Too Cheap to Govern

Then Phil Zimmermann's source code was published as a hardback book, because books have the First Amendment, and the Ninth Circuit found in Bernstein that code was speech. The regime did not lose the argument. It lost the chokepoint. There was nothing left at the border to inspect.

Eric Levitz, writing this week in Vox and syndicated to RealClearPolitics, notes that Chinese labs are shipping models — Kimi K3, Qwen3.8 Max — that land near the frontier and publish their weights as free downloads. He frames it as a threat to Anthropic and OpenAI, each valued near a trillion dollars. The financial story is the smaller one. The entire regulatory architecture assembled since 2023 rests on the same premise and inherits the same exposure.

Look at what that architecture actually grips. Compute thresholds — the 10^25 and 10^26 FLOP lines in the EU AI Act and the American executive orders — regulate training runs, which are expensive and traceable through the electricity bill. Export controls regulate advanced chips, which are physical, fabbed in about three places, and countable at customs. Pre-deployment evaluations regulate frontier labs, which are a handful of companies with lawyers and something to lose.

Every one of those is a clamp on a chokepoint. That is not a design flaw; it is the only thing law has ever been able to do. You cannot regulate a diffuse practice. You regulate the narrow place it has to pass through. The scarcity that created those narrow places was a market condition, not a law of nature, and nobody wrote that caveat into the statutes.

The pattern is well documented, and what the cases share is not their shape but their economics. Ecclesiastical licensing of printers held while presses were capital equipment owned by identifiable guildsmen, and collapsed when presses got cheap. Music-industry enforcement scaled beautifully against Napster, which had servers, and not at all against BitTorrent, which had a protocol. In each case the governed thing was expensive to acquire and then approximately free to use — a text once set, a file once copied, a cipher once compiled runs anywhere at no marginal cost. Law clamps acquisition. When acquisition gets cheap there is nothing downstream left to clamp, because use leaves nothing to inspect.

Nonproliferation is the standing counterexample and it holds for that reason inverted: enrichment is expensive to acquire and expensive to sustain, slow, and visible from orbit. The treaty is not what stops proliferation. The centrifuge cascade's price tag is. Law does not create governability; it exploits governability that physics and economics already supplied, and expires quietly when the supply runs out.

Which is the test this argument owes itself, and it does not come back clean. Free weights are not free capability. A downloaded model still has to run, and running one at frontier scale means chips — physical, fabbed in about three places, countable at customs, which is to say the chokepoint listed four paragraphs up. Export controls do not care whether the FLOPs are spent training or serving. The honest boundary: the chokepoint did not dissolve, it moved down the stack from the training run to the inference bill, and anything needing a datacentre stays as governable as it ever was. What escapes is everything below that line — a distilled thirty-billion-parameter model, quantised, running on a consumer GPU, with no training run to threshold, no customs entry, no vendor to serve papers on. The question is not whether capability goes free. It is how much of what people actually want from a model ends up under the line. On present trend, most of it.

The joke is that the same fact reads as triumph or catastrophe depending which question you brought. Cheap, open, widely available: that is the answer to enclosure, the thing everyone said they wanted when they warned that five firms would own the substrate of collective cognition. It is also, in the identical sentence, the end of anyone's ability to recall a model or require an evaluation before release.

But it answers enclosure of possession, which is the cheaper half. What most people's thinking touches is not determined by who may download weights. It is determined by defaults — which assistant the phone ships with, which model the operating system surfaces, which provider the app store approves and the hyperscalers route through. Commoditising the model layer does not distribute that. It pushes the margin up the stack to whoever owns the surface and concentrates it there. Enclosure did not end. It changed floors, from the thing to the doorway, and nobody has written a statute about doorways. The people whose cognition gets arranged by which button is nearest appear nowhere in this argument, which is a fairly precise description of where the control went.

So the migration is already scheduled: from the model to the deployment, from licensing the builder to policing the use, from pre-release evaluation to post-hoc liability. Slower, messier, and the only remaining surface with a chokepoint attached — because deployments have customers, and customers have addresses.

Watch for the second-order comedy while you wait. The labs that lobbied hardest for compute-threshold regulation were, whatever else they believed, building a moat. A rule that binds only organisations capable of a 10^26 FLOP training run fences out everyone who cannot afford one. It is a strange kind of loss to have the regulation you helped design stop mattering because the thing it regulated became free.

Regulatory capture requires something to capture. The interesting question this year is not whether the labs captured the regulators. It is whether either party has noticed that the object of regulation did not walk out the door as a torrent file. It went upstairs, to the floor where the defaults live, and left the statute downstairs guarding an empty room.