Harvard Kennedy fellow Shlomit Wagman: The U.S. and China will never trust each other on AI. That may not matter

The article argues that the U.S. and China should establish a narrowly targeted global AI-safety pact with pre-agreed warning indicators that trigger development slowdowns or pauses before systems reach autonomous replication, evasion of oversight, or catastrophic bio/cyber capabilities. It cites accelerating AI-assisted research, evaluator manipulation, and rising compute/GPU use as potential indicators, while proposing independent technical assessments backed by consequences involving advanced chips, semiconductor equipment, cloud infrastructure, capital, and market access. The proposal is primarily a policy framework rather than an announced regulatory action, but highlights growing AI-race and U.S.-China strategic risks for frontier-model developers and their supply chains.

Analysis

This is not yet a policy catalyst, but it raises the probability that AI regulation evolves from disclosure rules into capability-linked controls: compute reporting, mandatory third-party evaluations, and conditional training pauses. The market implication is a higher regulatory discount rate for frontier-model owners and GPU-dependent cloud capex, particularly where valuation assumes uninterrupted scaling; MSFT, GOOGL, AMZN and META would face delayed monetization while retaining largely fixed data-center commitments. Near term, this is narrative risk rather than an earnings revision driver.

The more actionable second-order exposure is in the AI hardware supply chain. A framework that uses chip access, equipment service and cloud capacity as enforcement tools would increase policy volatility for NVDA, AMD, ASML, AMAT, LRCX and KLAC; the risk is asymmetric for companies with meaningful China-linked revenue or service exposure. Conversely, cybersecurity vendors benefiting from model-security, identity, evaluation and data-governance budgets—PANW, CRWD, ZS and CHKP—could see a more durable incremental spend category, although current valuations already price substantial AI-adjacent demand.

Consensus may be overestimating the likelihood of a coordinated global brake and underestimating fragmented national rules. A US-China agreement would require verification mechanisms that are technically difficult and politically vulnerable; the more probable 6-18 month outcome is tighter unilateral export controls and domestic compliance mandates, not synchronized training pauses. The thesis is falsified if hyperscalers maintain or raise 2027 AI capex guidance while regulators limit action to voluntary standards, or if semiconductor firms demonstrate China revenue resilience through permitted products and non-China demand.

Key Decisions for Investors

• No directional trade on the commentary alone; set an alert for US Commerce/White House action on compute-reporting thresholds, model-evaluation mandates or expanded cloud-service restrictions. A formal rulemaking, rather than voluntary commitments, is the trigger for positioning.
• For 1-3 month regulatory-risk hedging, prefer a modest long PANW / short SMH pair: security and governance spend can accelerate while the semiconductor ETF carries export-control and AI-capex-duration risk. Reassess if SMH outperforms PANW by more than 10% after a policy announcement without corresponding estimate cuts.
• Reduce concentrated exposure to China-sensitive semiconductor equipment until next earnings disclosures clarify service backlog and China mix; ASML, AMAT, LRCX and KLAC are more exposed to policy-driven revenue discontinuities than to a gradual AI-demand slowdown. Cover the underweight if China revenue is retained through compliant-node demand and management raises full-year guidance.
• Watch MSFT, GOOGL, AMZN and META for a divergence between AI capex guidance and disclosed AI revenue/usage metrics over the next two quarters. If capex remains elevated while monetization lags and mandatory evaluation requirements emerge, use puts or relative shorts versus equal-weight software rather than outright shorting into continuing AI demand.