China vs US: Who is winning the AI race, in four charts

The US retains a substantial AI-infrastructure lead, holding nearly 75% of global AI computing power versus China’s just over 14%, with US hyperscalers projected to spend $764bn in 2026 against $102bn for China’s major platforms. China is narrowing the model gap—DeepSeek V4 Pro was estimated at roughly eight months behind leading US systems—and Chinese models lead OpenRouter usage, aided by lower costs and open-weight availability. China also produced more than 27% of global AI publications in 2024 versus 12% for the US, while the planned Trump-Xi summit may address an AI national-security incident notification mechanism alongside broader trade and geopolitical issues.

Analysis

The investable asymmetry remains in the AI supply chain rather than model rankings. NVDA’s scarcity value is supported by the combination of hyperscaler capex commitments and the cost of delayed deployment, but export restrictions convert China from an incremental revenue opportunity into a catalyst for indigenous substitution. Huawei’s gains pressure NVDA’s long-duration China TAM while benefiting the non-China installed base; the more important read-through is that Chinese buyers will optimize around lower-cost, open-weight models, reducing the compute required per application and eventually challenging inference pricing across cloud platforms.

For AMZN, MSFT, GOOG, META and ORCL, the next 1-3 months hinge on whether AI infrastructure spend translates into disclosed utilization, cloud backlog and incremental revenue rather than capex escalation alone. The market is likely underpricing divergence: MSFT and AMZN have the clearest enterprise monetization channels, while META’s return depends on ad-ranking/productivity gains and GOOG faces the greatest risk that lower-cost open models commoditize search-adjacent AI features. BABA and BIDU gain from local model adoption, but ADR holders retain a policy discount that can overwhelm operating progress around bilateral negotiations.

A bilateral AI incident-notification framework would be modestly risk-positive for globally exposed semis and Chinese ADRs if it lowers the probability of abrupt restrictions; it does not alter the strategic trajectory toward a bifurcated hardware stack. The contrarian risk is that consensus treats all hyperscaler capex as durable demand: a slowing in cloud AI workloads or evidence that customers shift to cheaper models would compress infrastructure returns before revenue catches up. Falsify the constructive US-infrastructure view if the next earnings cycle shows capex growth without backlog/RPO acceleration or if NVDA guides data-center revenue below hyperscaler deployment plans.

Key Decisions for Investors

• Maintain a 6-12 month long NVDA / short BABA pair in equal beta-adjusted dollars: NVDA retains exposure to global premium compute while BABA carries China-specific hardware, pricing and ADR-policy risk. Reassess if NVDA’s next data-center guide misses consensus or if US export rules materially broaden permitted high-end accelerator sales to China.
• Prefer long MSFT and AMZN over GOOG for the next 1-3 earnings cycles; size only after verifying AI cloud backlog and utilization disclosures. The thesis is enterprise workload monetization, not model leadership; exit relative longs if Azure/AWS growth decelerates despite elevated AI capacity additions.
• Avoid adding broad hyperscaler exposure solely on capex headlines. Set an alert for a second consecutive quarter in which AMZN, MSFT, GOOG and META raise capex while aggregate cloud revenue growth and RPO/backlog fail to improve; that setup favors reducing AI-infrastructure beta and raising downside hedges on QQQ.
• For higher-risk China exposure, use a small 6-18 month BABA/BIDU basket only following evidence of improving cloud AI revenue or margin disclosure, not leaderboard usage metrics. A new US restriction on cloud access, model weights, or semiconductor intermediates is a stop condition.