The US Commerce Department has drafted a rule that would require export licenses for virtually every AI chip shipment Nvidia and AMD make — to any country on Earth. Bloomberg first reported the draft on March 5, 2026, and the scope of it is almost absurd: not just China, not just adversaries — everyone. Allies included.
The stated goal is to prevent advanced AI accelerators from reaching China through third countries. The practical effect would be to hand Washington a veto over Nvidia’s entire global business. That’s a significant escalation from where this started.
Four Years of an Increasingly Messy Trade War
The US first banned exports of high-performance AI chips to China in October 2022, targeting anything as capable as the Nvidia A100. China responded by accelerating domestic alternatives. Then the US tightened again in October 2023, closing workarounds. Then Nvidia designed the H20 specifically for China — lower specs, legally exportable — and Chinese companies ordered millions of them.
In April 2025, the Trump administration clamped down on H20 exports too, requiring special licenses. By August 2025, China hit back differently: Chinese authorities told domestic tech firms to stop buying Nvidia chips entirely, and state media declared them a security risk. Nvidia halted H20 production. The chip it built specifically for China became a chip China officially refused to use.
Then Trump reversed course. He met with Nvidia CEO Jensen Huang in December 2025 — Huang had been lobbying against export controls for months — and green-lit H20 sales again, this time with a 25% fee flowing to the US government. Congress wasn’t happy and pushed back hard in February 2026, advancing a bipartisan bill to limit the White House’s ability to make those calls unilaterally. Now the Commerce Department’s new global licensing draft lands on top of all that — and nobody seems sure what the actual policy is.
China Didn’t Wait Around
While Washington debated itself, Chinese AI labs started shipping frontier models trained entirely on domestic hardware. Zhipu AI released GLM-5, a 745-billion-parameter open-source model approaching Claude and GPT-level performance — trained on Huawei’s Ascend chips using Huawei’s MindSpore framework, zero Nvidia dependency. Earlier, Zhipu’s GLM-Image became the first major Chinese model trained end-to-end on domestic silicon. We also covered China’s Moonshot AI releasing a trillion-parameter model — the pace of Chinese AI development is not slowing down.
Then at MWC Barcelona 2026, Huawei unveiled the Atlas 950 SuperPoD — 8,192 Ascend NPUs linked into a single system, its first overseas showcase of the hardware. That’s not a domestic product anymore. That’s a global pitch to any country that wants AI infrastructure without US strings attached.
The Unintended Consequence Nobody Wants to Say Out Loud
The original argument for export controls was that restricting chip access would slow China’s AI development. The evidence from 2025 suggests the opposite happened: the restrictions accelerated China’s investment in domestic alternatives, and those alternatives are now competitive enough to train frontier models and be marketed internationally.
The new global licensing rule, if it passes, would extend that same logic worldwide — giving every country that buys AI infrastructure a reason to ask whether there’s a more reliable supplier. Jensen Huang has called export controls “a failure.” Congress thinks he’s wrong. The Commerce Department just proposed something affecting the entire planet, and the only thing clear right now is that nobody is steering this ship.