The chip export ban: who's winning, who's losing, and where the rest of the world stands in line

The premise sounded clean. If China can't buy the world's most powerful AI chips, it takes longer to catch the United States in artificial intelligence. That was the bet behind the export controls Washington tightened through 2025. A year and a half on, the bet has gotten expensive and the scoreboard is a mess. Nvidia itself calls the policy a failure, part of Congress wants to harden it again, and the Trump administration ended up turning a national-security barrier into a revenue stream. Worth following, because the side effects reach far past Beijing.

The timeline nobody can keep straight

Start at the beginning, because the whiplash is part of the story. In January 2025, under the Biden administration, the Bureau of Industry and Security (BIS), the arm of the Commerce Department that decides what the US can and can't export, published the AI Diffusion Rule, a framework that sorted the world into three country tiers and set caps on raw compute.¹ It was meant to take effect in May.

It never did. On April 9, 2025, the government told Nvidia that selling the H20, the chip Nvidia had purpose-built to stay inside the rules and still ship to China, now required a license.² AMD got the same treatment on its MI308. In May, Commerce simply rescinded the AI Diffusion Rule before it ever bit, saying it stifled American innovation.³

Then came the strangest turn. In August 2025 the White House confirmed that Nvidia and AMD had agreed to hand 15% of their revenue from AI-chip sales in China back to the US government in exchange for the licenses.⁴ Trump said he'd asked for 20% and "negotiated" down to 15%.⁵ The awkward constitutional footnote: the US Constitution bars taxes on exports. Derek Scissors of the American Enterprise Institute put it bluntly: "There's no precedent for this, probably because export taxes are unconstitutional."⁶

December brought another reversal. BIS began allowing case-by-case sales of stronger chips (Nvidia's H200, AMD's MI325X) to China, under conditions like independent third-party testing in the US and buyer-side compliance.⁷ And in June 2026, the bureau closed a loophole: the licensing requirement applies to "all businesses with headquarters or a parent company in China," even if the subsidiary sits physically abroad.⁸ Blackwell, the flagship, stays barred from anyone Chinese-owned.

Ban in April, charge a toll in August, loosen in December, plug a leak in June. Small wonder the technical crowd trusts none of it.

Who's losing: Nvidia, in the short run

Nvidia's cost has a number on it. The company booked a $4.5 billion charge in the first quarter of fiscal 2026, tied to H20 inventory and purchase commitments that the restriction turned worthless, and left another $2.5 billion of H20 revenue unshipped that quarter.⁹ For the following quarter, guidance projected losing roughly $8 billion in H20 revenue.⁹

The caveat that keeps the story honest: Nvidia is nowhere near hurting. Total revenue that first quarter was $44.1 billion, with $39.1 billion in data center alone.⁹ The China hole is real, but it's a hole in a company whose demand everywhere else won't quit. CEO Jensen Huang has said Nvidia's share of the Chinese market collapsed and that export policy "has already largely backfired."¹⁰

Who's winning: the China that was forced to improvise, and TSMC

Here's the core irony. By shutting Nvidia's door, the US pushed Chinese buyers toward local suppliers, chiefly Huawei, and handed Beijing the best possible argument to pour state money into a homegrown supply chain. A report from ITIF, a Washington tech think tank, says it in the title: the controls "helped Huawei and hurt US firms."¹¹

But Chinese self-sufficiency carries a big asterisk, and the better-informed corner of the debate won't let it slide. Teardowns by TechInsights, the firm that cracks chips open to see what's inside, found that nearly every Huawei Ascend accelerator they examined used dies (the silicon wafers themselves) fabricated on TSMC's own 7-nanometer node, not China's SMIC, obtained through sanctions-evasion schemes.¹² And there's a physical bottleneck money can't fix overnight: HBM, the high-bandwidth memory every AI GPU needs, which Chinese makers still can't produce at scale.¹³ China is running. It hasn't arrived.

And whoever sells the factory to both sides gets paid either way. TSMC, the Taiwanese foundry that physically builds chips for Nvidia, AMD, and, through the side door, Huawei, closed Q4 2025 with $33.7 billion in revenue, up 26% year over year, and set a monthly record in January 2026.¹⁴ High-performance computing, which includes AI, became 58% of sales.¹⁴ The house always wins.

The second-order risk: everyone who isn't in the fight

The part the headlines skip. The export ban is a US-versus-China fight. Most of the world (Brazil, India, the Gulf, Southeast Asia) isn't a target, isn't named, isn't at the table. Which is exactly why the risk to them is easy to miss.

Take Brazil as the worked example. It makes no advanced logic chips. None. Every piece of silicon in a Brazilian data center, whether server CPU, GPU, or AI accelerator, is imported from Taiwan, South Korea, the US, and Israel. Incentive programs like PADIS gave tax breaks for design and assembly but never attracted modern-node fabrication; local design houses stay in analog and power management.¹⁵ Meanwhile data-center investment is booming: AWS announced roughly $1.8 billion to expand its operations in the country,¹⁶ and Rio AI City, from Elea Data Centers, was unveiled as Latin America's largest campus, with Oracle and Nvidia signing memorandums of intent.¹⁷

Put the two ends together and the problem shows. That whole build-out runs 100% on a single chain nobody outside it controls: Nvidia designs, TSMC fabricates. When global GPU demand already outstrips the fab, the northern-hemisphere hyperscalers set the order of the queue. A country off the US-China axis has no leverage over allocation or price. The export ban doesn't target these markets, but it tightens the global supply they draw their AI silicon from. It's a risk of empty shelves and high prices, not of sanctions.

What the community is saying

Across hardware and geopolitics communities the mood is skepticism and frustration, not hype, and the word that keeps coming back is "backfire": the sense that the controls did the opposite of what they promised. When the 15% deal landed, the tone curdled into mockery, and in a rare moment the partisan lines blurred. National-security hawks and free-market types hammered the same point for opposite reasons.

In geopolitics forums, the take that sticks is that measuring success by "how many chips we didn't sell" is the wrong metric; the right one is "how fast China goes independent." In hardware circles the skepticism cuts both ways: neither "the controls worked" (China found another road) nor "China is already self-sufficient" (the teardowns show TSMC dies and an HBM bottleneck) survives scrutiny. And the applied-AI crowd raises the angle that travels furthest: "what about the rest of the world that also buys Nvidia?" The worry is allocation. If global demand won't fit in the fab, everyone off the US-China axis ends up at the back of the line. All of this is community opinion, not fact, and tellingly, the markets most exposed to that allocation risk barely come up in the debate at all.

Verdict

The export controls had a clear goal, slowing Chinese AI, and delivered a tangled outcome: China was pushed to build its own chain, Nvidia ate a billion-dollar loss while still growing, TSMC profited from both sides, and the US government ended up charging a toll on the very barrier it raised. For anyone watching from outside the two combatants, the lesson isn't to root for a side in the chip war. It's to reckon with dependence: the AI build-out everywhere is rising on silicon imported 100% from a single chain, at the exact moment that chain is the most contested asset on the planet. The risk out here isn't a sanction. It's being last in line when the line gets long.