Investors rotate to China’s chipmakers as DeepSeek intensifies AI competition

Chinese chipmakers rallied on the DeepSeek-V4 launch, with Semiconductor Manufacturing International up 10% and Hua Hong Semiconductor up 15% in Hong Kong as investors bet on stronger demand for computing power. AI application developers moved the other way, with Zhipu AI down 9.1% and MiniMax down 9.4% amid concerns the new model intensifies competition. The backdrop is escalating US-China AI rivalry and tighter US export controls on advanced Nvidia chips, which are pushing Chinese firms toward domestic alternatives.

Analysis

This is less a clean “China AI up” signal than a forced reallocation inside the ecosystem: the market is rewarding the picks-and-shovels layer while marking down software-layer monetization. That makes sense in the near term because a better frontier model increases the implied compute intensity of the entire stack, but it also compresses the scarcity premium for domestic application vendors whose moats were largely narrative-driven. The second-order winner is any domestic supplier that can sell more wafers, packaging, memory, power management, and AI-server subsystems into a buildout cycle even if end-demand visibility remains poor.

The more interesting angle is that export controls may have unintentionally improved the relative positioning of Chinese semiconductor incumbents by narrowing the set of feasible substitutes. If domestic AI firms now need more inference and training capacity to keep pace, procurement urgency could persist for several quarters even if headline AI sentiment cools. That creates a setup where capex beneficiaries outperform application names, while the weakest link is still advanced-node exposure: any rally in local foundry names is likely to be strongest in mature nodes and adjacent equipment rather than in bleeding-edge logic.

The consensus is probably underestimating how quickly this can reverse if the new model is proved cheaper to run than assumed. If DeepSeek’s efficiency translates into lower tokens-per-dollar economics, the bull case for incremental compute demand becomes less linear, and the trade can flip from “more AI equals more chips” to “better AI equals less unit demand per workload” within 1-3 months. For now, the move looks tactically extended in application names and still under-owned in domestic hardware, but the durability depends on whether this sparks genuine enterprise deployment or just another short-lived sentiment rotation.

Key Decisions for Investors

• Long SMIC/Hua Hong on 2-6 week horizon; use pullbacks to build. Reward is continued re-rating on domestic substitution flows, but trim if the basket gives back >8-10% from the first breakout because the move is sentiment-sensitive.
• Short a China AI application basket versus long Chinese semis for 1-2 months. The cleaner expression is long SMIC or a semiconductor ETF proxy / short BABA-adjacent AI platform names where valuation has run ahead of monetization; target 15-25% relative outperformance if the rotation persists.
• Buy NVDA downside hedges rather than outright shorting for 1-3 months. The article is not a near-term demand shock to NVDA, but it reinforces China revenue risk and strengthens the case for a 5-10% hedge against any export-control headline or China substitution narrative.
• For BABA, avoid chasing the long side here; if anything, use strength to sell covered calls over 30-45 days. The company benefits from broader AI enthusiasm, but the market is likely to cap multiple expansion until there is evidence that AI drives incremental cloud and commerce monetization.