DeepSeek R2 Delayed by Huawei Chips

Chinese startup DeepSeek has postponed the launch of its LLM R2. The fault appears to lie with Huawei's Ascend chips, which the company used to train the model instead of components from the American company Nvidia, at the explicit suggestion of Chinese authorities. The Financial Times reports that R2 was supposed to launch as early as May.

Earlier this year, DeepSeek stunned the AI industry with the launch of R1, a free, open-source LLM that was as capable as OpenAI 's GPT o1 model, if not more, but trained at a much lower cost. Following R1's success, the Chinese government pushed the company to seek alternatives to Nvidia's H20 chips for training its new models. The choice fell on Huawei's Ascend chips, but they proved to be less capable than their overseas counterparts. According to some internal sources, the Chinese giant's components suffer from stability issues and slower inter-chip connection speeds than the H20s. The training management software also reportedly isn't up to par with Nvidia 's CUDA toolkit.

Chip mix

To address the issue, Huawei reportedly sent its engineers to work closely with DeepSeek's colleagues, but even direct intervention by the technicians failed to resolve the issue. For this reason, R2 was ultimately retrained with Nvidia chips, but the startup will still use chips from fellow company Huawei to handle the inference processes, i.e., calculating output based on new data entered by users.

If all goes as planned, R2 should finally arrive in the next few weeks, about three months later than the originally planned launch date. This may not seem like a long time, but in a rapidly growing sector like generative AI, it's a huge window of time, giving competitors (including Chinese ones) the opportunity to consolidate their position, launch new features, and gain significant market share.

The Nvidia node

Last Tuesday, August 12, the Trump administration announced an unprecedented agreement with Nvidia and AMD. Both can sell their chips to Chinese companies, provided they pay the US administration 15% of the proceeds from sales in China. The measure is likely unconstitutional, but both Nvidia and AMD accepted the terms, so as not to lose access to a crucial AI market in terms of volume and growth potential. For its part, Beijing has warned its companies to proceed with caution when purchasing versions of Nvidia H20 chips intended for the Chinese market. According to experts cited by central authorities, the components integrate tracking and remote disabling systems. In short, they could be Trojan horses and pose a geopolitical risk for the entire AI sector in China.

Under the agreement with the Trump administration, Nvidia can currently only export H20 chips, while the more advanced and powerful chips from the new Blackwell series, announced in 2024, will only be available in a pared-down version in China.