Does It Work? The Strategic Balance Sheet

The hardest question in this field is also the most important, and it does not have a settled answer. Do compute export controls actually achieve their goals? A practitioner who claims certainty here is selling something. What a serious analyst can do is lay out the balance sheet honestly, weighing what the controls plausibly accomplish against what they plausibly cost, and then reason about the conditions under which the ledger tips one way or the other.

Start with the case for. The core claim is that controls buy time. If strategically significant AI capability scales with compute, and if the most advanced compute runs through supply chains the United States can influence, then denying or slowing an adversary's access to frontier accelerators delays that adversary's most capable systems. Even a delay of a few years, the argument goes, is strategically valuable, because it preserves an advantage during a critical period and forces the adversary to divert enormous resources into catching up on hardware rather than pushing ahead on applications. On this view, controls are not expected to stop anyone permanently. They are expected to impose cost and buy time, and by that standard they can be judged effective even if the restricted actors eventually develop alternatives.

Now the case against. The first objection is that controls provoke exactly the independent capacity they aim to prevent, as the previous lesson described, potentially trading a temporary lead for a permanent rival. The second is that they impose real costs on American and allied companies, who lose sales, cede market share in contested regions, and fund their eventual competitors' rise by first teaching them the market exists. The third is that controls strain relationships with allies and partners who resent the extraterritorial reach and the imposition of quotas, weakening the very coalition the strategy depends on. The fourth is that leakage through gray markets and the cloud may blunt the controls enough that they impose cost and provoke retaliation without actually denying much capability, the worst of both worlds. Each objection is serious, and none can be dismissed.

The honest synthesis is that the answer depends on variables that are still resolving. It depends on how wide the gap really is between frontier and constrained hardware, and whether that gap matters as much as the compute-scaling premise assumes, because if capability turns out to depend less on raw scale and more on efficiency and algorithms, the entire chokepoint theory weakens. It depends on how fast restricted regions can build credible domestic alternatives, which determines whether the controls buy years or merely months. It depends on how well the leaks can be managed, which determines whether the controls have real teeth or merely symbolic ones. And it depends on whether the coalition holds, because controls the United States imposes alone are far weaker than controls its allies enforce alongside it. These variables are not yet settled, which is precisely why the strategic verdict remains open.

There is also a deeper uncertainty about the premise itself. The entire edifice rests on the belief that compute is the decisive chokepoint for strategically significant AI. If that belief is correct, controlling compute is controlling the future, and the costs are worth bearing. If it is wrong, if capability comes to depend more on data, on algorithmic efficiency, on talent, or on approaches that need less frontier hardware than assumed, then the controls may be fighting the last war, imposing large costs to guard a chokepoint that turns out not to be decisive. Betting the strategy on a single premise about how AI capability scales is the boldest and most questionable aspect of the whole enterprise, and no one yet knows whether the bet is sound.

Faraday's read: hold this question genuinely open, because the evidence is not yet in and confident answers in either direction are premature. The most defensible position is conditional. Compute controls are a bet that frontier hardware is the decisive lever and that a lead of a few years is worth the cost of accelerating competitors, straining allies, and sacrificing sales. If the compute-scaling premise holds and the coalition stays intact, the bet is reasonable and may look prescient. If capability decouples from raw compute, or if the controls mainly succeed in birthing independent supply chains and alienating partners, the bet will look expensive and shortsighted. A practitioner's job is not to pick a side in advance but to watch the resolving variables, the hardware gap, the speed of domestic alternatives, the durability of the chokepoints, the cohesion of the coalition, and update as the evidence arrives. The one certain thing is the thesis this course began with: whatever the ultimate verdict, export controls have already reshaped the global compute map, and that map is now a permanent feature of the terrain anyone in this field must navigate.