Chinese AI firms push beyond Nvidia as DeepSeek turns to Huawei

DeepSeek said its latest AI model has been optimized to run on Huawei chips for inference, while still relying on Nvidia chips for training, underscoring China’s shift toward a domestic AI hardware stack. Huawei said it plans to release a training chip this year, but acknowledged it will take another year to match Nvidia’s current performance. The article highlights how US export controls are accelerating China’s push for self-sufficiency and could deepen the bifurcation of the global AI chip market.

Analysis

The market implication is not simply that China is “de-Nvidia-fying” its stack; it is that the bargaining power in the ecosystem shifts from chip vendors to model developers and local integrators. If inference workloads migrate first, the revenue hit to Nvidia is initially modest, but the strategic damage is larger: every optimized Chinese model reduces the addressable market for premium GPU attach rates, software lock-in, and future training demand. The second-order effect is a faster commoditization of inference hardware globally, which could compress pricing power even outside China over the next 12-24 months.

For TSM, the near-term read-through is mixed rather than outright bearish. On one hand, reduced dependence on Taiwan-fabricated leading-edge chips in China is a structural negative for incremental China demand; on the other hand, any successful domestic Chinese alternative still depends on a broader ecosystem of design tools, packaging, memory, and yields that remain constrained. That means the displacement curve is likely nonlinear: weak today, more meaningful if Huawei/SMIC improve high-volume output, but not enough to erase TSM’s leadership in the foreseeable 1-2 years. The bigger risk is not lost unit demand alone, but margin pressure if geopolitical fragmentation forces more duplicative capex across the supply chain.

The contrarian point is that export controls may be creating a worse outcome for U.S. chipmakers than the market currently prices: they may be accelerating a China-specific architecture standard that becomes self-reinforcing. That said, the bullish counterargument for NVDA is that the frontier model race still runs on best-in-class training hardware, and China’s domestic stack remains behind on efficiency, yields, and scale. So the immediate setup is less “China replaces Nvidia” and more “Nvidia’s China optionality gets capped while its rest-of-world franchise remains intact.”

Catalysts over the next 1-3 months are mostly political: summit optics, any signal on H200 enforcement, and whether Beijing formalizes procurement preferences for domestic chips. Over 6-18 months, watch for evidence that Chinese model performance degrades less than expected when inference is fully local, because that would validate the alternative-stack thesis and pressure all AI hardware multiples.

Key Decisions for Investors

• Short NVDA vs long semis basket excluding China-sensitive exposure for 1-3 months; thesis is capped China optionality and multiple compression if policy headlines stay restrictive. Use tight stop if U.S.-China talks produce credible H200 approval or easing.
• Reduce tactical exposure to TSM on any China-led AI hardware enthusiasm; the risk/reward is asymmetrically worse if domestic Chinese fabs begin to prove volume scalability. Favor a wait-for-pullback approach over adding here.
• Optionality trade: buy 3-6 month NVDA downside puts financed with out-of-the-money call spreads to express headline risk around export controls and summit outcomes, while limiting premium burn if the market shrugs off the news.
• Pair trade: long AI software/application beneficiaries with minimal hardware supply exposure vs short AI hardware names most levered to China demand. The trade benefits if the market starts valuing ecosystem capture over chip unit growth.