Mutually Assured Compute Destruction

Mutually Assured Compute Destruction (MACD) is the deterrence mechanism proposed in AI 2040 Plan A for stabilizing an international slowdown in frontier AI development. The idea is to make strategic AI compute mutually vulnerable so that no major power can defect from the agreement and retain a decisive compute advantage.^{source: ai-2040-plan-a-2026.md}

Under the proposal, major powers place new frontier datacenters in third countries where rival powers could seize or destroy them if the treaty collapses. Hardware also depends on continuing authorization from participating powers. This turns ai-compute-infrastructure into a hostage-like commitment device: the shared compute base is useful while cooperation holds, but cannot be captured intact by one defector.

MACD is intended to make capability scaling more reversible than publishing new algorithms. Consortium members can collectively expand monitored compute for safety work and economic use while preserving a credible way to prevent that infrastructure from powering a unilateral intelligence explosion.

Strategic logic

Deterrence: Defection should not yield exclusive control of the shared compute base.
Verification: Declared chips, datacenters, energy use, and workloads are monitored across borders.
Reversibility: Physical infrastructure can be disabled; widely published algorithmic knowledge cannot be recalled.
Diffusion: Multiple countries and firms can access frontier-scale systems without trusting a single state or corporation.

Risks

The same mechanism creates severe tail risks. Treaty failure could destroy trillions of dollars of infrastructure, interrupt civilian services, and coincide with military escalation. If destruction fails, the enlarged compute base could instead enable a faster intelligence explosion. The regime also requires intrusive inspection, trusted hardware, international coordination, and confidence that covert compute remains too small to matter.

The concept is highly speculative and has no demonstrated real-world implementation at the scale proposed. It should be read as a design for scenario scrutiny, not an established governance technology.

Resources

AI 2040: Plan A