China’s AI chips rush to switch to domestic

Artificial intelligence (AI)-related mega-companies in China are stepping up their efforts to switch their chips to Chinese production. The reason for this is that US-based NVIDIA products were mostly used in the past, but procurement has become difficult due to the US government’s export controls. While Huawei Technologies and others are attracting attention as alternative companies, they could negatively affect the development of China’s AI industry due to concerns regarding performance, among other things.

“Happy New Year!” In late January, a video of NVIDIA’s chief executive officer (CEO) Jensen Huang proliferated and circulated on Chinese SNS (social networking sites). China’s official media outlet Xinhua News Agency reported that Jensen Huang was the topic of conversation as he interacted with employees in Shanghai and elsewhere on the eve of the Chinese New Year holiday in mid-February, greeting them in Chinese and dancing in his national costume.

Behind the scenes of its joyfulness, in China, NVIDIA is facing a strengthening of the headwinds. NVIDIA previously held about 90% of the market share in China’s AI chip market. However, the U.S. Department of Commerce has initiated export controls on Chinese AI chips for fear of conversion to military use, and NVIDIA’s AI chips such as “H100” will be the subject of controls in the fall of 2022.

NVIDIA began exporting products with reduced performance to China in order to circumvent the control, but due to the new regulations in October 2023, the export of these chips was also banned.

The string of embargoes has affected major Chinese internet companies such as Alibaba Group, Tencent Holdings and Baidu as well. In order to develop and deliver generative AI, it is necessary to have AI chips on their own corporate servers, and the trend of securing stocks of NVIDIA-made chips strengthens in 2023.

NVIDIA plans to export chips to China that are compliant with the latest embargo, but at the moment, most major Chinese Internet companies are accelerating their shift to Chinese production. This is because the risk of using NVIDIA chips is increasing further as the US Department of Commerce is expected to continue to impose new controls in the future.

Tencent’s president, Martin Lau, told analysts in November 2023 that it had stockpiled enough NVIDIA chips to develop several more generations of its own general-purpose big models, and that the new chip restrictions would not affect the development of Tencent’s AI capabilities in the short term, but explained that it would be necessary to look for domestic Chinese suppliers in the future.

While some of NVIDIA’s products have bypassed controls and are thought to circulate within China, it is difficult to secure purchasing quantities from large corporations.

In the spotlight as an alternative is Huawei. Huawei has teamed up with Chinese chip giant SMIC (Semiconductor Manufacturing International Corporation) to push its “7-nanometer” high-performance chips into practical use and launched smartphones equipped with the chips in August last year, amid U.S. sanctions. Huawei is selling devices and services that utilize the AI chip “Ascend”.

Executives from major Chinese Internet companies said of Ascend, although the performance is not as good as NVIDIA’s products, we can use them without any problems.

Other Chinese chip companies also see NVIDIA’s inability to enter China as a business opportunity and are taking action. Haikuang Information Technology will begin supplying its new AI chip, Deep Computing II, in the fall of 2023. Hikari told investors in late 2023 that it was “working with Baidu and Alibaba to advance the construction of AI infrastructure”. Emerging Moore Threads Intelligent Technology also released its AI chip “MTTS4000” in late 2023, claiming that it is easy to transfer data from NVIDIA.

The localization of AI chips is highly anticipated, but on the other hand, the subject is mass production. The U.S. Department of Commerce has targeted a number of Chinese chip companies, including Haikuang, for export controls. These companies do not have their own production facilities, and deals with the world’s chip-making mega-corporations such as TSMC are being restricted.

The move to mass production also has a long way to go for Huawei. Charlie Chai, an analyst at Shanghai-based research firm 86Research, noted for Huawei’s Ascend that while the performance is similar to that of Nvidia’s products, the gap could widen in the future due to issues such as manufacturing processes. Some reports have suggested that the Ascend’s yield is just over 20 percent.

China’s AI industry will have a global market share of more than 30 percent by 2035, a Chinese government-owned research firm said in January, presenting a forecast. However, if the localization of chips stagnates, the technology gap with countries like the U.S. will widen and AI performance will fall behind. How far companies like Huawei can go with their technology development in the face of headwinds will shape China’s AI industry in the future.