What it is
AI export controls have traditionally focused on hardware: the advanced chips, like Nvidia GPUs, that are needed to train large models. A newer and more contested idea is to control the weights themselves — the billions of numerical parameters a model learns during training, which together constitute the trained model’s actual intelligence. Weights are typically stored as a set of files; copy them, and you have copied the model’s capabilities, without needing the computing power, data, or engineering time that produced them.
Controlling weights as an export category means requiring a government license before they can be transferred across borders — whether sold, licensed, or simply moved to a server in another country. This is a different target than controlling chip sales or cloud access: a chip export control limits who can build a frontier model in the first place, while a weights export control limits who can acquire one that has already been built.
The United States moved first. In January 2025, the outgoing Biden administration’s Bureau of Industry and Security (BIS) issued the “Framework for Artificial Intelligence Diffusion,” which created a new export control classification (ECCN 4E091) for the weights of closed-source models trained on more than 10^26 computing operations — a threshold BIS estimated fewer than five models worldwide then exceeded. Licenses for such transfers were to be reviewed under a presumption of denial. The rule exempted “published” open-weight models. The incoming Trump administration rescinded the rule in May 2025 before it took full effect, and as of late 2026 no formal replacement had been issued — the underlying authority remains in regulatory limbo, announced as unenforced but not struck from the books.
China has now proposed the mirror image. On October 16, 2026, China’s Ministry of Commerce and Ministry of Science and Technology published a draft adding 23 AI-related items — alongside photovoltaic and biomedicine technologies — to the catalogue of exports requiring a license. The AI provisions target the weights, checkpoints, source code, and training methods of large models exceeding 2 trillion parameters, 10^26 training FLOPs, or a global top-10 ranking, plus autonomous agents capable of running other agents or orchestrating large numbers of tools. Like the US rule, it exempts models already made public through open-source release or academic publication. An interim registration requirement took effect immediately; public comment runs to November 15, 2026.
Why it matters for AI governance and narratives
Both regimes share a conspicuous gap: neither restricts access to a model delivered as a hosted service — an API or consumer app that any developer, anywhere, can call. A government can require a license to hand over the weights file, but it has no equivalent mechanism for the case where the model never leaves the host country’s servers at all. This is the asymmetry the observatory’s editorial flagged: one government (the US) has built its control regime around who can purchase the chips needed to train a frontier model; the other (China) is now proposing to control who can receive the finished weights. The hosted API — the channel through which most people outside either country actually touch frontier AI — sits outside both frameworks.
That gap matters for how AI capability diffuses globally, and for the narrative contest over who is “restricting” AI versus who is “open.” A regime that controls weights but not hosted access can still see its most capable models used worldwide via API, while claiming to have restricted their proliferation — a claim worth scrutinizing rather than accepting at face value.
Key facts and dates
- January 13, 2025: US BIS “AI Diffusion Rule” takes effect, creating ECCN 4E091 for closed-weight models trained above 10^26 operations.
- May 13, 2025: BIS announces rescission; enforcement ends May 15, 2025. No replacement rule issued as of October 2026.
- October 16, 2026: China’s MOFCOM and Ministry of Science and Technology publish a draft catalogue update covering weights, checkpoints, and training methods of models above 2 trillion parameters or 10^26 FLOPs, plus agent-orchestration technologies. Public comment open until November 15, 2026; interim registration requirement already in effect.
- Both the US rule (as issued) and the Chinese draft exempt models whose weights have already been published openly.
Where to learn more
- Federal Register: Framework for Artificial Intelligence Diffusion — the official US rule text.
- Sidley Austin: New U.S. Export Controls on AI Model Weights — Seven Key Takeaways — detailed legal breakdown of ECCN 4E091’s scope and exemptions.
- Akin Gump: BIS Rescinds Its AI Diffusion Rule and Issues Compliance Guidance — on the rule’s current unenforced status.
- Global Times: China adds 40 technologies in AI, biomedicine and PV for export controls in draft revised catalog — Chinese state-media reporting of MOFCOM’s October 2026 draft.