Nvidia’s Huang Warns DeepSeek on Huawei Chips ‘Horrible Outcome’ for US
CEO's admission that Chinese AI stack threatens American dominance signals export controls losing strategic leverage as frontier models bypass US hardware dependency.
Nvidia CEO Jensen Huang called DeepSeek V4’s planned deployment on Huawei chips ‘a horrible outcome for our nation’ in an interview Wednesday, marking the first explicit acknowledgment from US tech leadership that China’s AI infrastructure independence poses an existential competitive threat.
The statement, delivered on the Dwarkesh Podcast on April 16, carries dual significance. It validates DeepSeek’s cost-efficiency breakthrough as a genuine competitive threat rather than marketing, and confirms Chinese semiconductor alternatives are bypassing US Export Controls effectively enough to run frontier AI models. Coming from an executive with financial incentive to minimise competitor threats, the warning represents both competitive defense and credible technical assessment of a shifting landscape.
“DeepSeek is not an inconsequential advance. The day that DeepSeek comes out on Huawei first, that is a horrible outcome for our nation.”
— Jensen Huang, Nvidia CEO
DeepSeek V4 is expected to launch later this month running on Huawei’s Ascend 950PR chips, according to Tech Startups. The model will feature 1 trillion total parameters with approximately 37 billion active per inference pass. Alibaba, ByteDance, and Tencent have placed bulk orders for hundreds of thousands of Ascend 950PR chips ahead of the launch, with TechWire Asia reporting prices rose roughly 20% in recent weeks.
Breaking the CUDA Lock-In
The strategic shift isn’t just about hardware substitution. DeepSeek has rewritten core model components in collaboration with Huawei and Cambricon to run natively on Chinese silicon, moving away from dependency on Nvidia’s CUDA ecosystem. This represents a fundamental architectural divergence — Chinese developers are no longer attempting to optimise software for US hardware under export constraints, but rather co-designing the entire stack for domestic chips.
Huang acknowledged the risk explicitly: “Suppose it optimized for Huawei. Suppose it optimized for their architecture. It would put ours at a disadvantage.” The concern isn’t that Huawei chips match Nvidia’s raw performance — they don’t. Current Huawei Ascend 910C delivers roughly 60% of an H100’s inference throughput, per The Next Web, and US chips are projected to be 17 times more powerful than Chinese equivalents by 2027. The threat is ecosystem fragmentation: if leading Chinese models optimise for a different hardware stack, American chips lose relevance in the world’s largest AI market.
Market Signal
DeepSeek withheld early access to V4 from US chipmakers including AMD and Nvidia while granting exclusive optimisation time to Chinese suppliers, according to reporting in early April. The move signals strategic alignment with domestic semiconductor champions rather than continued dependency on American hardware — even where US chips remain accessible under relaxed export rules.
The Biden administration shifted policy on January 15, easing restrictions from presumption of denial to case-by-case review for certain AI chips destined for China, per Morgan Lewis. The change imposed a 50% volume cap and 25% tariff on H200 exports but opened a controlled pathway for American chips to reach Chinese buyers. DeepSeek’s choice to bypass that pathway entirely for V4 underscores the strategic value Beijing places on supply chain independence, even at a performance cost.
DeepSeek emerged in January 2025 claiming its V3 model cost just $5.6 million to train using 2048 Nvidia H800 GPUs over 55 days. SemiAnalysis later estimated total capital expenditure at $1.3-1.6 billion including infrastructure, R&D, and operations, per CNBC. Regardless of the true figure, the efficiency gains validated that algorithmic innovation could partially overcome hardware constraints — a thesis V4’s deployment now extends to silicon independence.
Congressional Response
House lawmakers called on April 17 for adding DeepSeek, Moonshot AI, and MiniMax to the export control entity list, according to the South China Morning Post. The bipartisan MATCH Act, introduced April 15, would restrict sales of deep ultraviolet lithography tools to China. A House Select Committee on the Chinese Communist Party investigation published April 16 characterised DeepSeek as a “profound threat” and cited its use of export-controlled Nvidia chips acquired through unclear channels, per a Mintz analysis of the report.
Committee leadership stated: “DeepSeek’s ability to develop cutting-edge AI models suggests that loopholes or indirect channels may still exist in US export controls on Nvidia’s chips.” The finding highlights a paradox: tighter restrictions may accelerate rather than slow Chinese self-sufficiency if they force developers toward domestic alternatives before the performance gap widens to prohibitive levels.
- Nvidia faces margin pressure in non-US markets as Huawei becomes commercially viable for training workloads
- CUDA ecosystem lock-in — Nvidia’s primary moat — weakens if Chinese models set developer norms outside American stack
- Export controls lose leverage if frontier performance becomes achievable on sanctioned hardware through software innovation
- US policy response likely to tighten upstream restrictions on lithography and chip design tools rather than downstream chip sales
What to Watch
V4’s actual performance at launch will determine whether Huang’s warning proves prescient or premature. Key metrics to track include:
- Inference latency and throughput on Ascend 950PR versus Nvidia H100/H200 benchmarks
- Adoption rates among Chinese enterprises currently using Nvidia hardware
- Congressional action on entity list additions and lithography tool restrictions in the next 60 days
- Nvidia’s Q2 2026 earnings guidance for China revenue, reported in late May
Huang framed the stakes in ecosystem terms: “If future AI models are optimized in a very different way than the American tech stack, 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 reveals US tech dominance depends less on absolute hardware superiority than on ecosystem lock-in — and that lock-in is breaking faster than export controls can contain it. DeepSeek V4 will test whether algorithmic efficiency and vertical integration can close a 5x performance gap enough to matter commercially. If the answer is yes, Washington’s strategic calculation shifts from denying China access to competing with China’s standards.