Nvidia's Jensen Huang cautions that DeepSeek operating on Huawei chips would result in a 'terrible consequence' for the United States.

      In summary, Nvidia CEO Jensen Huang warned during the Dwarkesh Podcast that if DeepSeek optimizes its AI models for Huawei's Ascend chips instead of American hardware, it would be a “horrible outcome” for the United States. This concern arises as the Chinese AI lab is preparing to launch its V4 foundation model on Huawei's Ascend 950PR processor. The shift from Nvidia's CUDA to Huawei's CANN framework could disrupt the software-hardware dependency that has sustained American AI supremacy, even as US lawmakers advocate for placing DeepSeek on the entity list for export controls.

      Huang emphasized on the Dwarkesh Podcast that if DeepSeek tailors its new AI models to function on Huawei chips rather than on American technology, it poses a direct challenge to the technological advantage that has supported US AI leadership for the last decade. He warned that should future AI models be optimized differently from the American tech stack, and as "AI diffuses into the rest of the world" using Chinese standards, China “will become superior to” the US. This statement gains significance coming from the head of a company that has thrived under the current system, where nearly every leading AI model worldwide utilizes Nvidia GPUs with the CUDA framework.

      DeepSeek is set to launch its V4 multimodal foundation model later this month. Reports indicate that V4 will operate on Huawei’s latest Ascend 950PR processor, while another source suggested it was trained on Nvidia’s Blackwell chips, potentially violating US export controls. However, these claims can coexist: a model can be trained on one hardware set and deployed on another.

      The significance of Huawei's integration lies in the software transition involved. DeepSeek has dedicated months to rewriting its core software to be compatible with Huawei's CANN framework, moving away from the CUDA ecosystem that Nvidia has developed over the past two decades. CUDA's dominance has provided a layer of American control over AI, extending beyond just hardware. Export restrictions can limit Nvidia hardware availability in China, but if Chinese labs continue to develop software for CUDA, they remain reliant on Nvidia's ecosystem, even when using different processors. DeepSeek's shift to CANN disrupts that reliance.

      DeepSeek’s V3 model, launched in late 2024, was trained on 2,048 Nvidia H800 GPUs, a chip designed for the Chinese market but banned from sale to China in 2023. The company has already shown it can produce competitive models with fewer resources than its American counterparts. Its R1 reasoning model matched or exceeded the performance of significantly more expensive models to train. V4 aims to further validate that the company can achieve this without any American hardware.

      Regarding hardware performance, Huawei’s chips currently do not match the superiority of Nvidia’s best offerings. The Ascend 910C, the predecessor of the 950PR, provides about 60% of the inference performance of Nvidia’s H100, a chip already two generations behind Nvidia’s latest. American chips are roughly five times more powerful than their Chinese counterparts today, with projections suggesting the gap could widen to 17 times by 2027. Huawei aims to ship 750,000 AI chips in 2026; however, its total production accounts for only 3 to 5% of Nvidia’s total computing power.

      Yet Huang’s concern transcends the current performance gap. He mentioned on the podcast that despite having inferior chips, China could still catch up in AI development due to its “abundant energy” and “large pool of AI researchers.” This suggests that raw hardware performance is just one aspect, and factors like software optimization, talent availability, and energy resources could compensate for deficiencies in silicon. If V4 demonstrates strong performance on Ascend chips, it legitimizes an alternative route for AI development that does not require Nvidia throughout the supply chain.

      The situation illustrates a conflict central to American chip export policy. Nvidia has restarted production of the H200, a more advanced chip available for sale in China, as confirmed by Huang in March. However, China has been hindering H200 imports to protect Huawei’s domestic chip industry, resulting in Nvidia's CFO stating there has been no revenue from H200 sales in China. The very restrictions intended to inhibit China’s AI capabilities may be accelerating the growth of a competitive Chinese alternative.

      DeepSeek's experience with its R2 model reflects both the potential and limitations of this Huawei initiative. The R2 model faced multiple delays due to training issues on Huawei hardware. While Chinese authorities encouraged the company to use domestic chips, DeepSeek encountered stability problems that compelled it to revert to Nvidia GPUs for training, utilizing Huawei chips only for inference. This distinction is critical: training is the most computation-heavy segment of AI development, and the inability of Huawei chips to reliably support it indicates a tangible hardware gap. However, inference, which generates commercial value, shows that Huawei’s chips are sufficient for that aspect.

      Meanwhile, US lawmakers are advocating for tighter restrictions. Recently, lawmakers and experts accused China of acquiring