenses has been expanded to more than 40 countries. The aim is to prevent advanced AI chips from entering China from other countries.
Third, it has imposed licensing requirements for chip-making equipment in 21 countries and expanded the list of equipment prohibited from entering these countries. The move is aimed at limiting China's ability to manufacture advanced chips below 14 nanometers.
Source: U.S. Department of Commerce website
On October 16, Chinese Foreign Ministry spokesman Mao Ning said at a regular press conference: "China has repeatedly stated its position on US chip export controls to China, and we believe that the United States should stop politicizing, instrumentalizing and weaponizing economic and trade science and technology issues, and stop disrupting the stability of the global production and supply chain." China will pay close attention to relevant developments and resolutely safeguard its own rights and interests. ”
The U.S. Semiconductor Industry Association, in which NVIDIA (NASDAQ:NVDA), Intel (NASDAQ: INTC), AMD (NASDAQ: AMD) and others participate, said in an October 17 statement on the U.S. Department of Commerce Export Control Rules that too broad unilateral controls could hurt the U.S. semiconductor ecosystem because they encourage overseas customers to look elsewhere. The U.S. Semiconductor Industry Association also urged the U.S. government to strengthen coordination with allies to ensure a level playing field for all companies.
It is widely believed that the update of US government regulatory rules will have a limited impact on the industry in the short term, but in the medium and long term, the regulation will affect the speed of China's development in the AI competition. A U.S. lawyer, who asked not to be named, told reporters that the industry was assessing the scope of the bill's impact. The US control style has always been selective law enforcement, that is, although the provisions of the bill have fallen, the specific expansion depends on the scale of law enforcement, and this also involves more games at home and abroad.
A number of semiconductor practitioners interviewed at home and abroad revealed that before the introduction of the bill, there were rumors in the industry that the United States would upgrade its control measures, but the actual content of the bill still greatly exceeded the industry's expectations.
A senior industry source said that compared with last year's relatively vague regulatory content that used process nodes as the basis for restrictions, the pertinence and accuracy of this clause are clearer, which is a new change worth paying attention to. In addition to the transmission speed, such restrictions on high-end GPUs also increase the dimension of performance control, not only NVIDIA and other American companies to provide special version of chips to China will be restricted, but also means that the space for Chinese companies to use some other means to bypass US regulations is narrower, which reflects the determination of the United States to limit China in AI development.
Restrictions are tightened
Affected by the US Department of Commerce export control rules, such as NVIDIA's A100/H100, A800/H800, AMD's MI25 series chips, Intel's Gaudi2 series data centers will be prohibited from being sold to Chinese customers.
NVIDIA announced on October 17 that some of the company's chips exceeded certain performance thresholds (including but not limited to A100, A800, H100, H800, L40, L40S and RTX4090). Licensing requirements may affect a company's ability to complete product development in a timely manner, support existing customers, or provide services to affected regions. If a customer requires a license to regulate a product, the company may seek a license for the customer, but there is no guarantee that the U.S. government will grant it. Given the strong global demand for NVIDIA products, no meaningful short-term impact on financial results is expected.
The U.S. Department of Commerce's Bureau of Industry and Security rules for the regulation of advanced chip and semiconductor manufacturing were first promulgated on October 7, 2022, and at that time mainly included two contents: First, restrictions on the export of advanced chips to China. These include NVIDIA's A100/H100 series AI chips; Second, restrict the export of semiconductor manufacturing equipment to China. These include limiting China's ability to make logic chips below 14 nanometers.
After the promulgation of the above regulatory rules, NVIDIA's A100/H100 chips were banned from supplying the Chinese market. NVIDIA A100/H100 chips are currently the world's most powerful data center chips, which are designed for AI inference or training scenarios such as autonomous driving, high-end manufacturing, and medical and pharmaceutical applications.
The US chip companies represented by NVIDIA are not willing to lose their share of the Chinese market, and have launched AI chips A800/H800 that reduce the speed of network interconnection for the Chinese market according to the export control rules of the US Department of Commerce. Its peak performance is in line with the A100/H100. Due to the lower memory bandwidth, the A800/H800 will take significantly longer to train large AI models. However, the A800/H800 is still recognized in the Chinese market as the best data center chip for AI training/inference.
In 2023, AI large model technology will become popular, and the market demand for NVIDIA computing power chips will rise sharply. In the United States, American companies including Microsoft, Tesla, Orcale, etc. are forming data center clusters with 10,000/H100 chips.
In China, cloud computing manufacturers such as Ali, ByteDance, Tencent, and Baidu, ICT solution providers such as Lenovo and Xinhua3, and Xpeng are also purchasing A800/H800 series chips on a large scale. A800/H800 is currently in short supply, and the reporter learned from a number of server manufacturers that NVIDIA's A800/H800 chip orders have been scheduled to the middle of 2024.
In August this year, authoritative media (the British Financial Times) reported that Chinese companies, including Alibaba, Tencent, ByteDance, Baidu, etc., had ordered a total of $5 billion in A800 chips, which are expected to be delivered in 2023-2024. If the unit price of an A800 chip is 100,000 yuan (RMB), the demand for A800 chips in the Chinese market exceeds 50,000.
A chip technician commented that the latest version of the US Department of Commerce's export control rules, chip performance density provisions are particularly stringent. This has blocked the possibility that U.S. chip companies will use other technology strategies to provide special products for the Chinese market. He believes that if only peak performance is limited, chip companies can also launch small volume products with low performance and power consumption for the Chinese market, and then form clusters through high-speed connections. The result is the expected cluster performance. But the terms of chip performance density block this path.
A semiconductor investor told reporters that after the upgrade of the regulation, it will become more difficult for chips within the scope of control to enter the Chinese market in the future, and the price of these chips will also rise. The reporter learned that the legal profession generally believes that this is an important change in the logic of restrictions.
Domestic chip companies have limited impact
Although domestic Huawei Ascend, Moore Thread, and Wall Technology announced the launch of AI chips whose theoretical performance can be compared to NVIDIA A100, the operating performance in actual scenarios remains to be seen. Even so, Moore Thread and Wall Technology were still included in the Entity List by the US Department of Commerce this time.
For Walltech and Moore's thread, being on the Entity List means that if you can't get a license, you can't not only import U.S. technology or products, but also can't use fabs based on U.S. technology or equipment to foundry chips for them.
On the evening of October 17, the Moore thread issued a statement saying that it was concerned about the news that the US Department of Commerce included the Moore thread in the "entity list" on October 17, local time, and strongly protested against this. Moore Thread said that the company is actively communicating with all parties and the impact on the matter is being evaluated. Bicheng Technology also said that the company expressed strong opposition to the US Department of Commerce's move, and will actively complain to the relevant US government departments, and call on the US government to re-examine.
Founded in October 2020, Moore Thread is a chip company focusing on full-featured GPU chip design. A person from Moore Thread told reporters that this morning, the company has held an all-hands meeting on the matter. The message from management internally is that despite the impact of this matter, the overall impact on the company's operations is still manageable. At present, in addition to high-end GPUs, Moore Thread is also increasing investment in new product lines that can generate revenue faster.
Founded in 2019, Bicheng Technology is a star company of domestic high-end GPU chips. In the past four years, it has obtained more than 5 billion yuan in financing, and investors include Qiming Venture Capital, IDG Capital, Hillhouse Venture Capital, etc. In March 2022, the valuation has grown to a staggering 17.2 billion.
The reporter learned that Bicheng Technology is currently preparing to land on the Hong Kong stock market. An FA (venture capital financial advisor) person close to Bicheng Technology told reporters that the current chip mass production and revenue of Bicheng Technology are not ideal, and the speed of business development is not as expected, and the reason why it can be listed on the Hong Kong stock market is because the Hong Kong Stock Exchange officially implemented the listing rules "Chapter 18C" for professional technology companies this year. This lowers the barrier to listing for hard technology companies.
The reporter learned that in order to attract a new generation of information technology, advanced hardware and other specific industry companies to list in Hong Kong, the Hong Kong Stock Exchange has lowered the revenue threshold that enterprises need to meet for listing, and paid more attention to the valuation of enterprises and the proportion of R&D investment. In other words, technology companies that are not profitable but have potential can also go public and raise funds.
For GPU companies, chip research and development is only the first step, the next supply chain supporting and ecological supporting is an important factor affecting product commercialization, a senior semiconductor investor believes that the domestic track around 2020 has ushered in the climax of private enterprises, but the development to the present, can produce products are few, let alone benchmark NVIDIA. From this point of view, the impact of this restriction upgrade on these companies is limited, even less than the impact of the ebb of the capital bubble.
U.S. chip companies raised concerns
The US Department of Commerce has always adopted a "license" strategy for chip sanctions against China - upstream and downstream enterprises in the US chip industry chain need to pass a license issued by the US Department of Commerce's Bureau of Industry and Security to sell products to China.
In other words, when the US government regulates the export of American chip companies, American companies will also suffer from performance. At present, about 50% of the revenue of chip manufacturers such as NVIDIA, Intel, and AMD comes from data center chips. The largest markets for data center chips are the United States and China.
The Chinese market has long been the largest market for U.S. chip companies outside the United States. The reporter counted that in 2022, the three main data center chips of NVIDIA, Intel and AMD accounted for 21.4%, 27.2% and 22.1% of the revenue in Chinese mainland (including Hong Kong), respectively.
The real revenue contribution of the entire Chinese mainland market is much greater than the financial data. A number of server manufacturers told reporters that PC companies in Taiwan, such as Asus, Gigabyte, MSI, etc., integrate NVIDIA's chips, and a large number of end customers are still Chinese mainland. According to the total statistics of the two markets of Chinese mainland and Taiwan, the revenue of NVIDIA, Intel and AMD in the Chinese market accounted for 47.3%, 40.3% and 32.1% respectively.
The US government's sanctions against China's chip field have triggered concerns about US chip companies. The Semiconductor Industry Association (SIA) accounts for 99% of the U.S. semiconductor industry and nearly two-thirds of the world's chip companies. Its members include Intel, NVIDIA, Qualcomm, AMD, Ampere and dozens of other companies. The U.S. Semiconductor Industry Association issued a statement on October 17 on the U.S. Department of Commerce's Bureau of Industry and Security export control rules, saying it is evaluating the impact of updated export controls on the U.S. semiconductor industry. The statement content displays:
"We recognize the need to protect national security and believe that maintaining a healthy U.S. semiconductor industry is an important part of achieving this goal." Overly broad unilateral controls could hurt the U.S. semiconductor ecosystem without promoting national security because they encourage overseas customers to look elsewhere. Therefore, we urge the government to strengthen coordination with allies to ensure a level playing field for all companies. ”
At the end of August this year, NVIDIA management bluntly said in an investor conference call in the second quarter of fiscal 2024 (early May to late July 2023) that the Chinese market demand accounted for 20%-25% of NVIDIA's data center revenue. In the long run, the ban on the sale of advanced data center chips to China, once implemented, would cause the United States to lose the opportunity to compete and lead in one of the world's largest markets.
By the close of trading at 4 p.m. ET, Nvidia shares were down 3.2 percent to close at $446.0 with a total market value of $1,085.27 billion. Intel shares fell 1.4% to close at $36.1, with a total market value of $151.02 billion. AMD shares fell 1.2% to close at $105.1 with a total market capitalization of $169.87 billion.
Responsible Editor | Zhang Yufei
Title map | Visual China