The Three Axes: Item, End User, Destination

If the previous lesson gave you the philosophy, this one gives you the map. Every export control decision, no matter how baroque, resolves into three questions. What is the item? Who is the end user? Where is the destination? Get these three coordinates right and you can locate almost any transaction inside the system. Get one wrong and you can turn a routine sale into a federal violation.

Start with the item. Controls do not apply to all chips, only to those that cross defined technical thresholds. Historically the key measures have been processing performance and the density of interconnection between chips, because both matter enormously for training large models. A single fast accelerator is useful; thousands of them wired together with high-bandwidth links are transformative, because the bottleneck in large training runs is often not raw computation but how quickly chips can share results. Regulators learned this and wrote it into the rules, controlling not just chip speed but the interconnect bandwidth that lets many chips behave as one giant machine. This is why a control threshold can catch a data-center accelerator while leaving a laptop graphics chip untouched, even when both come from the same design lineage.

The item axis also reaches upstream and downstream. Upstream, it covers the tools and know-how needed to make advanced chips: the lithography equipment, the design software, the specialized inputs. Controlling the means of production is often more powerful than controlling the product, because a country that cannot manufacture at the frontier stays dependent no matter how many finished chips it smuggles. Downstream, the rules increasingly reach the memory that advanced accelerators depend on. High-bandwidth memory, the stacked memory that feeds data to a processor fast enough to keep it busy, has itself become a control target, because without it even a powerful chip starves. Controlling a complement can be as effective as controlling the core product.

The second axis is the end user. The same chip can be perfectly legal to sell to a university and illegal to sell to a specific company down the road, because that company sits on a list of restricted parties. The best-known of these is the Entity List, a roster of named organizations that require licenses for transactions that would otherwise be routine, with those licenses often reviewed under a presumption of denial. The end-user axis is where intent re-enters a system built to avoid judging intent. Regulators cannot read minds, but they can name names, and once an organization is named, everyone in the supply chain must screen against it. This creates a compliance burden that ripples outward: distributors, cloud providers, and financiers all must know not just who they are selling to but who their customer is ultimately serving.

The end-user axis also carries the system's sharpest anti-evasion tools. Rules exist to stop a controlled item from reaching a restricted party through an intermediary, a front company, or a diversion through a permissive country. Knowing that a shipment will ultimately end up somewhere prohibited can make the sale illegal even if the immediate buyer is entirely legitimate. This is why sophisticated firms now investigate their customers' customers, a practice that would have seemed paranoid a decade ago and is now simply prudent.

The third axis is destination. This is where the country chart lives. For each reason an item is controlled, the chart specifies which destinations trigger a license requirement. Some countries are close allies and face almost no friction. Others face targeted requirements. A small set faces comprehensive restriction. The crucial insight is that destination is not just about the buyer's nationality but about where the compute will physically operate. A company headquartered in a permissive country can still trigger controls if it plans to install the hardware in a restricted one. Physical location of the silicon, not the flag on the letterhead, is what the rules increasingly chase.

These three axes interact, and the interactions are where the difficulty lives. A powerful item sold to a clean end user in a friendly destination is easy. A powerful item sold to a clean end user who will operate it in a sensitive destination is hard. A moderate item sold to a listed end user is hard regardless of destination. The system is essentially a three-dimensional grid, and a transaction is permitted only if it clears on all three axes at once. Failing any single axis can stop the deal.

Faraday's read: the strategic direction of travel is toward the end-user and destination axes and away from pure item thresholds. Item-based controls have a built-in weakness. Hardware improves, so any fixed performance line is eventually crossed by chips that are trivial to make, forcing endless revision. Controls keyed to who and where are more durable, because they track the thing regulators actually care about, which is not the chip but the capability it confers on a specific actor in a specific place. Expect the future of this system to lean harder on knowing your customer and knowing your geography, and less on chasing a performance number that hardware keeps outrunning. For the practitioner, that means compliance is becoming less an engineering question about specifications and more an intelligence question about people and places.