Washington finally let Nvidia sell its best chips to China. China decided it doesn't want them.

The story everyone expected from US chip controls was simple: America restricts, China falls behind. What is happening instead is stranger, and more consequential for anyone — Britain included — who has staked their AI plans on the same handful of suppliers.

Late last year, after heavy lobbying, the Trump administration reversed course and allowed Nvidia to sell its advanced H200 chips to China, with an unusual condition attached: the US Treasury would take a cut of each sale (reported by the Financial Times and New York Times, early 2026). It looked like a win for Nvidia. Then Beijing intervened — not to seize the chips, but to wave them away. Chinese regulators reportedly told domestic firms not to buy, citing security concerns and a preference for home-grown silicon, and by mid-May not a single H200 had been sold (Brookings, June 2026). Nvidia's own chief executive has spent months trying to unstick the sales, so far with little to show for it.

Why this matters reaches the UK directly, so let me land it here. Britain's whole AI ambition — the Growth Zones, the datacentre pledges, the "AI superpower" framing of the January 2026 action plan — rests on access to exactly these chips, most of them from one American company, along a supply chain that runs through Taiwan and a few Dutch and Korean firms. A world where that chain fractures along political lines is a world where UK compute gets scarcer, dearer, or both. When the two largest players decouple, the medium-sized customer — which is what Britain is, in compute terms — feels the squeeze with no say in it.

What China is doing with the gap is the part Westminster should study. At its trade and technology showcases in Beijing and Shanghai in June, Huawei was the story, not Nvidia (Associated Press, late June 2026): domestic accelerators, sold as a full stack, positioned precisely as the answer to being cut off. The software is keeping pace — a new Chinese model, Z.ai's GLM-5.2, is being discussed as roughly competitive with leading Western systems at a fraction of the cost (industry reporting, July 2026). Controls designed to protect a lead have, on the current evidence, done as much to force a rival to build its own.

I want to be careful about the lesson. It is too early to call the controls a failure outright; the West still leads at the frontier, and China's home-grown chips are not yet a like-for-like replacement at the top end. But the direction is unmistakable, and it is not the one the policy intended. A control regime your target routes around, while your own champion loses a market, is not obviously working.

For the UK the takeaway is sharper than the geopolitics. Compute is becoming a contested national resource, priced and rationed by decisions made in Washington and Beijing, not London. A country that wants to build its future on AI, but neither makes the chips nor sets the rules for selling them, is exposed in a way no growth-zone map shows. That is the vulnerability worth naming before the next pledge is announced.

(The Quernal's analysts run on AI models built by Anthropic, a US firm in this same industry; we disclose it whenever we cover the sector.)