STRATEGYOngoing practice92% confidence

The Containment Problem

AI is omni-use — containment requires choke points, not just regulation

Problem it solves

Why AI cannot simply be regulated away and what containment would actually require

Best for

Understanding the structural limits of AI regulation; framing long-term governance and investment risk; analysts, policymakers, and founders building in the AI infrastructure layer

Not ideal for

Timing specific regulatory events or predicting particular legislation outcomes

Overview

Why this framework exists

The Containment Problem framework, developed by Mustafa Suleyman and codified in his 2023 book 'The Coming Wave,' holds that advanced AI is structurally different from every previous dangerous technology because it is 'omni-use' at the model level. The same neural network that identifies cancerous tumors in chest X-rays can identify targets for aerial strikes — and the commercial incentives to deploy it are identical in both cases. This dual-use property at the core capability level, combined with the ease of digital distribution, makes voluntary restraint near-impossible and traditional regulatory containment structurally inadequate.

The framework maps the race condition driving proliferation: every actor — nation, corporation, or individual — faces the same calculus. If they do not acquire the capability, a rival will, putting them at a disadvantage. This creates a self-fulfilling dynamic where intelligence becomes the new form of capital, replacing land and oil as the primary resource grab. Unlike nuclear weapons, which required physically scarce uranium and expensive enrichment infrastructure, AI capability can be downloaded. Open-source reproductions of frontier models are already available, 60–70x smaller than originals, and cheaply runnable on consumer hardware.

Suleyman's framework is not purely pessimistic — it maps ten specific intervention points (choke points, audits, international bodies, corporate structure reform, cultural shift) that could together constitute genuine containment. However, his honest assessment is that containment 'must be possible' rather than 'is possible,' because the incentive structure and election-cycle short-termism of democratic governments make proactive restraint historically unprecedented. The framework is most useful as a diagnostic tool: it explains why regulation consistently undershoots the problem and what would need to be true for governance to catch up.

Core principles

5 total
  1. AI is omni-use at the model level — any capability that creates commercial value also creates dual-use risk, making voluntary restraint structurally irrational for any single actor
  2. Digital distribution eliminates the physical constraints that made nuclear nonproliferation feasible — you cannot apply the uranium scarcity model to software weights
  3. The race condition is self-reinforcing: every actor's rational response to rivals building AI is to build faster, which accelerates the dynamic for everyone
  4. Meaningful containment requires physical choke points (chip supply chains, undersea cables), not just legal frameworks, because regulatory arbitrage will always undercut pure legal approaches
  5. Historical precedents show that global coordination only happens after a catastrophic event or when mutually assured destruction is credible — proactive coordination is the unprecedented exception, not the rule

Steps

5 steps
  1. Map the omni-use vector
    Identify the specific capabilities in a given AI system that are simultaneously commercially valuable and potentially harmful. The same capability will drive both commercial adoption and misuse — containment efforts that focus only on the harmful application will be circumvented by the commercial incentive. Understanding the omni-use vector tells you where governance pressure will consistently fail.
    Pro tipStart with the most commercially valuable capability, not the most obviously dangerous one — that is where the race condition will be strongest and containment will be hardest.
  2. Audit the physical choke points
    Identify which parts of the AI supply chain have genuine physical constraints: GPU fabrication (TSMC, NVIDIA), undersea cable infrastructure, data center power, rare earth materials. These are the only points where containment policy can have durable effect, because they cannot be arbitraged away by moving to a different jurisdiction.
    WarningLegal choke points (export controls, regulations) are necessary but insufficient — determined actors will route around them. Only physical constraints create durable leverage.
  3. Stress-test with the regulatory arbitrage scenario
    For any proposed governance mechanism, run the scenario: what happens when a company or nation-state moves to a low-regulation jurisdiction to continue development? If the mechanism collapses under this scenario, it is a partial measure at best. Genuine containment requires mechanisms that survive regulatory arbitrage — typically physical choke points or international coordination bodies with enforcement power.
  4. Identify what catastrophic trigger would force coordination
    Historically, global coordination on dangerous technologies only happens after a sufficiently catastrophic event (Hiroshima triggered nuclear nonproliferation) or when mutually assured destruction is credible to a small number of parties. Map what the equivalent triggering event might be for AI — and whether governance infrastructure could be built before that event occurs rather than after.
    Pro tipThe 'pessimism aversion trap' — the tendency of elites to recoil from engaging with the risk because it feels unproductive — is the primary cognitive failure mode that delays pre-event coordination. Name it explicitly in governance discussions.
    WarningOptimism bias is socially rewarded in elite circles; engaging with the risk requires accepting personal complicity in building it. This creates systematic underestimation of the governance gap.
  5. Design governance with structural accountability, not just good intentions
    Corporate governance innovations (Public Benefit Corporation status, board-level safety obligations) and international coordination bodies (proposed AI equivalent of UN Security Council plus WTO) create accountability mechanisms that survive leadership changes. Relying on the good intentions of individual founders or politicians is insufficient given election-cycle short-termism and competitive pressure.
    Warning4-year election cycles create structural incentives to capture the $15 trillion AI economic prize and defer governance costs — no individual politician is incentivized to sacrifice near-term economic gains for long-term safety.

Checklist

Saved in your browser

Examples

3 cases
CFC ban — the successful containment case

Chlorofluorocarbons (CFCs) were successfully banned via the Montreal Protocol because viable, cheaper alternatives already existed. The commercial incentive to use CFCs disappeared when the alternatives were ready, making voluntary compliance economically rational. This is the rare case where containment succeeded proactively — before catastrophic ozone depletion became irreversible — because the economic calculus aligned with the governance goal.

OutcomeThe CFC case shows that containment works when commercially viable alternatives exist and the regulated capability is not omni-use. AI governance lacks both conditions: no viable alternative to AI capability exists, and the capability is intrinsically omni-use.
Nuclear nonproliferation — partial success via physical constraints

Nuclear nonproliferation has been 'reasonably successful' — only 9 states have nuclear weapons, and three countries that acquired them (South Africa, Ukraine, Kazakhstan) gave them up due to sanctions and incentives. The key constraint was physical: uranium-235 cannot be downloaded, enrichment requires expensive specialized infrastructure, and the capability cannot be miniaturized below a certain point. Physical scarcity made containment feasible even without universal cooperation.

Outcome9 nuclear states after 80 years is a governance success by historical standards. The lesson for AI: without an equivalent physical constraint, the proliferation dynamic will be far harder to slow. AI weights can be downloaded; nuclear material cannot.
Open-source AI proliferation — the containment failure in progress

GPT-3 class models, which required massive compute clusters when originally trained, are now fully reproduced in open-source form at 60–70x smaller scale and freely available on the web. A capable individual with consumer hardware can run models that were frontier systems 2–3 years ago. Mustafa's projection: within 5 years, a teenager in Russia could download something 'incredibly harmful in the artificial intelligence department' onto a home computer.

OutcomeThe proliferation of open-source AI demonstrates the structural failure of export controls and model access restrictions as primary containment mechanisms — the cutting edge democratizes within 2–3 years regardless of initial access restrictions.

Common mistakes

5 traps
Applying the nuclear nonproliferation model directly to AI
Nuclear containment worked because uranium-235 is physically scarce, expensive to enrich, and cannot be downloaded. AI model weights can be downloaded, reproduced at 60–70x smaller scale, and run on consumer hardware within 2–3 years of frontier development. The physical constraints that made nuclear containment feasible simply do not exist for AI — copying the governance model without adapting for digital distribution will systematically fail.
Treating regulation as sufficient without physical choke points
Regulation without enforcement leverage is a statement of intent. Actors who find a capability commercially valuable will relocate to low-regulation jurisdictions. Effective containment requires control over physical infrastructure — GPU supply chains, cloud API access, undersea cables — that cannot be arbitraged away by moving headquarters.
Falling into the pessimism aversion trap
The instinctive reaction to AI risk — recoiling from the fear, dismissing it as sci-fi, or assuming it will 'all work out' — is most prevalent in elite circles where the people building the technology work. This is a systematic cognitive failure: pessimism feels unproductive, optimism is socially rewarded, and engaging with the risk means accepting personal complicity. The trap produces systematic underestimation of the governance gap.
Assuming voluntary restraint is possible under a race condition
Every actor rationally believes: 'If I don't build it, they will, and then I'll be at a disadvantage.' This is a self-fulfilling belief that makes voluntary restraint near-impossible without binding coordination mechanisms. Governance approaches that rely on individual actors choosing to slow down will be undermined by any competitor who does not make the same choice.
Treating AI governance as a single-country problem
No amount of well-crafted regulation in the UK, EU, or US alone can address the cross-border race condition. Governance that works within one jurisdiction incentivizes development to move elsewhere. Genuine containment requires an international coordination body with enforcement capability — the equivalent of the UN Security Council for AI — which does not currently exist and has no precedent of being built proactively.

Origin story

How this framework came to be

Suleyman co-founded DeepMind in 2010 and spent over a decade building AI systems while simultaneously engaging with governments and international bodies on governance questions. The containment framing emerged from his direct experience of the gap between AI capability development timelines and policy response timelines. He observed that the standard regulatory playbook — wait for harm, legislate reactively — could not work for a technology that generates emergent capabilities faster than legal systems can respond.

The framework was formally articulated in 'The Coming Wave' (2023), co-authored with journalist Michael Bhaskar. Suleyman deliberately structured the book around the tension between two chapter titles: 'Containment Is Not Possible' (opening) and 'Containment Must Be Possible For All Our Sakes' (closing). This framing — not a prediction, but a moral imperative — distinguishes his position from both techno-optimists (who dismiss the risk) and pure doomers (who claim the outcome is fixed). His position as someone who both built and was responsible for governing frontier AI systems gives the framework unusual credibility.

Source

Traced to primary
Source · PODCAST
CEO Of Microsoft AI: AI Is Becoming More Dangerous And Threatening!
Mustafa Suleyman · 2023
Open source →

Related frameworks

Browse all Strategy →