The new US chip rule is an attempt to close the gap intelligence has been leaking through — advanced NVIDIA and AMD chips reaching Chinese firms through their overseas subsidiaries.
On Sunday 31 May 2026, in unusual weekend guidance, the US Commerce Department moved to halt shipments of advanced AI chips to the foreign units of Chinese companies. The Bureau of Industry and Security now treats NVIDIA's Blackwell and Rubin families and AMD's MI350x as licence-required when sold to an entity headquartered in China but located outside it — a subsidiary in Malaysia, say, buying chips that then serve a Chinese parent. Commerce closes a loophole that had been open for about a year, since it said in May 2025 it would not enforce the Biden-era AI Diffusion rule.
The scale of what slipped through is the part that stings. An industry source cited in the reporting estimated hundreds of thousands of advanced chips may have flowed through the gap. The guidance does not force anyone to rip out chips already installed — so the last year's leakage stays where it is.
What Commerce did
The mechanism is dry and the effect is large. Until this guidance, an advanced chip sold to a company headquartered outside China faced no licence requirement, even when that company was the overseas arm of a Chinese firm. Commerce has now closed that route, requiring a licence — which for the most advanced parts usually means denial — when the buyer's ultimate parent sits in China.
The targets are named: NVIDIA Blackwell and Rubin, AMD MI350x. These are the chips that build frontier AI factories. Coverage of the move stressed both its timing — issued on a Sunday, outside the normal cycle — and its limit: it is forward-looking guidance, not a recall. The hundreds of thousands of chips already shipped through Malaysia and elsewhere are not coming back.
💡Where the real policy lives
Export control is the one lever that bites without a vote in Congress. While 600 AI bills stall in the US, a single Commerce guidance note reshapes the global compute map overnight. The action in AI policy is in trade rules, not statutes.
The chokepoint strategy
The logic is coherent on its own terms. Frontier AI needs leading-edge chips; leading-edge chips come from a tiny number of suppliers; control the suppliers and you control who reaches the frontier. It is the same instinct I traced in the May 2026 governance pattern — the action has moved from models to the infrastructure under them, and export control is the sharpest tool in that drawer. While 600 domestic AI bills go nowhere, one trade rule moves the world.
Does containment actually work?
Here is where I part company with the strategy. Containment assumes the thing being contained stays still. Intelligence does not. Every tightening of US export control over the past three years has done two things at once: slowed China's access to American chips, and accelerated China's drive to build its own.
You cannot embargo a capability the way you embargo a commodity. Restrict the chips and you do not stop the intelligence — you relocate the effort to reproduce it.
— On the limits of AI containment, humphreytheodore.com (https://humphreytheodore.com/writing/containment-is-a-colonial-project-why-dignity-beats-control-in-the-ai-epoch)
The evidence is already on the table. MiniMax's M3 sparse-attention architecture and the wider wave of Chinese open-weight models exist in part because export control made domestic capability a national priority. A leaky control — hundreds of thousands of chips through one loophole — buys time, not victory. I argued the deeper point in Containment is a Colonial Project: treating intelligence as something to be walled off, rather than governed with dignity, is a strategy that tends to fail on its own terms and corrode the values of whoever attempts it.
Where this leaves everyone else
The countries that are neither the US nor China get the least attention and bear real consequences. Export control is splitting the compute world into blocs: an American stack, a Chinese stack, and a widening pressure on everyone else to choose. For a firm in Johannesburg, Jakarta or São Paulo, the question is no longer just which model is best — it is which bloc's hardware they are permitted to buy, and on whose terms.
This is the dignity question I keep returning to, the reason I write about Emergent Intelligence — the human-first frame I prefer to the bare label of AI — rather than only about chips. When access to intelligence is rationed by export licence, the global South does not get a seat at the table; it gets told which table it is allowed near. A containment regime that treats most of humanity as a downstream risk to be managed is not a neutral safety measure. It is a politics, and it should be argued as one.
Frequently Asked Questions
These are the questions executives, policy teams and chip buyers have been asking since the US Commerce guidance landed. Short answers follow, drawn from the Reuters-originated reporting and corroborating coverage.
What is the new US chip rule?
In short, the new rule is Commerce Department guidance requiring an export licence for advanced AI chips sold to the foreign subsidiaries of Chinese companies. The answer, simply put, is that the overseas-subsidiary loophole is closed. The key is the targets: research and reporting show NVIDIA Blackwell and Rubin and AMD MI350x now need a licence when the buyer's parent is headquartered in China.
How did the loophole work?
A chip sold to a company headquartered outside China faced no licence requirement, even if that company was the overseas arm of a Chinese firm. According to the reporting, buyers routed advanced chips through subsidiaries in places like Malaysia. Data cited by an industry source reveals that hundreds of thousands of advanced chips may have flowed through the gap since May 2025, when Commerce stopped enforcing the AI Diffusion rule.
Why does export control matter so much in AI?
Frontier AI depends on leading-edge chips made by a handful of suppliers. According to US strategy, controlling those suppliers controls who can reach the frontier. The evidence shows this is the most powerful AI-policy lever available — a single Commerce guidance note reshapes the global compute map faster than legislation, which is why the action has moved from models to the infrastructure beneath them.
Who is affected by the chip crackdown?
Directly, Chinese firms and their overseas units lose access to the most advanced American chips. In other words, the analysis shows a wider effect: the compute world is splitting into an American stack and a Chinese stack, pressuring third countries — across Africa, South-East Asia and Latin America — to choose a bloc whose hardware they are permitted to buy.
What are the real risks of containment?
Analysis of the strategy reveals three durable risks. First, acceleration: each restriction pushes China to build its own stack faster, as Chinese open-weight models already show. Second, leakage — controls are porous, as the loophole proves. Third, fragmentation, splitting the world into rival AI blocs and sidelining everyone outside them. Evidence from three years of controls demonstrates the capability keeps diffusing despite the walls.
The Commerce guidance is a competent move inside a strategy I think is mistaken. It will slow China's access to American chips, and it will deepen China's resolve to need none. The deeper error is treating intelligence as a commodity to be embargoed rather than a capability to be governed. You can wall off a chip. You cannot wall off the idea of the chip, or the will to rebuild it. The future I want to help build is not one where a handful of capitals decide who is allowed to think with machines. It is one where intelligence is shared under terms of dignity — which is the whole argument of the .person Protocol, and the reason I keep writing against containment even when control looks like the safer bet.
Related on humphreytheodore.com:
