Jensen Huang, chief executive of Nvidia, warned on the Dwarkesh Podcast that DeepSeek’s decision to optimize its next‑generation V4 foundation model for Huawei’s Ascend 950PR processor would be “a horrible outcome” for the United States. The comment underscores a growing concern in Washington that China could erode the software‑hardware moat that has kept American firms at the forefront of artificial‑intelligence development for more than a decade.
DeepSeek, China’s most advanced AI lab, is set to launch V4 later this month. According to The Information, the multimodal model will run on Huawei’s latest Ascend 950PR chip, while Reuters reported that training may have taken place on Nvidia’s Blackwell GPUs. The two statements are not mutually exclusive; a model can be trained on one platform and deployed for inference on another. What makes the move significant is DeepSeek’s migration from Nvidia’s CUDA framework to Huawei’s CANN software stack.
CUDA has become the de‑facto development environment for AI. Virtually every frontier model outside China is built on Nvidia GPUs using CUDA, giving the United States a layer of control that extends beyond the chips themselves. By rewriting its core code for CANN, DeepSeek is breaking that dependency. If the V4 model performs well on Ascend silicon, it would demonstrate that competitive AI systems can be built without any Nvidia hardware, challenging the premise of current export‑control policies.
Huawei’s chips lag behind Nvidia’s top offerings. The predecessor Ascend 910C delivers roughly 60 % of the inference performance of Nvidia’s H100, and the performance gap is projected to widen to 17 times by 2027. Nonetheless, Huang argued that raw hardware power is only one variable. China’s abundant energy, large pool of AI researchers, and now a viable software stack could allow it to catch up even with inferior silicon.
DeepSeek’s track record adds weight to the debate. Its V3 model, released in late 2024, was trained on 2,048 Nvidia H800 GPUs—chips specifically designed for the Chinese market but banned from sale to China in 2023. The lab’s R1 reasoning model matched or exceeded the performance of far more expensive Western counterparts. Yet the company’s R2 model stumbled during training on Huawei hardware, forcing a return to Nvidia GPUs for the most compute‑intensive phase. The distinction matters: training demands the highest performance, while inference—where commercial value is realized—may be adequately served by Huawei’s chips.
U.S. lawmakers are watching the situation closely. On Thursday, several members of Congress and technology experts called for DeepSeek, along with Moonshot AI and MiniMax, to be placed on the entity list, arguing that the lab’s shift toward domestic hardware could undermine U.S. export restrictions. Nvidia itself has restarted production of its H200 chip for the Chinese market, but Chinese buyers have largely blocked imports to protect Huawei’s domestic chip business.
Huang’s warning ultimately targets the software‑hardware co‑design that has made Nvidia the linchpin of the global AI supply chain. CUDA’s dominance means that governments, enterprises, and researchers all buy Nvidia GPUs because the software ecosystem demands it. A parallel ecosystem built around CANN could shift that calculus, reducing demand for Nvidia’s data‑center GPUs and weakening the justification for stringent export controls.
The stakes are concrete. Nvidia’s market capitalization exceeds $3 trillion, and its data‑center revenue grew 93 % year over year in the most recent quarter. If DeepSeek’s V4 proves competitive on Huawei silicon, the argument for continued reliance on Nvidia hardware—and the policies built around that reliance—could lose traction.
While Huawei’s chips remain far behind Nvidia’s in raw performance, the real concern is trajectory. Should DeepSeek’s model run effectively on Ascend processors, other Chinese labs may follow, eroding the CUDA moat that has kept Nvidia at the apex of the AI hardware market. Huang’s public admonition suggests he believes the risk is no longer theoretical but imminent.
Este artigo foi escrito com a assistência de IA.
News Factory SEO ajuda você a automatizar conteúdo de notícias para o seu site.