Superintelligence Strategy is a March 2025 national-security paper that applies Cold War deterrence theory to advanced AI. It proposes a three-part framework for great-power AI policy built on deterrence, nonproliferation of AI chips, and domestic competitiveness, and frames its recommendation as anti-Manhattan-Project: it argues states should refrain from racing to AGI rather than try to win such a race.
Provenance
The paper is authored by:
- Dan Hendrycks — Director, Center for AI Safety
- Eric Schmidt — former Google CEO/Chairman
- Alexandr Wang — founder/CEO Scale AI (now Meta CAIO)
It was published March 5, 2025 at https://www.nationalsecurity.ai/, with an expert version on arXiv at https://arxiv.org/pdf/2503.05628. It appears in two formats: a standard version of roughly 10 pages presented as a condensed policy brief, and an expert version of roughly 32 pages containing the extended analysis.
Argument
The paper holds that rapid AI advances reshape national security, stating that "destabilizing AI developments could rupture the balance of power." Against that premise it proposes a three-part framework analogous to Cold War nuclear strategy: deterrence via MAIM (Mutual Assured AI Malfunction), nonproliferation of AI chips, and competitiveness in domestic AI chip manufacturing.
| Pillar | Core proposal |
|---|---|
| Deterrence | Espionage and sabotage against destabilizing AI projects |
| Nonproliferation | Chip location tracking; denial to rogue actors |
| Competitiveness | Domestic chip manufacturing (CHIPS Act alignment) |
MAIM
MAIM is described as:
"A deterrence regime resembling nuclear mutual assured destruction (MAD) where any state's aggressive bid for unilateral AI dominance is met with preventive sabotage by rivals."
The paper argues that mutual awareness that aggressive racing invites sabotage is meant to produce stability, by analogy to how MAD stabilized nuclear postures during the Cold War. The nonproliferation pillar pairs chip location tracking with denial of chips to rogue actors, and the competitiveness pillar ties domestic chip manufacturing to alignment with the CHIPS Act.
Positioning
The paper presents itself as the first high-profile attempt to import Cold War deterrence theory into AI policy, and it proposes sabotage as a legitimate counter-move in how states treat rival frontier AI programs. Its framing is deliberately anti-Manhattan-Project, arguing states should "don't race to AGI" rather than "win the race." On posture it contrasts with Situational Awareness: The Decade Ahead, which implied US nationalization and racing; the two share the premise of fast frontier AI progress.
Relationships
- contradicts: Situational Awareness: The Decade Ahead on posture; both share the premise of fast frontier AI progress.
- supports: Export Controls (AI), Compute Governance, Chip Smuggling and Export-Control Evasion.
- depends-on: Semiconductor Supply Chain.
- related: CSIS — DeepSeek, Huawei, Export Controls, and the Future of the U.S.-China AI Race (Allen, March 2025), CSET + DOJ + CSIS — US-China AI National-Security Axis (composite source summary).
Sources
- Primary: Superintelligence Strategy