Xi Jinping has reiterated that he wants AI self-sufficiency, but China continues to be in dire need of quality chips (which it will not easily have) - Analysis

The U.S. blockade of the flow of the most advanced microprocessors to China-decided in 2022 by the Biden administration and then amplified by Trump-has so far succeeded in slowing and complicating Beijing's technological rise.

GettyImages

Although there is no shortage of voices to the contrary, this is at least what many experts claim, who argue that export controls will have an even greater impact in the future for two reasons: first, because U.S. companies will benefit from Nvidia's next-generation chips; and second, because China is likely to remain stuck at the 7nm technology node, at least until it develops its own EUV lithography capability.

But net of this, Chinese AI continues to advance.

It was especially Huawei's recent chip developments-with the new Ascend 910D processor-that showed an acceleration toward Beijing's technological self-sufficiency, particularly in the AI sector.

On the other hand, China's successes in terms of heavy government investment, smuggling of chips, exploitation of gaps in U.S. export control coverage, completion of equipment transfers within the country, recruitment of expert talent by leading international companies in reverse-engineering foreign technologies, exploitation of state-backed economic espionage, and production of genuine domestic innovation represent a formidable combination for pursuing the goal of AI “self-sufficiency and self-empowerment” reiterated (also) recently by Xi Jinping.

In any case, for all that has happened in the past year, there is no doubt that China has narrowed the AI development gap with the United States. It was especially the Chinese artificial intelligence startup DeepSeek that attracted the world's attention, launching in January an AI-based reasoning model that it said was trained with less advanced chips and was cheaper to develop than its Western rivals.

Beijing has also made progress in infrastructure software engineering, but DeepSeek's announcement especially challenged the assumption that U.S. sanctions were holding back China's artificial intelligence sector in the midst of a strong geopolitical technological rivalry and that China was lagging behind the United States after OpenAI's groundbreaking ChatGPT launch in late 2022.

Despite these restrictions, China's Yangtze Memory Technology Corporation (YMTC) has managed to produce high-density memory chips comparable to those produced by its Korean rivals.

China is also innovating on how to use less efficient chips more effectively. Earlier this year, a Chinese research team won an award at a prestigious international conference for just that: using less powerful chips to outperform high-end hardware.

“We must continue to strengthen basic research, focus our efforts on mastering key technologies, such as high-end chips and basic software, and build an independent, controllable and collaborative artificial intelligence hardware and basic software system,” Xi said, adding that artificial intelligence regulations and laws should be accelerated to create a 'risk warning and emergency response system to ensure that artificial intelligence is safe, reliable and controllable.

During a Feb. 17 meeting between CCP Chairman Xi Jinping and Chinese technology executives, including DeepSeek CEO Liang Wenfeng, Huawei founder Ren Zhengfei told Xi that his earlier concerns about the lack of domestic production of advanced semiconductors and the harmful impact of U.S. export controls have been alleviated thanks to recent breakthroughs by Huawei and its partners. Ren also said he heads a network of more than 2,000 Chinese companies that are working collectively to ensure that China achieves more than 70 percent self-sufficiency in the entire semiconductor value chain by 2028.

And it is likely that within the next year or two this will happen. Indeed, Beijing will be able to produce millions of microprocessors, such as Huawei's Ascend for artificial intelligence.

But the issue is not just quantitative, for it is mainly on the quality side that China's biggest problems lie. These chips, in fact, have significantly lower performance than Nvidia's in training advanced AI models and are, in addition, supported by a much weaker software ecosystem, with many complex problems that are likely to take years to solve.

Achievements in AI chip production

Although Beijing is not yet capable of producing the most advanced chips like Nvidia's, there is no doubt that its AI chip makers have achieved milestones that seemed unimaginable just a few years ago.

During a recent closed-door meeting with U.S. lawmakers, Nvidia CEO Jensen Huang expressed concern about Huawei Technologies' advances in artificial intelligence. The discussion with the U.S. House of Representatives Foreign Affairs Committee focused on Huawei's AI chips and the potential effects of U.S. restrictions on Nvidia chips in China.

Huawei is reportedly preparing to launch its most powerful AI processor, the Ascend 910D, with the goal of overtaking Nvidia's H100. Several Chinese technology companies have been contacted to test its technical feasibility, and the first units are expected to be available by the end of May.

Huawei's focus is clearly on developing chips for artificial intelligence, a rapidly growing sector in China and considered crucial to its technological independence. At present, the Ascend product line focuses on the Ascend 910B and Ascend 910C, the latter of which has two Ascend 910B logic arrays per integrated unit (meaning that more space is needed on each silicon wafer to produce an Ascend 910C than a 910B).

Huawei plans to begin massive shipment of its advanced 910C AI chip to Chinese customers as early as May 2025. The company also impressed analysts with the new CloudMatrix chip cluster it began delivering in April. It is a rack-scale artificial intelligence system consisting of 384 Ascend 910C processors arranged in an all-optical, all-to-all mesh network. The system spans 16 racks, including 12 processing racks with 32 accelerators each and four network racks supporting high-bandwidth interconnections using 6,912 LPO 800G optical transceivers.

For some, CloudMatrix would outperform Nvidia's popular NVL 72 cluster, although it consumes more power.

The Kirin 9000S chip-manufactured by China's SMIC using a 7-nanometer process and found in the Mate 60 series smartphones-has also shown significant progress, although it lags slightly behind industry leaders such as Intel and TSMC.

The recent alliance between Huawei (artificial intelligence chip designer), SMIC (artificial intelligence chip maker) and CXMT/XMC (high-bandwidth memory makers), aimed at developing its own artificial intelligence chip ecosystem, is also significant in the logic of increasing domestic chip production.

Quality limits

At the moment, it is mainly the performance of Chinese microprocessors for AI that worries local manufacturers, because it is considered not yet up to par with those of competitors.

DeepSeek, for example, evaluated Huawei's Ascend chips and found that they are not suitable for AI model training: each Ascend 910C offers about 60 percent of the performance of an Nvidia H100 for AI model inference.

This is a significant gap, given that in the coming years an increasing part of the processing requirements for advanced AI models will be devoted to inference, the stage where the model is able to apply what it has learned from training to concrete situations.

Barclays, an investment bank, estimates that by 2026, 70 percent of computing demand for artificial intelligence will come from inference.

In 2024, Nvidia's CEO estimated that the use of its chips will consist of “40 percent inference and 60 percent training,” up from 90 percent training in 2016.

Chinese Vulnerabilities

Despite indigenous chip production, China's supply chain continues to have significant vulnerabilities, particularly:

• in the design of AI chips;

• In the production of advanced node logic chips;

• in the production of high-bandwidth memory (HBM) in advanced nodes.

AI chip design

There are many companies working on AI chip design in China, including Huawei, Cambricon, and Biren.

Huawei holds a prominent position with its Ascend line of chips for AI, but its design and manufacturing strength is not all domestic.

For this reason, some allege that the Chinese company circumvented U.S. export controls by obtaining high-end chips from Taiwan. According to the allegations, Taiwan Semiconductor Manufacturing Company (TSMC), in particular, allegedly produced large quantities of Huawei Ascend 910B chips on behalf of Huawei's shell companies and then shipped them to China in violation of U.S. regulations (Huawei and TSMC, however, deny all allegations).

According to some sources, Huawei acquired more than 2 million artificial intelligence chip arrays in this way, also allowing it to stockpile HBM (the high-bandwidth memories built into processors) for more than a year.

According to consulting firm SemiAnalysis, even Huawei's latest AI processor, the Ascend 910 C, still contains many components supplied by foreign companies.

In addition to Huawei, Cambricon Technologies, a publicly traded and partly state-owned Chinese technology company, has reportedly already delivered the chip intended to replace Nvidia's A100 to its customers.

Hygon, another local chip maker, has just finished testing an alternative to Nvidia and is expected to begin deliveries in the coming months.

The production of advanced node logic chips

China has no domestic alternative to the state-of-the-art lithography tools produced by the Dutch company ASML.

Currently, the most advanced logic chip manufacturer is SMIC. The SN2 plant in Shanghai is the only one with an active 7-nm logic chip production line, in operation since July 2021, which is more than a year before the first tranche of the Biden administration's semiconductor equipment export controls went into effect.

SMIC and Huawei are, however, working to bring a 5-nm node into large-scale production, but they must do so without access to EUV lithography equipment because, as mentioned, Beijing has no local manufacturer of such machines.

In addition, export controls have prevented such machines from being exported to China.

However, according to U.S. think tank CSIS, the bottleneck in the expansion of production capacity to 7 nm (which is referred to as “N+2” in SMIC's node naming system) is not so much lithography as it is U.S. etching, deposition, inspection and metrology equipment, the shortage of which has been a major factor preventing SMIC from improving its yield, even though the company is expected to be able to reach 50,000 WPM at 7 nm by the end of 2025.

Those working to produce a domestic alternative to ASML's EUV technology is Huawei. In fact, the company currently has two research and development centers for semiconductor manufacturing tools, one in Shanghai and one in Shenzhen. According to a report, Huawei is investing $1.66 billion in the Shanghai facility alone and has hired a large number of chip industry veterans with experience in companies such as Applied Materials, Lam Research, KLA, ASML, TSMC, Intel and Micron.

The production of high-bandwidth memory (HBM) in an advanced node

China is also making progress in the area of high-bandwidth memory (HBM) units integrated into AI processors such as those from Nvidia. The market is currently dominated by South Korean SK Hynix.

Other major manufacturers include Samsung, another South Korean company, and the American Micron. In December, the United States began restricting HBM sales in China, but CXMT, a Hefei-based company founded in 2016, is believed to be rapidly catching up in this technology.

The lack of compatible artificial intelligence software

Another problem for Chinese manufacturers is the software used by developers to program AI chips.

Huawei's Ascend chips, in particular, continue to face problems related to the lack of compatible AI software, causing low utilization of the purchased chips.

CUDA, Nvidia's platform, is the one used by almost all AI programmers. It allows developers to harness the power of GPUs to accelerate applications, offering superior performance in various fields such as image processing, deep learning, numerical analysis, and computational science.

However, CUDA has one limitation: it only works with Nvidia chips, so those who leave these chips must leave the CUDA ecosystem, which requires solving many extremely complex software problems (for which CUDA already provides free solutions).

Huawei has created a replacement for CUDA, called CANN, that programmers can use for its Ascend chips. However, the software is years behind Nvidia's and is full of bugs.

According to a September 2024 report, even Huawei employees found the product “difficult and unstable to use” and prone to frequent crashes. Industry sources reported that DeepSeek's latest evaluation of CANN was very negative.

The situation, however, could change because Huawei recently joined the PyTorch open source foundation in an effort to increase Ascend and CANN compatibility of the PyTorch AI development framework.

The situation, however, may be changing as Huawei recently joined with the PyTorch open source foundation in an effort to increase the Ascend and CANN compatibility of the PyTorch AI development framework.

A recent analysis concludes that the current shortage of computing resources in China is due more to practical challenges and suboptimal implementation-many unused small and medium-sized data centers rather than a small number of fully utilized superclusters-than an absolute shortage of chips.

For CSIS, a combination of recent changes in Chinese government policies, such as the “divestment of unused computing resources to cloud providers” and strategic moves by major Chinese technology companies to favor DeepSeek models even over domestic ones, “suggests that the situation may change in the near future.”

The difficulties of decoupling

Because the global semiconductor industry is transnational, decoupling manufacturing remains very complex.

In fact, chip manufacturing relies on specialized and complex, globally distributed supply chains involving many stages of production, including specialized materials, production equipment, design and related software, fabrication, testing and packaging.

Taiwan, South Korea, the United States, Japan, Singapore and China are the main chip-producing countries. There are also major facilities (often subsidiaries of a leading manufacturer) in Europe, Southeast Asia, South America, and Israel.

Many segments of the supply chain are highly concentrated, with one or a few suppliers dominating a particular process or area of interest. One of the most visible areas is photolithography equipment, where only the Dutch ASML supplies EUV equipment and the top three suppliers (ASML, Nikon, Canon) account for virtually all of the total market share. Eighty percent of the foundries' market share is in Asia, almost all of it in Taiwan. The latter is home to the world's most advanced semiconductor manufacturer, Taiwan Semiconductor Manufacturing Company (TSMC), and 20 percent of global semiconductor manufacturing capacity.

U.S. companies such as Qualcomm and Nvidia, although they specialize in advanced semiconductor design, outsource production to third-party foundries, including those owned by TSMC. This is because the United States lacks state-of-the-art manufacturing capacity. According to the Semiconductor Industry Association, the U.S. share of global semiconductor manufacturing has declined from 37 percent in 1990 to 12 percent in 2020.

In summary, the global semiconductor value chain is transnational. It is based on a high division of labor between companies and regions and is defined by strong interdependencies and various bottlenecks at the level of production steps, suppliers and chip types. In addition, no one region can control all the production steps and supplies needed for state-of-the-art semiconductor manufacturing.