Compute is the new nuke: Inside AI’s MAD doctrine

For four decades, the thing that kept the United States and the Soviet Union from using nuclear weapons was not restraint. It was the guarantee that using them would destroy the user too. Neither side needed to trust the other's intentions because intentions stopped mattering once both had enough warheads to survive a first strike and retaliate. Deterrence rested on symmetry of catastrophe, not goodwill.

A new report on AI governance, AI 2040: Plan A, borrows that logic wholesale and applies it to something that has never needed a doctrine before: computing power.

The report comes from the AI Futures Project, the group behind the widely read AI 2027 scenario, and the authors' backgrounds matter because the proposal depends heavily on forecasting and technical credibility. Daniel Kokotajlo, who leads the project, left OpenAI's governance division in 2024 over disagreements about the company's safety priorities, giving him firsthand knowledge of how a frontier lab operates. Eli Lifland sits atop the RAND Forecasting Initiative's all-time leaderboard, a position earned through audited forecasting tournaments rather than résumé claims. Ryan Greenblatt is chief scientist at Redwood Research, where his published work with Anthropic on alignment-faking informs much of the report's treatment of AI deception. Romeo Dean, trained in computer science and hardware at Harvard, built the compute and verification models that turn the report's central metaphor into arithmetic instead of rhetoric.

More from Opinion

That metaphor is mutually assured compute destruction.

Under the proposal, the United States and China would build their newest AI datacentres not on home soil but in territory least defensible against the other's military: China's in Canada and America's in Mongolia. The logic mirrors Cold War second-strike doctrine. If the agreement collapses, each side destroys its own compute before the other can capture it, just as a nuclear power would rather lose a weapon than allow an adversary to seize it. Dean's modelling gives the idea weight. Total AI compute grows from roughly 20 million chip equivalents in 2026 to 60 billion by 2034, and the arrangement works only if that compute can be destroyed on demand rather than merely declared off-limits.

The comparison to nuclear deterrence goes further than metaphor because the doctrine aims to deter something different. Nuclear MAD tries to prevent weapons from being used. Compute MAD accepts that AI development will continue. Its purpose is to stop either side from pulling ahead in secret.

The report argues that the real danger is not a single catastrophic launch but the gradual concentration of power in whichever country, company, or individual reaches superintelligence first without meaningful oversight. Its answer is radical transparency in research. AI development is divided into research, which becomes visible across borders, and inference, which remains private. Neither government needs to trust the other's intentions. Both verify the other's chips.

The report strengthens its case by modelling four ways the agreement could fail to materialise and comparing its proposal against each of them. One scenario lets the United States race toward an intelligence explosion. Another slows American development in the hope that China follows. A third assumes China secretly stockpiles smuggled compute inside a hydropower tunnel. The last halts frontier AI research altogether.

Each scenario receives its own chapter, its own estimate of catastrophe, and its own assessment of how power concentrates among governments or AI companies. Plan A emerges as the preferred option not because it is safe but because it performs better than the alternatives the authors actually modelled and stress-tested. That reflects both Larsen's advocacy background and Kokotajlo's experience inside OpenAI. They treat governance the way engineers treat a bridge design: look for the failure points before anyone has to drive across it.

The consequences extend well beyond deterrence because any compute-sharing agreement has to survive the economic disruption advanced AI creates. Payroll and income tax shrink as AI replaces human labour. Corporate tax raises less revenue because capital expensing shields much of firms' profits even as those firms absorb a growing share of the economy.

The report's solution is to cap and trade compute and robot production much like carbon permits, distributing the proceeds as a citizen's dividend that begins around $45,000 per year and rises into the millions within a decade. The authors do not present this as an inevitable future. They present it as the political price of making a compute-deterrence regime sustainable after AI has transformed the economy around it.

The same logic pushes the agreement beyond Washington and Beijing. A bilateral arrangement gives every other country an incentive to build hidden compute stockpiles if either superpower appears close to breaking away. The report's Consortium model expands membership, along with permit revenue, to middle powers and the Global South in an effort to remove that incentive.

For countries such as India, this is not a great-power negotiation happening elsewhere. The report treats their participation as necessary because a deterrence system that excludes major players becomes less stable, not more.

The report's own footnotes complicate its confidence in ways worth taking seriously. The six authors disagree sharply over the probability of a misaligned AI takeover, with estimates ranging from roughly 20 per cent to 70 per cent. They also paid a former collaborator to publish a public critique of the report alongside its release.

Few advocacy documents invite criticism so directly, and that openness increases confidence in the authors' intellectual honesty. It does not solve the report's central weakness. The technical modelling behind compute destruction, developed by researchers with genuine forecasting and AI control credentials, is considerably stronger than the geopolitical assumption beneath it: that Washington and Beijing would voluntarily negotiate this agreement before a crisis forces the issue.

Nuclear deterrence was never a blueprint for peace. It was a strategy for surviving the absence of peace, grounded in the physical reality that warheads could be counted and destroyed.

Plan A wagers that compute can play the same role. If chips can be verified, counted, and destroyed, then deterrence becomes credible without requiring trust between rivals. Whether that wager succeeds depends less on the elegance of the doctrine than on whether two governments that struggle to cooperate on almost anything else will accept an international system of chip inspectors.

The report does not solve that problem. It is the first proposal detailed enough to show exactly where it begins.

(Sreejith Sreedharan is a technology analyst and author of ‘Future of Work – AI Augmented Autonomous Decentralised’. He works on organisational AI readiness and created the AI Instinct Index®, a psychometric diagnostic designed to assess behavioural readiness for AI adoption and adaptive capacity in constraint-heavy environments. Views expressed in the above piece are personal and solely those of the author. They do not necessarily reflect Firstpost’s views.)