By Jinduan
Last Friday, a letter forced the world's two most powerful AIs offline simultaneously. A U.S. Department of Commerce export control order: it prohibited any foreign citizen from accessing Anthropic's Fable 5 and Mythos 5 models. Unable to verify user nationality in real-time, Anthropic did the only thing it could: shut down these two models, released just three days prior, to everyone, globally.
For the first time, humanity has placed an intelligent entity existing as bits under the same export control framework as enriched uranium.
The Enriched Uranium Playbook Is Repeating
Historically, export controls have been applied to only two categories: hardware and recipes. As we know, things like uranium enrichment centrifuges, high-end lithography machines, and military-grade encryption algorithms.
Their commonality is physical scarcity. Without the equipment, without the blueprints, you cannot build them. Controls work because the controlled items have physical boundaries; they can be intercepted at customs, traced through supply chains.
But Fable 5 is a set of weight parameters. It can be copied infinitely, bypasses customs, requires no shipping containers, and the physical act of "smuggling" doesn't exist. Theoretically, it could awaken on any server worldwide. The U.S. Commerce Department found all traditional tools ineffective; it couldn't stop Fable 5 at the border, so it had to cut it off at the source.
The true object of control is the "capability density" concentrated within this set of weight parameters: code generation capability, reasoning and planning capability, cross-domain knowledge invocation capability. When these capabilities were scattered across countless engineers' minds, they never triggered controls. But when compressed into a single model, when one person can invoke all these capabilities with a single prompt, the compression itself constitutes a threat.
This is the precise mapping of the enriched uranium logic into the digital world.
Uranium ore is widespread in the Earth's crust, never controlled; but when enriched to a certain concentration, it becomes one of the most closely monitored substances on the planet. Models are similar: when individual capabilities are scattered across open-source code repositories, technical Q&As, and academic papers, they are free. When all these capabilities are "enriched" into a single, callable interface, they cross a threshold—and the price for crossing that threshold is being shut down.
The history of enriched uranium provides a mirror for understanding this event.
In 1938, Hahn and Strassmann discovered nuclear fission in Berlin. In the following decade, uranium transformed from an obscure laboratory element into the most sensitive strategic material on Earth. In 1946, the U.S. passed the Atomic Energy Act, bringing all nuclear technology, materials, and knowledge under government control. Private capital was expelled from the nuclear field, international scientific exchange was severed, and even basic physical data were classified. A pure natural element was henceforth shackled with political chains.
The control eighty years ago was based on a grim logic: some forces are too powerful to be held by any entity not ultimately accountable to national interests. Eighty years later, the same logic may be reactivated, this time targeting not the fission of atomic nuclei, but the forward propagation of neural networks.
Three Things Will Happen in the Next Decade
The control of enriched uranium in the 1950s spawned an entirely new international governance structure: the International Atomic Energy Agency, the Nuclear Non-Proliferation Treaty, the Supplier Group. Once technological control is established, it irreversibly moves toward institutionalization, multilateralization, and permanence.
AI will be no exception. In the next decade, it is highly likely that three things will happen.
1. Capability review will become institutionalized.
Every new frontier model will not only undergo safety red-teaming before release but also government-authorized third-party compliance reviews. The review standards will not come from within the companies.
Model capability assessment will shift from "benchmarking" to "checklisting." Each potentially misusable capability on the list will trigger additional control requirements. A model's "enrichment"—parameter size, reasoning depth, cross-domain generalization capability—will be precisely measured like the concentration of Uranium-235, with thresholds set. Models exceeding a certain threshold will automatically trigger export control clauses.
2. Jurisdictional boundaries will become blurred.
The Fable 5 order targets "foreign citizens," regardless of their geographic location. For the first time, the U.S. government's control reach has extended to every user globally.
A developer in Singapore, using an American company's API, is subject to U.S. export control laws. A company in Berlin finds that its AI vendor's compliance obligations are determined not by German law, but by a single order from the U.S. Commerce Department.
This unilateral expansion of jurisdiction will force non-U.S. enterprises to rethink their AI supply chains: the U.S. vendor you rely on could be ordered by the U.S. government to cut off your access one day. The Fable 5 incident gave a very clear answer: it can.
3. Technological paths will diverge.
When closed-source frontier models face repeated shutdown risks, the global AI industry will be forced onto a dual-track system.
One track comprises U.S. closed-source frontier models, subject to export controls, each release accompanied by the risk of being taken down. The other track consists of open-source models, locally deployed models, models under non-U.S. jurisdictions—they may be less advanced, but they are not threatened by U.S. government power cuts.
The market share of open-source models will be driven not just by performance, but by the security attribute of "not being unplugged." For the past three years, open-source models have been catching up in capability; in the next decade, they may form a structural advantage over closed-source models in the dimension of "reliability."
The Deepest Fissure Lies in the Property Rights System
All the above speculations are built upon a fundamental question that remains unanswered. The deepest crisis exposed by the Fable 5 incident is that digital civilization has yet to establish a property rights system for "intelligence."
Legally, a model is sold as a service. You pay, and I use my assets to perform tasks for you. My assets remain mine; what you buy is merely their output. This logic has worked in traditional service industries for millennia without issue.
But AI is different. When your company spends three months fine-tuning all internal tools based on Fable 5's specific behavioral patterns, training employees, writing hundreds of automation scripts dependent on that model's specific output format, you have, in practice, turned Fable 5 into your means of production.
But legally, it is still Anthropic's service. It can be recalled on any given day, and the compensation you receive will not exceed the subscription fee you paid in the past month.
You invested real means of production, yet receive only service-level legal protection. The gap between the two is the loss borne by all corporate clients worldwide when Fable 5 was taken down. This loss does not appear on any balance sheet, triggers no insurance claims, and is not covered by any legal clause.
Humanity spent three hundred years building a legal system around "property." A piece of land, a factory, a patent—all have clear ownership, transaction rules, and dispute resolution mechanisms. But this system has a default premise: property is tangible, or at least has a traceable carrier.
When a model is unplugged, you can do nothing. It is neither stolen nor destroyed. It still exists, but you are not allowed to use it. This is a new form of deprivation: it deprives not the object, but the right to use. And the right to use, legally, never belonged to you.
Enriched uranium has been controlled for eighty years and remains humanity's most sensitive technological asset. AI control has just begun; its endpoint may be a permanently fractured digital world. In this world, the smartest model may not be the usable one. The usable model will inevitably be the one with the clearest property rights. Not being taken away, at certain historical junctures, is far more important than temporary superiority.
