The latest news indicates that the Chinese artificial intelligence startup DeepSeek encountered technical malfunctions while using Huawei chips for training, leading to a delay in the release of its latest R2 model. This incident highlights the serious challenges Beijing faces in advancing localization of AI development.
According to reports from the Financial Times, after releasing the R1 model in January this year, DeepSeek was “encouraged” by regulatory authorities to switch to Huawei’s Ascend processor instead of Nvidia’s system.
However, during the training process for the R2 model, stability and performance issues with the Ascend chips persisted, forcing DeepSeek to adopt a compromise of “training using Nvidia, inference using Huawei.” The planned release in May was delayed, missing the market opportunity.
Training involves allowing the model to learn and optimize parameters through large amounts of data; while inference involves using the trained model to generate results or predictions, such as replies from chatbots.
Huawei had deployed an engineering team to DeepSeek’s office to assist in development, but they were unable to successfully complete a full training session on the Ascend chips. DeepSeek continues to work with Huawei to try to fully adapt the R2 model to the Ascend chips for the inference stage.
Sources indicated that DeepSeek’s founder, Liang Wenfeng, was dissatisfied with the progress of R2 research and development, and is pushing for increased resource investment to create advanced models that maintain a leading edge.
Nvidia recently reached an agreement with the U.S. government to pay 15% of its H20 chip sales revenue in China to the U.S. in exchange for export licensing. The company emphasized that a complete withdrawal from the Chinese market and developer community would weaken America’s global economic competitiveness and technological influence.
However, the Cyberspace Administration of China summoned Nvidia last month to explain whether the H20 chip has any security risks, such as remote tracking or shutdown backdoors, and to submit relevant proof.
On August 6, Nvidia issued a statement categorically denying that its GPU chips contain any backdoor functions, remote shutdown switches, or spyware.
According to Reuters, while the Chinese authorities have not officially banned the sale of H20 chips, they have specifically required major internet companies like Tencent, ByteDance, and Baidu to provide “reasonable explanations” when using H20, especially in government or sensitive applications.
Despite the demand for H20 in the Chinese market, ongoing political and regulatory pressures may prompt enterprises to reduce their procurement volume.