Tiers, Caps, and the Middle World
The single most ambitious move in the history of compute controls was the attempt to stop regulating chips one deal at a time and instead draw a map of the whole world. The AI diffusion framework, introduced in the final stretch of one administration and then reshaped by the next, tried to answer a question the piecemeal system had dodged: if compute is strategic, how much of it should each country be allowed to accumulate? The answer it gave was a tiered geography.
The structure was straightforward in concept. At the top sat a small group of close allies and partners, countries deemed trustworthy enough to receive advanced accelerators with minimal friction. These were the destinations where American compute could flow freely, on the theory that a chip in an allied data center strengthens the bloc rather than the adversary. At the bottom sat a set of countries of concern, destinations already under heavy restriction, for whom the presumption was denial. The novelty was not the top or the bottom. It was the middle.
Between the trusted few and the excluded few lay most of the planet: dozens of countries that were neither allies nor adversaries, ranging from wealthy technology hubs to developing economies with ambitions to build their own AI capacity. The old system had no coherent answer for these places. The diffusion framework tried to give one by assigning numerical caps, ceilings on how much advanced compute could be exported to a given country without special approval. A nation in the middle tier could import accelerators up to a threshold; beyond it, buyers needed licenses and, in some designs, needed to meet security and trust standards to qualify for larger allocations.
This was a genuinely new idea in trade policy. Ordinarily, controls say yes or no. The diffusion approach said how much, and in doing so it treated compute like a rationed commodity. The mental model was less a border checkpoint and more a global allocation table, with each country assigned a budget of frontier computing power. The logic was that raw prohibition was too blunt for a world where American firms wanted to sell and most countries posed no threat, while unrestricted sales risked letting adversaries accumulate compute through intermediaries. Caps were the compromise: a dial rather than a switch.
The framework also introduced the notion of trusted operators. A cloud provider or data center operator that met stringent security requirements could earn a status allowing it to deploy larger volumes of compute, even in middle-tier countries, on the theory that a well-secured facility run by a vetted company is safer than the same hardware in unaccountable hands. This shifted some of the burden from geography to institutions. It mattered not only where the compute sat but who was trusted to run it. In effect, the rule tried to build a class of pre-approved custodians who could carry compute across the tier lines that would otherwise constrain it.
The framework proved intensely controversial. Allies and partners in the middle tier objected to being assigned quotas at all, seeing it as a demotion and a constraint on their own development. Industry warned that complexity and caps would push customers toward non-American suppliers, ceding the market in the name of controlling it. Within the United States, critics argued the design was simultaneously too permissive, because it still allowed large flows, and too restrictive, because it alienated the very partners whose cooperation the strategy required. The result was that the incoming administration reworked significant portions of the approach, dropping some of the rigid tiering in favor of other mechanisms. The specific rule, in its original form, did not fully survive.
But the idea survived, and that is the point. The diffusion framework revealed how Washington now thinks about compute: as a resource to be allocated across a map, with abundance for the trusted, denial for the dangerous, and managed scarcity for everyone in between. Even as the mechanics were revised, that conceptual architecture kept reasserting itself, because it flows directly from the premise established earlier in this course. If compute is strategic, then its global distribution is a strategic variable, and strategic variables get managed.
Faraday's read: watch the middle tier, because that is where the real contest lives. The top and bottom are settled; nobody expects free compute flows to adversaries or heavy friction with the closest allies. The unresolved question is how the United States treats the large group of ambitious, unaligned, or swing states that want to build domestic AI capacity and are being courted by every supplier. Whoever wins the allegiance of the middle world shapes the compute map for a generation. The specific caps will be argued over endlessly. The strategic prize, the middle, will not change. For a practitioner, the lesson is that the most commercially important and politically volatile deals are increasingly in these swing markets, where the rules are least settled and change most often.