On March 5, 1957, the United States signed a nuclear cooperation agreement with Iran. The deal was part of President Eisenhower's “Atoms for Peace” program, a Cold War initiative designed to share civilian nuclear technology with allied nations under American supervision. The theory was elegant: if nuclear capability was going to spread regardless, better to spread it through a managed architecture with monitoring, safeguards, and relationship-building than to watch it proliferate through back channels. Iran would get a research reactor. The US would get influence, oversight, and a strategic partner in a volatile region.

Twenty-two years later, the Shah was gone, the American embassy was under siege, and the reactor we'd provided became the seed of a nuclear program we've spent the past four decades trying to contain. The sanctions regime, the Stuxnet cyberattack, the JCPOA negotiations and their collapse, the ongoing enrichment standoff — all of it traces back to a technology transfer that made perfect strategic sense at the time and appears catastrophic in retrospect.

Any other paper on AI export controls would use this story as a cautionary tale about technology transfer. Don't give away strategic capability. Don't enable future adversaries. The obvious lesson writes itself.

But the obvious lesson is wrong, or at least incomplete. Because during roughly the same period, the United States pursued a parallel nuclear strategy with dozens of other countries — the same technology sharing, the same civilian reactors, the same potential for weapons development. Most of those relationships didn't end in four decades of containment struggle. The difference wasn't the technology, but rather the governance architecture surrounding it.

We find ourselves now making the same category of decision about artificial intelligence, and we seem to be making it without a clear theory of which model we're following.

I.Two models for managing dangerous technology

The history of nuclear non-proliferation offers two distinct strategic approaches, and the US has used both.

The first is denial: export controls, sanctions, military containment. You identify the chokepoints through which capability flows and you close them. You make the components hard to acquire, the expertise hard to transfer, the finished product impossible to deploy without consequences. This is the model we've applied to Iran since 1979, to North Korea since the 1990s, to various other actors through the Wassenaar Arrangement and related regimes. It's expensive, it requires constant vigilance, and there's evidence that it works — at least for a while.

The second is governance: treaties, monitoring, managed sharing. You accept that capability will spread and focus on shaping how it spreads. You create international institutions with inspection authority. You establish norms about acceptable use. You build relationships with the technologists and policymakers in other countries so that when hard decisions arise, you have channels and leverage. This is the model that produced the Nuclear Non-Proliferation Treaty (NPT), the International Atomic Energy Agency (IAEA), and a world where dozens of countries have the technical capacity to build nuclear weapons and have chosen not to.

The obvious reading is that denial is for adversaries and governance is for allies. But that's an overly simplistic view. The NPT framework includes countries that are geopolitical competitors. The governance model worked because it created something more durable than bilateral relationships: an architecture that countries bought into because it served their interests, with monitoring that created transparency and accountability.

The question then becomes which model we're applying to AI, and the apparent answer is that we're attempting a path of denial without governance. We have export controls. We don't have anything resembling an international architecture for AI development.

II.The China parallel (or “we've done this before”)

The pattern with China follows the Iran template with a longer runway and higher stakes.

In 2001, the United States supported China's entry into the World Trade Organization. The theory was, again, elegant: integrate China into the global economic system, enable their development, and the resulting prosperity would create a middle class that demanded political liberalization. Trade would be the governance architecture. Economic interdependence would align interests.

Part of that integration involved technology transfer. American companies wanted access to Chinese markets. Chinese policy required joint ventures and knowledge sharing. For two decades, this was treated as acceptable cost of doing business. Semiconductor manufacturing, AI research, telecommunications infrastructure — all of it flowed through joint venture relationships.

Then the relationship context shifted. Not as dramatically as the Iranian revolution, but materially. Competition replaced partnership as the dominant frame. And suddenly the technology transfers that had seemed like smart business began to look like strategic errors.

The United States' response has been denial. Export controls on advanced chips. Restrictions on ASML lithography equipment. Investment screening for AI companies. Visa limitations for certain research fields. The explicit goal is to maintain a capability gap — to keep US and allied AI development ahead of Chinese development by restricting access to the compute infrastructure that enables frontier models.

This is a coherent strategy, and it may even work for a while. The question is: for what purpose? What are we building while we buy time?

III.The chokepoint addiction

There's a pattern in American strategic thinking that shows up across domains: find the narrow channel through which the thing you care about flows, and then find a way to control that channel.

For oil, it was the Strait of Hormuz. Twenty percent of global petroleum passes through a waterway twenty-one miles wide at its narrowest point. American naval presence in the Persian Gulf was designed to guarantee that flow — not because we needed the oil ourselves (US import dependency has declined) but because our allies did and because the global economy required it. The strategic architecture we built around oil assumed permanent chokepoint management.

For semiconductors, the chokepoint is Taiwan. TSMC manufactures over ninety percent of the world's most advanced chips. The fabs are concentrated on an island that China claims as territory and has repeatedly signaled willingness to take by force. American semiconductor strategy has been to defend this chokepoint — through diplomatic ambiguity, military posturing, and alliance maintenance — while slowly building domestic alternatives through the CHIPS Act.

For AI capability specifically, the chokepoint strategy has an additional layer: not just Taiwan's fabrication, but the Netherlands' lithography. ASML is the only company in the world that makes the extreme ultraviolet lithography machines required for cutting-edge chip manufacturing. There are no alternatives. The machines cost $380 million each and take years to build. US export control strategy has effectively deputized a Dutch company as an enforcement mechanism for American technology policy.

The pattern appears to work, until it doesn't. Hormuz works until Iran mines it or missile coverage makes transit uninsurable. Taiwan works until it doesn't. ASML works until someone develops an alternative or finds a way to achieve equivalent capability through different means.

Which brings us to the diffusion problem.

IV.The diffusion clock

In January 2025, a Chinese AI lab called DeepSeek released a model that performed at or near the level of frontier American systems at a fraction of the training cost.

The immediate response in American AI circles was skepticism — claims about training costs are hard to verify, and there were questions about whether DeepSeek had used distillation from American models as a shortcut. But the skepticism somewhat missed the real point. Even if the efficiency claims were overstated, the capability was real. Even if distillation played a role, the architectural innovations were genuine. The assumption underlying export controls — that hardware constraints would bottleneck capability development — was being tested.

Call this the diffusion clock, or the time between when you restrict a capability and when the capability becomes achievable through alternative means. For nuclear weapons, the clock ran decades because the physics was understood but the engineering was extraordinarily difficult. For AI capability, the clock may run faster because the limiting factor has been compute rather than fundamental knowledge, and compute constraints can be routed around through efficiency improvements, algorithmic advances, and massive scale-up of lower-capability hardware.

The irony is that American AI policy is currently running two contradictory strategies simultaneously. On hardware, we're pursuing denial: restrict NVIDIA chips, restrict ASML equipment, restrict semiconductor manufacturing know-how. On software, we're permitting diffusion: Meta's Llama models are open-source, research papers are published openly, the algorithmic techniques underlying frontier models are widely shared.

The theory seems to be that compute is the binding constraint and therefore compute is what should be controlled. But if DeepSeek suggests anything, it's that the binding constraint may be shifting. Efficiency gains are being substituted for raw compute, algorithmic improvements are compounding, and the diffusion clock is running faster than the denial strategy assumed.

V.Governance debt

There's a concept in software engineering called “technical debt”, which is the accumulated cost of shortcuts and deferred maintenance that compounds over time until the system becomes unmaintainable. A similar phenomenon exists in strategic technology management.

Governance debt is the gap between technological capability and the institutional architecture required to manage it safely. Every year that capability advances without corresponding governance development, the debt compounds. And like technical debt, governance debt is easier to ignore in the short term and catastrophic in the long term.

For nuclear technology, we paid down governance debt through decades of institution-building. The IAEA was established in 1957. The NPT was signed in 1968. The Comprehensive Nuclear-Test-Ban Treaty process began in the 1990s. None of these institutions were perfect, and the debt wasn't fully paid. But there's an architecture, there are norms and there are mechanisms for monitoring, verification, and consequence.

For AI, the governance debt is accumulating rapidly. Capability is advancing faster than any previous technology transition. The institutions that might manage it are nascent at best. The EU AI Act is a regulatory framework for one jurisdiction, not a global governance architecture. The Bletchley Declaration was a statement of intent, not a true enforcement mechanism. The US AI Safety Institute is still finding its footing. China has its own AI governance framework that's not interoperable with Western approaches.

Meanwhile, the denial strategy is being used to try and buy time. Export controls on semiconductors don't solve the governance problem, they simply delay the point at which we have to confront it. The question is whether we're using that time to build governance architecture or simply accumulating more debt.

The Iran analogy cuts both ways here. We're focused on the failure case: technology transfer that enabled a threat. But the success case is equally relevant: governance architecture that prevented dozens of other countries from going down the same path. The NPT wasn't a denial strategy. It was an agreement that offered something in exchange for restraint — access to civilian technology, security guarantees, membership in an international order. Countries that could have built nuclear weapons chose not to, in part because the alternative offered was more attractive than the isolation that pursuit would bring.

What's the equivalent offer for AI? What governance architecture makes restraint more attractive than racing? That question doesn't have an answer yet, and nobody in a position of authority seems to be working on one.

VI.The shape of the problem

The nuclear parallel clarifies the stakes, but it also obscures an important difference. Nuclear weapons have a relatively clear threshold: you either have them or you don't. The capability is binary in ways that matter. AI capability is continuous and multipurpose. There's no obvious line between civilian and military application, between beneficial and dangerous use, between a system that helps with research and one that poses strategic risk.

This makes the governance problem harder in some ways and easier in others. Harder because you can't simply monitor for enrichment facilities or count warheads. There's no equivalent of a “nuclear test” that announces capability. But it's also easier because the same continuity that makes bright lines difficult also means there's no single catastrophic threshold. The failure modes are more diffuse, which means the opportunities for intervention are more distributed.

The denial model assumes you can hold a line. The governance model assumes you can shape a trajectory. For nuclear technology, holding a line made sense because the capability was discrete and the consequences were existential. For AI, shaping a trajectory may be the only viable strategy because the capability is continuous and the consequences are pervasive.

But we're not doing that either. We're holding lines on hardware while letting trajectories run unshaped on software, norms, and institutional architecture. We're buying time without a theory of what we're buying it for.

VII.What the parallels suggest

The Strait of Hormuz took decades of naval architecture to secure. We built bases, rotated carrier groups, developed doctrine, cultivated relationships with Gulf states, and created a system that could guarantee oil flow through a chokepoint we didn't control. It was expensive, it required constant maintenance, and it worked — until the threat model evolved beyond what the architecture was designed for.

Taiwan's semiconductor concentration is a vulnerability we've known about for years and are only now beginning to address. The CHIPS Act is a start. The new fabs will take most of a decade to reach production. And in the interim, we depend on a chokepoint we can defend militarily but cannot replace industrially.

The AI chokepoints (advanced chips, lithography equipment, training compute) are defended through export controls rather than physical presence. The enforcement mechanism is regulatory rather than military. But the underlying logic is the same: control the narrow channel, maintain the gap, buy time.

The question that echoes across all three domains is what you do with the time you buy. For oil, we used the time to develop alternatives (eg. shale and renewables) that reduced dependency. For semiconductors, we're using the time to build domestic capacity. For AI, we're using the time to… what exactly?

The nuclear precedent suggests an answer: build governance architecture that outlasts the denial strategy. Create institutions that provide transparency. Develop norms that make certain behaviors costly. Offer something in exchange for restraint so that racing isn't the only rational strategy.

None of that is happening. The AI governance conversation is fragmented across jurisdictions, captured by industry in some domains and ignored by policymakers in others, and proceeding at a pace wildly mismatched to the capability development it's supposed to govern.

VIII.The receding tide

There's a problem with the nuclear analogy that the preceding sections haven't confronted: the institutions that made nuclear governance possible were products of a specific historical moment. And by all appearances, that moment has passed.

The NPT was negotiated when the United States had both the power and the inclination to build multilateral architecture. The IAEA was established when international institutions were understood as force multipliers for American interests rather than constraints on American action. The entire edifice of arms control treaties, inspection regimes, and managed sharing emerged from a postwar consensus that collective security arrangements served everyone's interests, including the hegemon's.

That consensus is unraveling. Not slowly, not subtly, but actively and explicitly. The country that built the architecture is now dismantling it.

The Iran nuclear deal is instructive here in a way that cuts against the framing I've offered. The JCPOA was, for all its flaws, a governance achievement: a multilateral agreement that constrained Iranian enrichment through monitoring, verification, and economic incentives. It wasn't a denial strategy. It was architecture. And it collapsed not because Iran cheated or because the inspections failed, but because the United States withdrew. The country that had spent decades building governance capacity decided that constraint wasn't worth the cost of commitment.

This isn't an isolated case. The pattern repeats across domains. Trade governance through the WTO has been paralyzed since the US blocked appellate body appointments. Climate governance through the Paris Agreement has been subject to withdrawal and re-entry based on domestic political cycles. Health governance through the WHO has been defunded. The institutional infrastructure that would be required to build AI governance architecture at global scale is being hollowed out at exactly the moment when the challenge demands it.

The optimistic read is that this is a temporary aberration — a political moment that will pass, after which institution-building can resume. The pessimistic read is that the postwar period was the aberration, a historically unusual window of multilateral capacity that has now closed. The realist read is somewhere between: the institutions won't disappear entirely, but they'll operate at reduced capacity, with reduced legitimacy, and with reduced American commitment.

Any of those reads changes the calculus for AI governance. If the NPT-style architecture isn't buildable because the political conditions that enabled it no longer exist, then what's the alternative?

IX.What remains

Bilateral arrangements between major powers are possible even when multilateral institutions are weak. US-Soviet arms control continued through periods of intense hostility because both sides had incentives to manage the most dangerous risks. A US-China understanding on AI development — even one that is informal, limited, focused on specific catastrophic risks — might be achievable even without broader institutional backing. But bilateral deals are fragile. They depend on relationships between specific leaders and can collapse when administrations change.

Bloc-based governance is the Cold War model applied to AI: the US and allies coordinate policy within their sphere, China and partners do the same within theirs, and the blocs manage competition at the boundary. This is arguably what's already emerging. The export control regime is a form of bloc coordination. The EU AI Act applies to one bloc. China's governance framework applies to another. The question is whether bloc-based approaches can manage global risks or whether they simply create racing dynamics between blocs.

Industry self-governance fills gaps when public institutions can't or won't. The major AI labs have safety teams, publish research on risks, and coordinate through forums like the Frontier Model Forum. The limitations are obvious: self-governance lacks enforcement, creates conflicts of interest, and tends to address reputational risks more effectively than systemic ones. But in the absence of public architecture, industry governance is what exists.

Doing nothing is also an option. Racing without coordination, competition without guardrails, capability development at maximum speed with governance as an afterthought. This is arguably the default trajectory if deliberate architecture isn't built. The diffusion clock runs, the governance debt compounds, and the consequences arrive.

None of these alternatives are as robust as the multilateral architecture that managed nuclear risks. All of them are more fragile, more contingent, more dependent on circumstances that could shift. But the choice isn't between the NPT model and these alternatives. The choice is between these alternatives and pretending the NPT model is still available when the conditions that created it have eroded.

X.The clock is running

In 1953, President Eisenhower gave his “Atoms for Peace” speech to the United Nations General Assembly. The core argument was that nuclear technology would spread regardless of American preferences, and the choice was between managed diffusion with governance or chaotic diffusion without it.

Seventy years later, we face a version of the same question with a faster clock and less capacity to answer it. The multilateral architecture Eisenhower helped build is fraying. The institutions that would need to govern AI are weakened or absent. The political conditions that enabled postwar institution-building — American hegemony combined with American commitment to collective frameworks — have shifted in ways that may not shift back.

The denial strategy is buying time. The diffusion clock is running. The governance debt is compounding. And the architecture that would need to exist when the clock runs out isn't being built — in part because we may no longer have the institutional capacity to build it.

The Iran case is remembered as a failure of technology transfer. But the deeper lesson may be about the fragility of governance itself. The architecture that constrained Iranian nuclear development for decades required constant maintenance, required sustained commitment, required the belief that multilateral frameworks were worth the cost of American constraint. When that belief wavered, the architecture collapsed faster than it was built.

We are now attempting to govern the most consequential technology transition in human history during a period of institutional recession. The tools that worked for nuclear — treaties, international agencies, managed sharing with monitoring — may not be available. The alternatives are weaker, more fragile, more contingent.

The honest conclusion is uncertainty. We don't know whether bilateral arrangements or bloc-based coordination or industry self-governance can manage the risks that AI development creates. We don't know whether the institutional tide will come back in or whether we're seeing a permanent shift toward fragmentation. We don't know whether the time we're buying with denial strategies is enough to build whatever architecture is still possible.

What we know is that the clock is running, the precedents are less reassuring than they first appear, and the question of what comes next doesn't have an answer yet.