China’s top internet companies have stockpiled billions of USD worth of Nvidia’s H20 artificial intelligence chips this year before the US cut off shipments of the components in April, Nikkei Asia has learned.
The H20 graphic processing unit (GPU) was specifically designed for sale in the Chinese market to comply with US export controls, but ByteDance, Alibaba, and Tencent began preparing last year for the possibility that Washington could curb their shipments, too, sources said.
The three companies had asked Nvidia to ship a total of around one million H20 chips—roughly a full year’s supply—as soon as possible, ideally by the end of May, one person briefed on the matter told Nikkei Asia. The actual number delivered fell short due to the Trump administration announcing in early April that these chips would need a license for export, the source added.
Another source said the value of those rush orders exceeded USD 12 billion, and that several billion USD worth of chips were shipped before the new curbs took effect.
ByteDance was among the most aggressive in snapping up as many Nvidia chips as possible, according to two of the sources.
These stockpiling efforts came amid a surge in demand for AI computing power in China, particularly since the emergence of DeepSeek early this year. In February, Tencent began integrating DeepSeek into its superapp WeChat, greatly boosting computing power demand.
“The H20 curb didn’t come as a surprise because it was widely anticipated across the industry,” a leading Chinese tech company executive said. “Every major Chinese tech company had been stockpiling H20 in advance. After all, it wasn’t banned at the time, and given its strong performance, why not?”
In addition to rush orders, Chinese companies have been looking into ways to buy Nvidia chips outside of China, where they are not subject to US export controls, sources said. Other efforts to secure access to AI hardware include exploring ways to set up subsidiaries or affiliates overseas or collaborating with industry partners such as telecom operators, multiple industry sources told Nikkei Asia.
“The [Chinese] clients are very calm,” said an executive with a supplier to ByteDance and Alibaba Cloud that met with the companies after the latest restrictions. “They knew it was coming and they have been prepared for this day. They told us that their aggressive goal to build more data centers this year remains unchanged.”
Besides mainland China and Hong Kong, Alibaba runs data centers in 13 countries, including two in the US, while ByteDance has data centers in various Southeast Asian and European countries, including Ireland and Norway.
Meanwhile, Chinese data center giants are speeding up the verification process for domestic GPU platforms, such as Huawei’s Ascend, according to multiple people with knowledge of the matter. In early April, Huawei unveiled its latest AI computing solution, the CloudMatrix 384, which connects 384 of its self-developed Ascend AI chips to rival the performance of Nvidia’s advanced GB200 NVL72.
The H20 is the twice-downgraded version of Nvidia’s H100 chip, which first hit global markets in the third quarter of 2022. In response to US export controls that year, Nvidia introduced the lower-powered H800 for the Chinese market. Nvidia later downgraded this to the H20 after the US further clamped down on AI hardware exports to China.
The H20, launched in the first half of 2024, is roughly one-tenth of the computing power of the original H100 for AI training and only 20% of its inference capability. However, it remains very popular in China.
Inference refers to an AI application making predictions or generating responses based on a trained model.
Eugene Lee, a Hong Kong-based AI engineer, said that while the H20 might appear to be a training GPU, its configuration is actually optimized for inference, in contrast with the high-performance training capabilities of Nvidia’s H100 and H800 chips. Thus, if supplies of the H20 run short, it can be partially substituted by domestic alternatives or cloud-based solutions for small and midsize models. However, the deployment and ongoing optimization of large models in such a situation could be significantly constrained.
Lee added that many large-scale AI models rely on previously procured H100s and H800s for training, with major cloud providers also using them to deliver inference services.
“If existing inventories of the H100 and H800 are depleted, it could seriously hinder the training of advanced models and the development of next-generation systems, posing a substantial threat to China’s competitiveness in high-end AI development,” Lee said.
Nvidia warned that the restrictions on the H20 will benefit its competitors as Chinese customers seek alternatives domestically or elsewhere. On April 15, the company said it expects to take a USD 5.5 billion quarterly hit from the curbs.
Following that announcement, Nvidia CEO Jensen Huang made an unexpected visit to Beijing, where he told officials he would make “every effort” to keep serving China.
In the fiscal year that ended on January 26, China accounted for about 13.1% of Nvidia’s total revenue, down from nearly 17% a year earlier, the company’s filing showed. Singapore’s share, by contrast, rose to 18% from 11.2% the previous year. Those numbers are based on customer billing locations, however, and Nvidia noted that many customers use Singapore to centralize invoicing even though products are typically shipped elsewhere. By actual shipments, Singapore made up less than 2% of its total revenue for that period.
Nvidia declined to comment. ByteDance, Alibaba, and Tencent did not respond to requests for comments.
This article first appeared on Nikkei Asia. It has been republished here as part of 36Kr’s ongoing partnership with Nikkei.