The real AI policy is written in export controls

The most important AI regulation of the last two years did not come from a legislature, a model evaluation, or anyone who has ever used the phrase "frontier safety." It came from the Bureau of Industry and Security, an obscure agency inside the Commerce Department, in the form of a rule about how many transistors you are allowed to bond together before the federal government wants a word. I say this as someone who spent years inside the regulatory machine and watched the AI debate float entirely free of it.

Here is the popular framing, stated fairly, because it deserves to be. The way states will govern artificial intelligence, the thinking goes, is by regulating the software: licensing powerful models, mandating evaluations, drawing red lines around capabilities, perhaps standing up a dedicated agency to gatekeep releases. It is a serious view, held by serious people, and the EU AI Act and a year of executive orders have given it institutional weight. If you want to know how AI will be controlled, this framing says, watch the model layer.

It is also, I want to argue, looking at the wrong layer. The instrument that is actually shaping who can build advanced AI, and how fast, is not model regulation at all. It is export control on the supply chain underneath the models, and specifically on the most boring-sounding part of it: advanced packaging and the memory that gets packaged. That is where the state is genuinely setting terms right now. The model-regulation debate is loud. The packaging rules are quiet, technical, and decisive.

The category error in the room

Start with the category error, because naming it is the whole game. We keep treating "AI policy" as a question about software, when the binding constraint is physical. A frontier model is not conjured from regulatory ether; it is trained on tens of thousands of accelerators, each of which is a logic die lashed to stacks of high-bandwidth memory by an advanced-packaging process. Roughly half the manufacturing cost of a leading AI chip is its HBM. There is no advanced AI without that assembly, and the assembly has a chokepoint with a name: CoWoS.

CoWoS — chip-on-wafer-on-substrate — is TSMC's technique for stitching a processor to its memory on a silicon interposer. It is the narrowest pipe in the entire AI build, and in 2026 it is sold out. TSMC has been scaling the line from roughly 35,000 wafers a month in late 2024 toward a projected 130,000 by year's end, an 80 percent compound growth rate, and demand still runs ahead of it. Nvidia alone has reportedly booked well over half of that capacity through 2027, with backend fabs quoting lead times north of a year. When one company can absorb most of the world's supply of a single packaging process, that process is not a manufacturing detail. It is a policy lever, and governments know it.

You cannot license a model into existence, but you can decide whether the silicon it runs on ever gets assembled. One of those is a press release. The other is the actual rule.

What the rules already do

Look at what BIS has actually written, not what the op-eds say it should. The December 2024 package and its successors did three precise things, and none of them touched a model.

1. Put high-bandwidth memory itself under control — U.S.-origin HBM and, via the Foreign Direct Product rule, foreign-made HBM with U.S. tooling in its lineage — gated by a memory-bandwidth-density threshold rather than by what the chip is "for."
2. Reached into advanced packaging directly, with attestation regimes keyed to whether a finished package contains HBM and to an aggregated transistor count — roughly 30 billion transistors for an HBM-bearing package — so the rule binds at assembly, the exact chokepoint.
3. Tightened controls on the semiconductor manufacturing equipment and named the foundries and "authorized IC designer" pathways, so that even a clean-room workaround inherits a license requirement.

Notice the design. The threshold is not "is this an AI chip," a question no rule can answer cleanly, because every definition either sweeps in your laptop or lets a determined actor reclassify around it. The threshold is transistor count and memory bandwidth density at the moment of packaging — physical, measurable, and tied to the one step that cannot be done at scale outside a handful of buildings. This is what good export control looks like when an agency actually understands the technology: it regulates the chokepoint, not the abstraction.

Steelman the free-trade objection

Now the strongest case against all of this, because if you cannot state it, you have not earned your conclusion. The free-trade objection is not naive, and it is not the libertarian reflex that government should simply get out of the way. It is the harder version: that unilateral controls on a globally integrated supply chain mostly relocate the activity they were meant to stop, while taxing the people who imposed them.

The evidence is real. CSIS analysts have documented that controls handed China a national mission rather than a wall. Beijing has poured something on the order of $150 billion into domestic chips and set a 70 percent localization target; the Chinese-made share of semiconductor equipment in its own market climbed from roughly a quarter to a third between 2024 and 2025. China now publishes about twice as many chip-design and fabrication research papers as the United States. Smuggling is not hypothetical — a 2024 ring moved an estimated $390 million in restricted GPUs, and Huawei reportedly obtained two million chiplets through shell buyers. Controls relevant only while you hold something the other side needs are a depreciating asset, and you are spending down the asset every time you tighten.

There is a domestic cost too, and it is the part the hawks skip. By fencing American toolmakers out of the largest equipment market on earth, the United States gave Chinese buyers a standing reason to design U.S. suppliers out of their supply chains permanently. That is not a one-year revenue dent; it is the deliberate construction of a parallel ecosystem that no future administration can un-build. The objection, fully steelmanned, is that export controls buy a few years of friction at the price of a permanent rival and a poorer domestic industry.

Why it is still the real instrument

And yet. Concede every point above and the conclusion does not flip, because the question was never "do export controls make China go away." They do not, and any official who promised that oversold the tool. The question is which instrument is actually allocating frontier AI capability today, and the honest answer is the one written in EAR citations, not model licenses. A few years of lead time at the frontier, in a field where capability compounds, is not nothing — it is most of what "governing AI" has concretely meant so far. The model-regulation debate has produced frameworks; the packaging controls have produced facts on the ground.

Watch where the political energy goes when something real is at stake, and you see the same thing. When the administration moved on Nvidia's H20 in 2025, it did not convene a model-evaluation panel. It pulled an export license, booked the company a $4.5 billion inventory charge, and three months later cut a deal to let the chips flow again in exchange for 15 percent of the China revenue — an arrangement so plainly a tax on exports that Nvidia flagged the constitutional risk in its own SEC filing. That is the state shaping AI in real time, with a license as the throttle and a chip as the unit of account. No one in that negotiation was arguing about system cards.

What follows, with the costs admitted

If the supply chain is the real governing surface, then the policy questions worth arguing about are not the ones dominating the hearings. They are these. Who actually enforces a packaging attestation, and how, when the finished module can be re-marked and re-routed through a third country? What is the realistic version of multilateral control when the Netherlands, Japan, South Korea and Taiwan each hold a different chokepoint and none of them will move at America's tempo for free? And critically: what is the off-ramp, the answer to a control regime whose own logic guarantees it expires the moment the other side reaches parity?

I will admit the cost of my own position, because the column is worthless if I don't. Treating export control as the real AI policy means accepting that we are governing the most consequential technology of the decade with a tool that is leaky, unilateral by default, and self-depleting — and that the rule writers at BIS, not any AI agency, are the de facto regulators. That should make the model-safety crowd uncomfortable, and it should make the free-traders uncomfortable too, because pretending the controls don't bind is as much a fantasy as pretending they're permanent.

But clarity about the instrument is the precondition for using it well. The debate we are having is about whether to put a fence around the software. The debate that matters is whether we understand the one we already built around the silicon — who it stops, who it merely annoys, and how many years of lead we are buying before the bill comes due. I would rather argue about the rule that exists than the one that photographs better. The frontier is not being governed in the conference room. It is being governed at the bonding station, one wafer at a time.