As AI leaders warn of catastrophe, US and China shun slowdown calls

US President Donald Trump and Chinese President Xi Jinping are expected to agree only limited AI-safety cooperation at their Washington summit, falling well short of industry calls for a coordinated slowdown in frontier AI development. US-China strategic rivalry, mistrust over verification, and Beijing's concerns that safety rules could mask chip and market-access restrictions remain major obstacles. Potential initial steps include a bilateral dialogue and a national-security threat notification mechanism, but no agreement is expected on binding capability limits or restrictions on recursive self-improving AI.

Analysis

The investable implication is not AI-safety policy itself, but a higher probability that the AI arms race remains capital-intensive rather than shifting toward a regulated, slower deployment cycle. That supports sustained accelerator, networking, power and data-center demand over the next 6-18 months; NVDA retains the cleanest operating leverage because incremental cluster spending reaches its revenue line before application-layer monetization is proven. META is directionally supported by freedom to deploy models, but its near-term P&L remains exposed to rising depreciation and inference costs before ad-productivity gains are fully measurable.

The non-obvious downside is that a narrowly framed security dialogue could make hardware controls more enforceable, not less restrictive. Any agreement around model testing, incident reporting, or national-security thresholds creates a policy architecture that can be paired with compute-tracking requirements and tighter end-use rules. That would be negative for NVDA's addressable market at the high end and positive for Chinese substitution efforts, even if global aggregate AI capex remains strong.

Consensus is likely too focused on a binary "cooperation equals de-risking" outcome. Limited coordination can reduce tail-risk headlines while preserving the competitive incentive to overbuild capacity, which is constructive for semiconductor revenue but ultimately raises the risk of a 2027 utilization and pricing reset. The key distinction is whether hyperscaler capex guidance continues to rise faster than AI revenue disclosures; if it does, NVDA can outperform for another 1-3 quarters, while META's multiple becomes increasingly dependent on demonstrated monetization rather than strategic narrative.

Key Decisions for Investors

• Maintain or add a 1-3 month tactical long NVDA only on post-summit weakness, rather than chasing a headline move. Thesis: continued frontier competition sustains accelerator orders; invalidate if management commentary or channel checks indicate a material reduction in China-related demand or hyperscaler capex guidance rolls over. Use a 8-10% stop from entry and target 15-20% upside.
• Express the supplier-versus-spender divergence through long NVDA / short META in equal dollar risk over the next two earnings cycles. NVDA captures incremental infrastructure spend sooner, while META must prove that higher AI depreciation and inference expense convert into advertising revenue; cover the short if META reports AI-driven ad pricing or conversion gains that exceed incremental infrastructure-cost growth.
• Do not treat a bilateral safety announcement as an all-clear for NVDA's China exposure. Set an event alert for language on compute reporting, model-capability thresholds, end-use verification, or joint security standards; these provisions would raise the probability of tighter export-control enforcement and warrant reducing NVDA gross exposure even if the initial equity reaction is positive.
• Watch META's 2027 capex and depreciation commentary over the next 1-3 months. A further capex increase without a corresponding improvement in ad-load, pricing, engagement, or operating-margin guidance is a more actionable bearish catalyst than the policy headlines themselves.