Recent reports from US media indicate that the US export control measures on chips have impeded the progress of the new model by the Chinese artificial intelligence company DeepSeek.
According to sources cited by The Information on Thursday, due to a shortage of Nvidia server chips in the Chinese market, this will limit the speed and scale at which Chinese cloud service providers can offer services for DeepSeek’s new model.
These sources are employees of a major Chinese cloud service provider that offers the DeepSeek model to corporate clients. They noted that the upcoming release of DeepSeek’s large language model R2 in China may not reach the same level of popularity and scale as its predecessor R1.
DeepSeek relies on Nvidia’s hardware and software for training, and performs best when running on Nvidia chips. The sources also revealed that although Chinese domestic chip companies, including Huawei, offer alternatives to Nvidia chips, DeepSeek’s models are optimized based on Nvidia’s hardware and software. Therefore, running these models on Chinese chips would be more challenging to manage and less efficient.
Employees of Chinese cloud computing companies mentioned that if the performance of the R2 model surpasses existing open-source models, the subsequent demand for it could overwhelm Chinese cloud service providers who are already struggling with the shortage of Nvidia chips.
They stated that the majority of cloud customers currently using the R1 model are running it on Nvidia’s H20 chips.
Despite the uncertainty about the release date of R2, employees of cloud computing companies revealed that the company has been in close communication with some Chinese cloud computing companies to provide them with technical specifications to guide them on hosting and distributing the model on servers. These plans include preparing enough H20 chips for customers who wish to use the full version of R2.
The US ban in April shattered hopes of Chinese cloud computing companies stocking up on more H20 chips. Depending on China’s existing inventory of H20 chips may limit the extent to which enterprises can use R2.
The US ban highlights China’s continued heavy reliance on US technology. The H20 chip was released in 2023 by Nvidia in response to tightened export restrictions by the US, which effectively banned Nvidia from selling its powerful Hopper series of graphics processor chips to China. Subsequently, Nvidia customized a reduced version of the H20 chip for sale in the Chinese market.
Before the release of R1, some Chinese cloud computing companies and developers had been avoiding H20 because of its limited specifications, which were insufficient to run large language models (LLMs).
Following the release of R1, the landscape changed, with almost all Chinese enterprises – from cloud computing companies and tech startups to state-owned enterprises and government agencies – eager to use R1.
To meet the surge in demand for R1, tech giants including ByteDance, Alibaba, and Tencent Holdings ordered $16 billion worth of H20 chips in the first three months of 2025, equivalent to 1.2 million chips. In comparison, Nvidia shipped a total of 1 million H20 chips to China in 2024, according to SemiAnalysis estimates.
Due to US export controls, the chip challenge posed a significant hurdle to China earlier this year when the demand for R1 surged. To fill the gap, many domestic enterprises purchased Nvidia chips designed for gaming, such as the RTX 5090 and RTX 4090, to power R1. These chips are also restricted from sale in China but are easier to obtain on the black market than acquiring Hopper chips.
According to two sources, DeepSeek has yet to confirm a specific release time for the R2 model. DeepSeek engineers have been working diligently on developing R2 over the past few months, but CEO Liang Wenfeng is reportedly not satisfied with the new model’s performance. Engineers are striving to perfect the R2 model before Liang approves its release.
In January of this year, DeepSeek, headquartered in Hangzhou, caused a stir in the tech community, raising questions about the suitability of US high-end chips. The company claimed that its R1 inference model was comparable to leading US models but at a fraction of the cost.
US company OpenAI accused DeepSeek of using their AI models during the training of R1.
A senior official from the US State Department revealed in an interview with Reuters on Monday that DeepSeek is bypassing US export controls through alternative means to obtain advanced chips manufactured in the US.
The official stated that DeepSeek has a “significant amount” of high-end Nvidia H100 chips. Since 2022, the US has completely banned the export of H100 chips to China out of concern that Beijing might use these chips to enhance military capabilities or gain an advantage in AI competitions.
“DeepSeek is attempting to evade export controls through shell companies in Southeast Asia and is seeking to access US chips remotely by entering data centers in Southeast Asia,” the official said.