In short: Nvidia CEO Jensen Huang warned on the Dwarkesh Podcast that DeepSeek optimising its AI models for Huawei’s Ascend chips instead of American hardware would be “a horrible outcome” for the United States, as the Chinese AI lab prepares to launch its V4 foundation model on Huawei’s Ascend 950PR processor. The migration from Nvidia’s CUDA to Huawei’s CANN framework threatens to break the software-hardware dependency underpinning American AI dominance, even as US lawmakers push to place DeepSeek on the entity list for export control.
Nvidia CEO Jensen Huang said on the Dwarkesh Podcast on Wednesday that if DeepSeek optimised its new AI models to run on Huawei chips rather than American hardware, it would be “a horrible outcome” for the United States. The warning frames the emerging partnership between China’s most capable AI lab and its most advanced chipmaker as a direct threat to the technological leverage that has underpinned American AI dominance for the past decade.
“If future AI models are optimised in a very different way than the American tech stack,” Huang said, and as “AI diffuses out into the rest of the world” with Chinese standards and technology, China “will become superior to” the US. The statement is notable because it comes from the CEO of the company that has benefited most from the current arrangement, in which virtually every frontier AI model in the world is trained on Nvidia GPUs using Nvidia’s CUDA software framework.
What DeepSeek is building
DeepSeek is preparing to launch V4, a multimodal foundation model expected later this month. The Information reported earlier in April that V4 would run on Huawei’s latest Ascend 950PR processor, while a separate Reuters report suggested the model had been trained on Nvidia’s Blackwell chips, which would constitute a violation of US export controls. The two claims are not necessarily contradictory: a model can be trained on one set of hardware and deployed for inference 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 the CUDA ecosystem that Nvidia has spent two decades building into the foundation of AI development. CUDA’s dominance has functioned as a second layer of American control over AI, beyond the chips themselves. Export restrictions can limit which Nvidia hardware reaches China, but as long as Chinese labs wrote their software for CUDA, they remained dependent on the Nvidia ecosystem even when using alternative processors. DeepSeek’s move to CANN breaks that dependency.
DeepSeek’s V3 model, launched in late 2024, was trained on 2,048 Nvidia H800 GPUs, a chip tailor-made for the Chinese market that was itself banned from sale to China in 2023. The company has already demonstrated that it can produce frontier-competitive models with fewer resources than its American rivals. Its R1 reasoning model matched or exceeded the performance of models that cost orders of magnitude more to train. V4 would extend that approach by proving the company can do it without American hardware at all.
The hardware gap and why it may not matter
On raw performance, Huawei’s chips are not competitive with Nvidia’s best. The Ascend 910C, the predecessor to the 950PR, delivers roughly 60% of the inference performance of Nvidia’s H100, a chip that is itself two generations behind Nvidia’s current best. American chips are approximately five times more powerful than their Chinese equivalents today, and that gap is projected to widen to 17 times by 2027. Huawei is targeting 750,000 AI chip shipments in 2026, but its total production represents only 3 to 5% of Nvidia’s aggregate computing power.
But Huang’s concern is not about the current performance gap. He said on the podcast that even if China had inferior chips, it could still catch up with the US in AI development given its “abundant energy” and “large pool of AI researchers.” The implication is that raw hardware performance is only one variable, and that software optimisation, researcher talent, and energy availability can compensate for silicon disadvantages. If V4 performs well on Ascend chips, it validates an alternative path for AI development that does not depend on Nvidia at any point in the supply chain.
The export control paradox
The situation exposes a tension at the centre of American chip export policy. Nvidia restarted production of the H200, a more powerful chip, for sale in China, as Huang confirmed in March. But China has been blocking H200 imports to protect Huawei’s domestic chip business, and Nvidia’s CFO has said the company has recorded no revenue from China H200 sales. The controls designed to limit China’s AI capabilities are instead accelerating the development of a Chinese alternative.
DeepSeek’s experience with its R2 model illustrates both the promise and the limits of the Huawei path. R2 was repeatedly delayed because of training failures on Huawei hardware. Chinese authorities had urged DeepSeek to train on domestic chips, but the company encountered stability issues that forced it to revert to Nvidia GPUs for training while using Huawei chips only for inference. The distinction matters: training is the most compute-intensive phase of AI development, and the fact that Huawei chips could not handle it reliably suggests the hardware gap is real. But inference, the phase where models serve users, is where commercial value is generated, and Huawei’s chips appear adequate for that purpose.
Meanwhile, US lawmakers are pushing to tighten restrictions further. On Thursday, lawmakers and experts accused China of buying “what they can” and stealing “what they cannot” in the AI industry, and called for the government to evaluate placing DeepSeek, Moonshot AI, and MiniMax on the entity list for export control.
What Huang is really warning about
Huang’s warning is ultimately about software-hardware co-design. Nvidia’s dominance rests not just on making the fastest chips but on CUDA’s position as the default development environment for AI. When researchers write code, they write it for CUDA. When startups build products, they build them on CUDA. When governments invest in AI infrastructure, they buy Nvidia GPUs because that is what the software requires. DeepSeek’s migration to CANN threatens to create a parallel ecosystem in which none of that applies.
The scale of Nvidia’s business makes the stakes concrete. The company’s market capitalisation exceeds $3 trillion. Its data centre revenue grew 93% year over year in its most recent quarter. Its chips power the training runs for virtually every major AI model outside China. If the most capable Chinese AI lab demonstrates that competitive models can be built without Nvidia, the argument for maintaining export controls weakens, the argument for buying Nvidia weakens, and the geopolitical assumptions that have shaped AI policy for the past three years come under pressure.
None of this means Huawei is about to overtake Nvidia. The performance gap is large and growing. The R2 training failures demonstrate that Chinese hardware is not yet ready for the most demanding AI workloads. But Huang is not warning about today. He is warning about a trajectory in which DeepSeek proves the concept, other labs follow, and the CUDA moat that has made Nvidia the most valuable company in the AI supply chain begins to erode.
The fact that the CEO of Nvidia is the one making this argument publicly suggests he believes the risk is no longer theoretical. DeepSeek’s V4 will be the first major test. If a multimodal foundation model runs competitively on Huawei silicon, the warning Huang issued on Wednesday will look less like corporate lobbying and more like the most consequential forecast in the AI chip war so far.
