Nvidia CEO Jensen Huang just issued a stark warning about the future of AI competition—and it's not about who builds the fastest chip. In a wide-ranging conversation with AI podcaster Dwarkesh Patel, Huang argued that if top-tier open-source models like DeepSeek launch first and perform best on Huawei's hardware, the US could face a "terrible outcome." (See: Microsoft Tests OpenClaw-Style AI Agents That Work Continuously for Users)
https://www.youtube.com/watch?v=Hrbq66XqtCo
The five-layer cake and China's AI advantage
Huang frames the AI industry as a five-layer cake, with chips as the foundational layer. He points out that China holds some serious cards: roughly half the world's AI researchers, massive 7nm chip manufacturing capacity, and abundant energy resources.
Sure, a single Huawei Ascend chip might only deliver one-half to one-third the performance of Nvidia's H200. But you can compensate for that by scaling up the cluster. Most current large models are trained on Hopper architecture anyway, and 7nm chips are already sufficient for that workload.
Why export controls backfire
Huang is unequivocally against extreme chip export bans. He believes they're forcing China to build a complete, independent technology stack from the ground up. The real danger? If a leading open-source model like DeepSeek gets deeply optimized specifically for Huawei's Ascend architecture, developers worldwide—especially in cost-sensitive markets like the Middle East, Southeast Asia, and Latin America—will flock to that ecosystem for better performance and lower cost.
Compute platforms have powerful lock-in effects. Moving from CUDA to PyTorch is already a massive migration cost for developers. If the best-performing Chinese AI models run on non-US hardware, American technical standards risk becoming irrelevant.
"Abandoning the world's second-largest tech market is like giving away the most important layer of the AI industry," Huang emphasized. Nvidia's philosophy, he says, is to "do what's necessary, and as little as possible."
He argues the US should maintain its lead through relentless innovation, not by retreating from competition with a "loser's mentality." He notes that Huawei posted its best-ever financial results in 2024, and a wave of Chinese chip companies have gone public—proof that export controls haven't stifled China's chip industry but instead accelerated its self-sufficient ecosystem.
AI safety and the road ahead
On AI safety, Huang acknowledges advanced models have potential for cyber attacks, but he believes the solution lies in dialogue and cooperation, not technological blockade. He dismissed comparisons of AI to enriched uranium as "a terrible, illogical analogy," stressing that chips and nuclear weapons are fundamentally different.
Huang also shared Nvidia's product roadmap: the Vera Rubin platform launches this year, Vera Rubin Ultra arrives next year, and the Feynman architecture follows the year after. Nvidia's moat, he contends, isn't just raw hardware performance. It's the CUDA ecosystem that runs on every major cloud platform and supports countless frameworks, plus the sheer efficiency measured in tokens per watt and performance per dollar.