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

      In brief: Nvidia's CEO Jensen Huang expressed on the Dwarkesh Podcast that if DeepSeek optimizes its AI models for Huawei’s Ascend chips instead of American technology, it would lead to “a horrible outcome” for the United States. This comes as DeepSeek is set to launch its V4 foundation model on the Huawei Ascend 950PR processor. The shift from Nvidia's CUDA to Huawei's CANN framework threatens to disrupt the dependency that has supported American AI leadership, even as U.S. lawmakers push to add DeepSeek to the entity list for export control.

      During the podcast on Wednesday, Huang stated that if DeepSeek adjusts its new AI models to operate on Huawei chips instead of U.S. hardware, it would pose a “horrible outcome” for the United States. This warning highlights the growing alliance between China’s leading AI laboratory and its top chip manufacturer as a direct threat to the technological advantage that has bolstered U.S. AI for the past decade.

      “If future AI models are optimized in a significantly different manner than the U.S. technology stack,” Huang noted, as “AI spreads globally” with Chinese standards, it could result in China becoming “superior to” the United States. This statement is significant coming from the CEO of a company that has thrived under the existing arrangement, where nearly every cutting-edge AI model globally is developed on Nvidia GPUs using Nvidia’s CUDA software framework.

      What DeepSeek is developing

      DeepSeek is set to introduce its V4 multimodal foundation model later this month. Earlier in April, The Information reported that V4 would utilize Huawei’s latest Ascend 950PR processor, while a separate Reuters article claimed that the model was trained on Nvidia’s Blackwell chips, potentially breaching U.S. export regulations. These two assertions are not inherently contradictory, as a model can be trained on one hardware and implemented on another for inference.

      The significance of Huawei integration lies in the accompanying software transition. DeepSeek has spent months adapting its core code to function with Huawei’s CANN framework, moving away from the CUDA ecosystem that Nvidia has built over two decades as the backbone of AI development. CUDA's predominance has provided a layer of American control over AI that extends beyond the hardware itself. While export restrictions can limit the Nvidia hardware available to China, as long as Chinese labs develop their software for CUDA, they remained reliant on the Nvidia ecosystem, even when using alternative chips. DeepSeek's switch to CANN disrupts that reliance.

      DeepSeek's V3 model, unveiled in late 2024, was trained on 2,048 Nvidia H800 GPUs, a chip specifically designed for the Chinese market that was prohibited from sale to China in 2023. The company has demonstrated its ability to produce competitive frontier models with fewer resources than its U.S. counterparts. Its R1 reasoning model matched or exceeded the performance of models that required significantly higher training costs. V4 aims to extend this strategy by proving the company's capacity to succeed without U.S. hardware at all.

      The hardware disparity and its potential implications

      In terms of raw performance, Huawei’s chips lag behind Nvidia’s leading models. The Ascend 910C, the predecessor to the 950PR, achieves roughly 60% of the inference performance of Nvidia’s H100, which itself is two generations behind Nvidia’s latest products. Currently, American chips are about five times stronger than their Chinese counterparts, and projections suggest that this gap could expand to 17 times by 2027. Huawei aims for 750,000 AI chip shipments in 2026, but that would only represent 3 to 5% of Nvidia’s total computing power.

      However, Huang's concern isn't just about the existing performance gap. He stated on the podcast that even if China has less powerful chips, it could still catch up with U.S. AI advancements due to its “abundant energy” and “large pool of AI researchers.” This suggests that raw hardware performance is merely one factor, and that software optimization, talent among researchers, and energy availability could offset disadvantages in silicon. If V4 performs well on Ascend chips, it could validate an independent pathway for AI development that does not depend on Nvidia at any point in the supply chain.

      The export control dilemma

      The scenario highlights a conflict within American chip export policy. Nvidia has resumed production of the H200, a more powerful chip intended for the Chinese market, as Huang confirmed in March. However, China has been blocking imports of the H200 to protect Huawei’s domestic chip sector, and Nvidia's CFO has stated that the company has made no revenue from H200 sales to China. The restrictions aimed at limiting China’s AI capabilities are instead propelling the growth of a Chinese alternative.

      DeepSeek's experience with its R2 model showcases both the potential and limitations of the Huawei strategy. The training of R2 faced multiple delays due to failures on Huawei hardware. Despite encouragement from Chinese authorities for DeepSeek to train using