The U.S. says China's AI progress is down to 'distillation.' But is it that clear cut?

Cohere CEO Aidan Gomez said Chinese AI models are now "world class" and, on some benchmarks, outperform leading U.S. models, arguing that illicit model distillation cannot fully explain China’s progress. His view challenges allegations from Anthropic and U.S. agencies that Chinese labs including Alibaba, Moonshot and DeepSeek have used industrial-scale extraction of proprietary model capabilities. Tight U.S. access to advanced Nvidia chips has also pushed Chinese developers toward leaner architectures and domestic semiconductor development, raising the risk that Western export-control policy underestimates China’s independent AI capability.

Analysis

The investable implication is not that U.S. frontier models immediately lose their lead, but that model capability is becoming a less defensible source of economic rent. If leaner architectures deliver comparable task performance on constrained hardware, inference cost—not absolute training compute—becomes the competitive battleground. That favors hyperscalers with distribution and enterprise workflow ownership, while pressuring standalone model providers whose valuations assume persistent benchmark leadership and premium API pricing.

For NVDA, a faster Chinese innovation cycle is a two-sided signal. Near term (days to 3 months), it reinforces global compute demand and validates the urgency of capacity procurement outside China; however, over 6-18 months, software efficiency and domestic accelerator substitution could reduce the strategic value of each marginal high-end GPU and permanently shrink the China-addressable opportunity. The key issue is whether Chinese developers achieve production-scale inference economics on local silicon, rather than isolated benchmark wins.

BABA has the clearest asymmetric optionality: a credible enterprise-AI offering can improve cloud utilization, attach rates and merchant productivity while challenging the market's assumption that its AI assets are structurally non-monetizable. But sanctions remain the binding valuation constraint, and capability claims are insufficient without evidence of paid cloud consumption, rising AI-related ARPU, and margin resilience. AMZN's more relevant risk is sovereign-cloud concentration: prolonged regional service disruption would increase redundancy capex and strengthen multi-cloud procurement behavior, modestly diluting AWS switching-cost economics.

Consensus likely overstates a binary U.S.-versus-China model race and understates fragmentation. Export controls may create two AI stacks with different hardware, tooling and compliance requirements; that raises implementation costs for multinational customers even as it creates localized winners. The falsifier for the China-efficiency thesis is persistent dependence on imported high-end accelerators, no measurable enterprise monetization, or widening capability gaps in real-world coding and agentic workloads over the next two quarterly model cycles.

Key Decisions for Investors

• Maintain a tactical long NVDA over the next 1-3 months, but hedge the 6-18 month China-substitution risk with a defined-risk put spread or reduce exposure into evidence that domestic Chinese accelerators are supporting large-scale inference. Thesis fails if hyperscaler capex guidance weakens or NVDA indicates a material China revenue/units reset.
• Put BABA on a catalyst watch rather than chase headline-driven strength: initiate only if the next two reporting periods show AI-cloud revenue acceleration, improving cloud margin, and disclosed enterprise adoption. A 6-12 month long is attractive only with confirmation; export-control tightening or flat cloud growth invalidates the rerating case.
• Pair trade for a 6-12 month horizon: long AMZN / short a basket of frontier-model-dependent private-market proxies where available, or underweight listed AI application vendors with limited proprietary distribution. The mechanism is margin migration from model suppliers toward cloud, data, and enterprise distribution layers; reverse if model API pricing remains firm despite open-source performance gains.
• Monitor Chinese semiconductor supply-chain indicators rather than model benchmarks: domestic accelerator shipment volumes, cloud inference pricing, and utilization are the actionable confirmation signals. If local inference pricing falls materially versus imported-GPU alternatives, reassess NVDA China terminal-value assumptions and upgrade China AI infrastructure exposure where liquid vehicles become available.