Can the US win the AI cold war with China?

The article argues that escalating US restrictions on China’s AI sector could backfire by accelerating Chinese self-sufficiency, straining allied cooperation and raising costs for US technology companies. China remains dependent on imported advanced chips and equipment, but it controls 91% of global rare-earth separation and refining capacity, giving Beijing potential leverage over US chipmaking, manufacturing and defense supply chains. Nvidia CEO Jensen Huang says export controls have cost US companies billions, while nearly 200 US startups warned that restricting Chinese open-source AI models could consolidate the domestic market among a small group of AI giants.

Analysis

The near-term market risk is not a binary loss of Chinese revenue but a recurring product-cycle tax on NVDA: each rule revision can strand tailored inventory, require engineering resources for compliant SKUs, and reduce the pricing power that has supported extraordinary gross margins. A further tightening would likely hit NVDA’s China-adjacent data-center revenue and guidance credibility within days, while the broader earnings effect emerges over the next 1-3 quarters. ASML is relatively better insulated because its service and installed-base economics provide recurring cash flow, but its order mix and medium-term backlog conversion remain vulnerable if allied policy alignment weakens.

The underappreciated second-order effect is that restrictions incentivize compute efficiency rather than merely reducing Chinese AI output. More efficient model architectures lower inference cost globally, which is disinflationary for GPU demand per unit of AI usage but potentially expands total AI adoption; that outcome favors hyperscalers with proprietary distribution (MSFT, GOOGL, AMZN) over merchant hardware suppliers if compute intensity falls faster than workloads grow. A forced split in technology stacks also raises duplication costs for semiconductor-equipment supply chains, supporting strategic domestic-processing assets but pressuring companies whose valuation assumes frictionless global TAM expansion.

Rare-earth exposure is a more credible tail-risk hedge than a core directional thesis. Any renewed disruption would initially re-rate non-Chinese processing capacity and expose defense-prime supply-chain vulnerabilities, but MP’s upside remains contingent on successfully converting domestic processing capacity into qualified, economic output rather than simply benefiting from spot-price headlines. The consensus appears too focused on headline chip controls and insufficiently focused on allied compliance fatigue: a negotiated carve-out or weaker enforcement would be a positive surprise for NVDA and ASML, making outright structural shorts unattractive without a concrete policy catalyst.

Key Decisions for Investors

• Maintain a 1-3 month tactical underweight in NVDA versus MSFT or GOOGL rather than an outright semiconductor short. Use any policy-driven NVDA rally to initiate the pair; the thesis is multiple and margin risk from compliant-product redesign, while the risk is hyperscaler capex remaining strong enough to offset China leakage. Cover if NVDA raises data-center gross-margin guidance or demonstrates no material revenue displacement in the next earnings release.
• Buy a small 6-12 month MP position only on confirmation of tighter rare-earth export licensing or a sustained rise in NdPr pricing; size as an event hedge, not a base-case commodity trade. Target asymmetric upside from strategic re-rating, but exit on processing-ramp delays, negative unit-economics disclosure, or evidence that alternative supply is readily qualified.
• Prefer ASML to NVDA on a relative basis over 6-18 months, but avoid adding before clarity on order-book conversion and China-related service exposure. The relative trade benefits if hardware export controls repeatedly compress NVDA’s addressable market; it fails if additional equipment restrictions materially reduce ASML bookings or if AI-capex demand reaccelerates NVDA faster than policy friction grows.
• Set a policy alert for coordinated US-EU-Japan restrictions on additional semiconductor equipment, cloud-compute access, or model distribution. Such coordination would justify adding downside protection on NVDA and reassessing ASML; unilateral US action without allied participation is more likely to be a short-lived sentiment event than a durable earnings reset.