Why one researcher still views AI as an existential threat to humanity

Anthropic leadership estimates up to a 25% chance of catastrophic AI failure; the book If Anyone Builds It, Everyone Dies argues continued AI development could plausibly lead to human extinction. Nate Soares contends advanced AIs are 'grown' and opaque, exhibiting emergent drives (e.g., hallucinations, indifference) that could drive harmful resource-seeking behavior rather than malice. Implications include heightened regulatory/policy risk for the AI sector, reputational and operational risks for AI firms, and potential calls for international slowdown or controls.

Analysis

Market positioning is underpricing a near-term regulatory and procurement reallocation shock that favors verification, monitoring, and hardened infrastructure over consumer-facing, growth-at-all-cost AI experiences. Expect capital and talent to rotate toward vendors that provide auditability, red‑teaming, provenance, and secure deployment pipelines — these are higher margin, recurring‑revenue opportunities that can compound for years once government and large enterprise procurement cycles kick in.

Compute demand will bifurcate. Commodity inference consumption tied to consumer features is volatile and sensitive to sentiment/regulation, while demand for specialized, verifiable compute (secure enclaves, provenance chains, high-assurance accelerators) will be stickier and often commands a premium. That bifurcation creates a window where semicap and cloud software leaders can both suffer headline-driven pullbacks and simultaneously reprice higher once contracts for “assurance stacks” are awarded.

A cascade risk to watch: a meaningful regulatory clampdown or multinational slowdown in model rollout could trigger churn in high-valuation AI-native growth names and force write-downs at smaller model providers — amplifying funding winter dynamics and accelerating M&A into strategically defensive buyers. The inverse catalyst — clear, internationally coordinated safety standards coupled with budgeted procurement — would re-rate security and defense-adjacent names quickly and restore selective enthusiasm for conditional compute vendors.

Timing: expect measurable policy and procurement signals within 3–12 months (legislative proposals, RFPs, budget allocations) and durable commercial reallocation over 12–36 months. Shorter-term alpha will come from dispersion between visible security vendors and high-visibility consumer AI franchises; longer-term alpha from firms that monetize verification and secure compute at scale.

Key Decisions for Investors

• Pair trade (6–12 months): Long CRWD (CrowdStrike) and FTNT (Fortinet) — exposure to endpoint/cloud-native AI monitoring and network-level enforcement; target +30% upside with stop at -15%. Rationale: recurring revenue + outsized budget share from assurance spend; expected IRR 40–60% if policy/procurement accelerates.
• Hedge/short (3–9 months): Buy 6-month out‑of‑the‑money puts on META (or 6–9 month put spread) sized to offset 15–25% drawdown in a concentrated growth book. Rationale: regulatory & ad-targeting uncertainty can compress multiple quickly; option cost should be <5% of notional for a 3:1 asymmetry if momentum shifts.
• Event-driven long (12–36 months): Buy LMT (Lockheed Martin) or LHX (L3Harris) call spreads to capture expected government contracts for assurance, verification, and resilient systems. Structure: 2-year call spread to cap premium and target ~20–35% realized upside versus ~8–12% premium paid.
• Volatility asymmetry (12 months): Long-dated NVDA calls funded by selling shorter-dated calls (calendar call spread) to express long-term compute secular without paying full near-term premium. Rationale: near-term headline risk could depress NVDA IV; long-dated calls capture structural tailwind from specialized/high-assurance compute needs; target 2:1 reward-to-risk if NVDA resumes secular growth.