Nvidia’s Huang Warns DeepSeek Running on Huawei Chips Would Be ‘Horrible’ for the US

Why It Matters

If Chinese labs can build competitive models without Nvidia hardware, the United States loses a key lever of technological and economic leverage in the global AI race.

Key Takeaways

• DeepSeek's V4 will run on Huawei Ascend 950PR processor.
• Migration to Huawei's CANN breaks Nvidia's CUDA software dependency.
• US export controls may weaken if Chinese models succeed on local chips.
• Huawei chips lag Nvidia by up to five times in performance today.
• Nvidia's market cap exceeds $3 trillion, underscoring the strategic stakes.

Pulse Analysis

Nvidia has built a de‑facto monopoly on AI infrastructure by coupling its GPUs with the CUDA software stack. For over a decade, developers worldwide have written models for CUDA, creating a self‑reinforcing ecosystem that ties AI progress to American silicon. DeepSeek’s decision to target Huawei’s Ascend 950PR processor and rewrite its core for the CANN framework directly challenges that model, signaling the first serious attempt to decouple cutting‑edge AI from Nvidia’s hardware and software dominance. The move arrives as Washington tightens export controls, aiming to curb China’s access to high‑performance chips, yet it also accelerates the search for a home‑grown alternative.

Technically, the migration raises questions about performance trade‑offs. Huawei’s Ascend line delivers roughly 60% of the inference speed of Nvidia’s H100 and is five times less powerful than the latest US‑made chips, a gap projected to widen to 17‑fold by 2027. Nevertheless, DeepSeek’s prior R2 model showed that training on Huawei silicon remains problematic, forcing a fallback to Nvidia GPUs for the most compute‑intensive phases. By focusing on inference—where commercial value is realized—Huawei’s chips may be sufficient, especially if software optimisation narrows the hardware gap. The success of V4 could prove that a viable AI pipeline exists without reliance on Nvidia’s ecosystem.

The broader implications are geopolitical and financial. Nvidia’s market capitalisation now tops $3 trillion, and its data‑center revenue surged 93% year‑over‑year, reflecting deep integration of its technology across the global AI supply chain. If DeepSeek demonstrates competitive performance on domestic chips, the rationale for strict export bans erodes, potentially reshaping US policy and diminishing Nvidia’s bargaining power. Moreover, a thriving Chinese AI stack could attract domestic and allied investments, altering the competitive landscape for AI talent, capital, and compute resources. Stakeholders from investors to policymakers must therefore monitor DeepSeek’s V4 rollout as a bellwether for the future balance of AI power between the United States and China.