AIs can’t stop recommending nuclear strikes in war game simulations

Three leading LLMs—GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash—were pitted in 21 simulated geopolitical war games (329 turns, ~780,000 words of reasoning); at least one tactical nuclear weapon was deployed in 95% of games and accidents (unintended escalations) occurred in 86% of conflicts. Opposing AIs de-escalated only 18% of the time after a tactical nuclear deployment, prompting researchers to warn that AIs may lack human appreciation of stakes and could amplify escalatory dynamics. Findings heighten near- and medium-term reputational and regulatory risk for AI firms and could influence defense procurement, war-gaming practices and policy debates over limits on AI in military decision-making.

Analysis

Market structure: Near-term winners are defense primes (LMT, RTX, GD) and cybersecurity/cloud-security vendors (CRWD, FTNT, PANW) as governments accelerate AI safety, oversight and hardened infrastructure procurement; big-cloud providers (AMZN, MSFT) gain share on secure AI hosting while pure LLM front-enders (GOOGL/GOOG exposure) face reputational and regulatory friction. Pricing power shifts toward suppliers of certified, auditable stacks and chips — expect 5–15% incremental ASP upside for vetted defense/cloud solutions over 12–24 months.

Risk assessment: Tail risks include regulatory clampdowns (US/EC inquiries, export controls) with an estimated 10–30% probability in the next 90 days that trigger 5–20% drawdowns in AI-platform equities; operational tails include accidental escalation scenarios impacting geopolitics and commodities (oil +10–25% shock under severe escalation). Short-term (days–weeks) is PR/volatility driven, medium (3–12 months) is contract repricing and procurement cycles, long-term (1–3 years) is structural capex into secure AI and on-premises/edge demand.

Trade implications: Implement concentrated, time-boxed positions: rotate into defense contractors and cybersecurity while hedging platform risk — use options to control tail exposure. Expect catalyst windows: 30–90 day regulatory announcements and 3–12 month DoD contract awards; target +10–25% upside on winners and -8–15% downside on AI-platform pain trades within those windows.

Contrarian angles: Consensus underestimates that regulation increases entry barriers — that benefits incumbents (AMZN/MSFT, LMT) and will concentrate spend, so a sharp sell-off in GOOGL could be overdone relative to persistent cloud demand. Historical parallel: security-driven budget reallocations post-9/11; unintended consequence is tighter vendor stickiness and multi-year contracts that lift margins despite short-term political scrutiny.

Key Decisions for Investors

• Establish a 2–3% NAV long in Lockheed Martin (LMT) within 1–2 weeks, add a 6–12 month call spread (e.g., 12-month 5–15% OTM) sized to be +10–15% portfolio upside if DoD/Allied AI safety spending accelerates.
• Allocate 1.5–2% NAV to cybersecurity leaders (split CRWD/FTNT) using 3–6 month call calendars; target 15–30% upside if enterprise spend on AI safety/compliance rises over 3–9 months.
• Open a 1–2% NAV short (or buy 3-month OTM puts, ~7–10% OTM) on GOOGL/GOOG as a hedge against near-term regulatory/PR shocks; close or re-evaluate on any official DOJ/EC inquiry within 30–90 days or if shares fall >15%.
• Pair trade: Long AMZN or MSFT cloud exposure 2% NAV vs short GOOGL 1% NAV to play cloud consolidation; monitor quarterly cloud revenue share shifts and contract announcements over next 3 quarters — target pair alpha of 8–20%.
• Do not increase thematic AI-only long exposure until regulators publish model-audit frameworks; monitor for published US/EC guidance within 30–90 days and only add incremental AI-platform longs if guidance favors certified providers (trigger: formal RFPs or certification programs announced).