AI Compute and the Limits of Mutually Assured Deterrence

Treating computing power as the strategic equivalent of a nuclear arsenal shifts the AI debate from voluntary restraint to enforceable capability. The comparison is useful because it asks whether competitors can be discouraged from pursuing dangerous advantages even when trust is absent. It is also hazardous if a memorable metaphor is mistaken for a complete doctrine.

The supplied Hindu Post article reports one proposal, AI 2040: Plan A, rather than presenting corroboration from several independent sources or the full report itself. The available material therefore supports an assessment of the proposal’s logic and unanswered questions, not a conclusion that compute-based deterrence is already workable or widely accepted.

What the nuclear analogy is meant to transfer

According to Hindu Post’s summary, the proposal borrows from the Cold War logic of mutually assured destruction. In the source’s account, deterrence between the United States and the Soviet Union did not depend primarily on confidence in benign intentions. It depended on each side retaining enough destructive capacity to retaliate after an attack, making aggression ruinous to the aggressor as well.

Applied to AI, the central insight is that governance must remain effective when competitors distrust one another. Promises of caution are fragile if a party expects a decisive benefit from breaking them. A deterrence framework instead tries to alter the calculation: the expected consequences of crossing an agreed boundary must outweigh the expected advantage.

Compute is attractive as a governance lever because advanced AI development depends on physical computational infrastructure. In the proposal as summarized, Romeo Dean’s compute and verification models give the analogy a quantitative foundation. That is potentially more concrete than trying to regulate intentions, but measurement alone does not create deterrence. A doctrine must still identify the prohibited conduct, determine who can verify it and specify what consequence follows a verified violation.

Compute is not a warhead

The analogy begins to weaken when the character of the underlying assets is considered. A nuclear weapon is itself a destructive instrument. Computing hardware is general-purpose infrastructure whose consequences depend on models, data, software, energy, operators and intended use. The same broad class of resource can support ordinary commercial activity, scientific work or the development and operation of advanced AI.

Compute also does not automatically supply the reciprocal catastrophe on which mutually assured destruction rests. One party’s possession of a large computing base does not, by itself, guarantee that an opponent will suffer an unacceptable cost after violating an AI agreement. The missing bridge is an enforceable response mechanism. Unless the framework explains what retaliatory or corrective capacity survives a violation and why its use would be credible, compute remains an object of competition rather than a deterrent.

The actors are different as well. The source emphasizes the authors’ experience in frontier laboratories, forecasting and hardware, implicitly highlighting that advanced AI is shaped by both governments and private organizations. A stable doctrine would therefore have to allocate obligations across states, laboratories, infrastructure providers and other relevant operators. Nuclear language that treats each country as a single decision-maker can conceal this institutional complexity.

Four tests for a credible compute doctrine

A precise red line

A deterrent cannot operate around an undefined offense. The regulated act might concern possession of hardware, concentration of capacity, a particular training activity, development of a specified capability or deployment in a prohibited setting. These are not interchangeable. A rule aimed at equipment could capture benign uses, while a rule based only on observed model behavior might activate too late. The supplied summary does not provide enough detail to determine where AI 2040: Plan A draws this boundary.

Reliable verification and attribution

Verification must establish more than the existence of computing equipment. It must connect observable evidence to the conduct covered by the agreement and distinguish a violation from permitted activity or measurement error. Attribution matters because consequences imposed on the wrong actor would weaken legitimacy and could intensify competition. The source reports that Dean developed compute and verification models, but the excerpt supplies neither their assumptions nor their error tolerances, so their adequacy cannot be evaluated from the available material.

A survivable and proportionate response

Traditional deterrence relies on the target retaining an effective response after an initial strike. A compute doctrine needs its own answer to survivability: what capacity to impose consequences remains available after a concealed breakthrough or attempted first-mover advantage? It must also avoid forcing every violation into an all-or-nothing response. Proportionate, predictable consequences are generally more credible than threats so extreme that decision-makers would hesitate to carry them out.

Resistance to deception and evasion

Hindu Post reports that Ryan Greenblatt’s work with Anthropic on alignment-faking informs the proposal’s treatment of AI deception. That background makes deceptive behavior a relevant design concern, but it does not prove that a compute-monitoring regime can detect every form of concealment. A robust system would need to consider both human evasion of governance rules and the possibility that advanced models could produce misleading evidence or strategically compliant behavior. Confidence should rest on tests and independent scrutiny, not on the assumption that monitored actors will cooperate with the monitor.

Key takeaways for evaluating the proposal

• The strongest part of the analogy is its distrust-aware logic: rules must influence capabilities and incentives, not depend on goodwill.
• Compute is a measurable input to AI development, but possessing it does not automatically create a retaliatory capability or mutual vulnerability.
• A workable doctrine needs an explicit red line, reliable attribution, survivable enforcement and consequences that decision-makers could credibly apply.
• Dual-use infrastructure and the involvement of private laboratories make AI governance more institutionally complex than a state-to-state weapons model suggests.
• The supplied evidence is one publication’s summary of a proposal; it does not establish independent validation, policy adoption or consensus about feasibility.

From a strategic metaphor to accountable governance

The source presents the proposal through the backgrounds of its authors. It reports that AI Futures Project leader Daniel Kokotajlo left OpenAI’s governance division in 2024 following disagreements over safety priorities; that Eli Lifland leads the RAND Forecasting Initiative’s all-time leaderboard; that Greenblatt is chief scientist at Redwood Research; and that Dean trained in computer science and hardware at Harvard. These credentials explain why forecasting, deception and verification feature prominently in the framework. They do not substitute for disclosure of assumptions, outside evaluation or evidence that governments and companies could implement the system under competitive pressure.

The next analytical step is consequently institutional rather than metaphorical. Any serious plan would need transparent definitions, reviewable measurements, safeguards against conflicts of interest, procedures for contesting findings and a calibrated path from suspected non-compliance to confirmed violation. It would also need to show that monitoring itself does not create dangerous incentives to conceal activity or accelerate development before restrictions take effect.

Compute accounting may become an important component of AI governance, but strategic stability will depend on the rules wrapped around the accounting. The durable question is not whether compute resembles nuclear weapons. It is whether observable limits, legitimate verification and credible consequences can be combined before competition makes mutual restraint harder to sustain.

References

• Hindu Post — Compute is the new nuke: Inside AI’s MAD doctrine