In short: Nvidia CEO Jensen Huang warned on the Dwarkesh Podcast that if DeepSeek optimizes its AI models for Huawei’s Ascend chips instead of American hardware, there will be “terrible consequences” for the US, as the Chinese AI lab prepares to launch a V4 foundational model on Huawei’s Ascend 950PR processor. The migration from Nvidia’s CUDA to Huawei’s CANN framework threatens to break America’s software-hardware dependence on artificial intelligence, even as US lawmakers seek to place DeepSeek on the entity list for export controls.
Nvidia CEO Jensen Huang said on the Dwarkesh Podcast on Wednesday that it would be a “terrible outcome” for the United States if DeepSeek optimized its new AI models to run on Huawei chips rather than American hardware. The warning describes the emerging partnership between China’s most capable artificial intelligence lab and its most advanced chipmaker as a direct threat to the technological leverage that has supported America’s dominance of artificial intelligence for the past decade.
“If future AI models are optimized very differently from the American tech stack“, Huang said and “Artificial intelligence is spreading to the rest of the world” China with Chinese standards and technology”will be superior to” US The statement is notable because it comes from the CEO of a company that benefits most from the current arrangement in which nearly every frontier AI model in the world is trained on Nvidia GPUs using Nvidia’s CUDA software framework.
What DeepSeek builds
DeepSeek is set to launch V4, its multimodal foundational model, which is expected later this month. In April, it was reported that the V4 would run on Huawei’s latest Ascend 950PR processor, while a separate Reuters report suggested that the model was powered by Nvidia’s Blackwell chips, a violation of US export controls. The two claims are not necessarily contradictory: one model can be trained on one set of hardware and deployed to produce results on another.
What makes the Huawei integration significant is the software migration behind it. DeepSeek has spent months rewriting its core code to work with Huawei’s CANN framework, moving away from Nvidia’s CUDA ecosystem, which it has built over two decades to underpin its AI development. CUDA’s dominance has functioned as a second layer of American control over AI beyond the chips themselves. Export restrictions may limit which Nvidia hardware reaches China, but as long as Chinese labs write their software for CUDA, they remain dependent on the Nvidia ecosystem even when using alternative processors. DeepSeek’s move to CANN breaks this dependency.
Released in late 2024, DeepSeek’s V3 was trained on the 2048 Nvidia H800 GPU, a chip specifically designed for the Chinese market, which was banned from being sold to China in 2023. The company has already demonstrated that it can produce. frontier competitive models With fewer resources than their American competitors. His R1 reasoning model matched or outperformed models that cost orders of magnitude more to train. V4 will expand on this approach by proving that the company can do without American equipment.
Hardware gap and why may not matter
In terms of raw performance, Huawei’s chips can’t compete with Nvidia’s best. The 950PR’s predecessor, the Ascend 910C, features roughly 60% of Nvidia’s current best chip, two generations behind Nvidia’s H100. American chips are about five times more powerful than their Chinese equivalents today, and the gap is projected to widen 17 times by 2027. Huawei aims to ship 750,000 AI chips in 2026, but its total production is only 3-5% of Nvidia’s total computing power.
But Huang’s concern is not about the current performance gap. He said in the podcast that even if China has inferior chips, it can catch up with the United States in the development of artificial intelligence.plenty of energy” and “Large pool of AI researchers”. The bottom line is that hardware performance is only one variable, and software optimization, researcher talent and power availability can compensate for silicon deficiencies. If V4 performs well on Ascend chips, it validates an alternative path for AI development independent of Nvidia at any point in the supply chain.
The export control paradox
The situation reveals a tension at the heart of America’s chip export policy. As Huang confirmed in March, Nvidia has restarted production of the more powerful H200 chip for sale in China. But China blocks Huawei’s H200 imports to protect its domestic chip business, and Nvidia’s CFO said the company has not made any revenue from Chinese H200 sales. Instead of controls designed to limit China’s AI capabilities, China is accelerating the development of an alternative.
DeepSeek’s experience with the R2 model shows both the promise and the limits of Huawei’s path. R2 was repeatedly delayed due to training failures on Huawei hardware. Chinese authorities urged DeepSeek to train on domestic chips, but the company ran into stability issues, forcing it to revert to Nvidia GPUs for training while using Huawei chips only for inference. The distinction is important: training is the most computationally intensive phase of AI development, and the inability of Huawei chips to handle it reliably suggests that the hardware gap is real. But the bottom line is that the stage where models serve users is where the commercial value comes in, and Huawei’s chips seem adequate for that purpose.
Meanwhile, US lawmakers are trying to tighten restrictions even further. On Thursday, lawmakers and experts accused China of the buyout.what they can do” and theft “what they can’t do” worked on the artificial intelligence industry and called on the government to consider placing DeepSeek, Moonshot AI and MiniMax on the export control enterprise list.
What Huang is really warning about
Huang’s warning is ultimately about software-hardware co-design. Nvidia’s advantage is based not only on making the fastest chips, but also on CUDA’s position as the standard development environment for artificial intelligence. When researchers write code, they write it for CUDA. When startups develop products, they build them on CUDA. When governments invest in AI infrastructure, they buy Nvidia GPUs because the software demands it. DeepSeek’s migration to CANN threatens to create a parallel ecosystem where none of these apply.
The The scale of Nvidia’s business specifies the stakes. The market capitalization of the company exceeds 3 trillion dollars. Its data center revenue grew 93% year over year in the last quarter. Its training chips work for almost every major AI model outside of China. If the most capable Chinese AI lab demonstrates that competitive models can be developed without Nvidia, the argument for keeping export controls weakens, the argument for buying Nvidia weakens, and geopolitical assumptions For the past three years, the policies that have shaped AI policy are under pressure.
None of this means that Huawei will overtake Nvidia. The performance gap is huge and growing. Failures in R2 training show that Chinese hardware is not yet ready for the most demanding AI workloads. But Huang does not warn about this day. He warns of DeepSeek’s proof-of-concept, a trajectory followed by other labs, and the CUDA moat that has made Nvidia the most. valuable company AI is starting to erode the supply chain.
The fact that Nvidia’s CEO made this argument publicly suggests that he believes the risk is no longer theoretical. DeepSeek’s V4 will be the first big test. If the multimodal foundation model works competitively on Huawei silicon, Huang’s warning on Wednesday will sound less like corporate lobbying and more like it. the most effective prediction so far in the AI chip war.
