US chip export controls reversed twice in five months
American export policy on AI chips loosened in January 2026 and tightened again in May. Both moves were made by the same administration, five months apart.
Anyone trying to plan a hardware purchase against that is planning against a moving target, and the direction of travel is less obvious than either headline suggests.
The January loosening
On 13 January the Department of Commerce published a regulation permitting sales of advanced AI chips to China, codifying a policy change announced the previous month.
The rule loosened restrictions on Nvidia’s H200 and AMD’s MI325X, both previously banned for export, and it arrived with an unusual attachment: a 25% export tax on the cleared H200 sales.
An export tax on chips is a striking instrument. It treats access to American silicon as something to be priced rather than simply permitted or denied.
A Bureau of Industry and Security rule effective 15 January put H200 and MI325X licences under case-by-case review, while the Blackwell-class B30A stayed blocked outright.
The Council on Foreign Relations was blunt about the design, calling the new policy strategically incoherent and unenforceable, which is a strong verdict on a rule that had barely taken effect.
The May reversal
Four months later the direction changed, and the mechanism changed with it.
Rules announced in late May 2026 require export licences for transfers of Nvidia’s most advanced processors to any entity headquartered in China or Macau, CNBC reported.
Read the wording carefully, because that’s the substantive shift. The test is where a company is headquartered, not where the chips are delivered.
That closes the workaround where a Chinese firm bought restricted hardware through an overseas subsidiary, and the ban now follows the company rather than the border.
Where chip export controls currently sit
| Hardware | Status for China | Since |
|---|---|---|
| Nvidia H200 | Case-by-case licence, 25% export tax | 15 January 2026 |
| AMD MI325X | Case-by-case licence | 15 January 2026 |
| Blackwell-class B30A | Blocked | Throughout |
| Top Blackwell parts to China-HQ firms anywhere | Licence required | Late May 2026 |
That shape is the actual policy, once you strip the headlines away. Sell the previous generation, price it, and keep the current one at home.
Whether that’s coherent depends on your theory of what the controls are for. As a way of keeping a capability lead it’s defensible; as a way of preventing Chinese AI development it clearly is not, since a generation-old chip still trains models.
What it has cost Nvidia
The financial effect is unusually easy to see, because a public company has to disclose it.
Nvidia took a $5.5 billion writedown and estimated $14 to $18 billion in lost annual sales, and by its Q1 FY2027 reporting the company carried no China data-centre revenue in its outlook at all.
Zero is the number worth pausing on. A company forecasting nothing from a market that size has concluded the policy is not going to stabilise in its favour, and its annual filing sets out the risk in its own terms.
That sits oddly next to the company’s position elsewhere in the industry, where it helped arrange up to $500 billion for AI infrastructure. Demand is not the constraint; permission is.
Why chip export controls keep moving
Two reversals in five months looks like indecision, and it’s better understood as three constituencies pulling in different directions.
National security officials want the tightest possible restriction. American chipmakers want access to one of the largest hardware markets on earth. Trade negotiators want leverage they can trade away.
The 25% export tax is what a compromise between those three looks like written into a regulation. It permits the sale, takes a cut, and keeps the option of withdrawing permission later.
None of those pressures went away in May, which is the reason to expect further movement rather than a settled position.
The argument that controls backfired
The strongest case against this policy is not that it’s unfair but that it produced the opposite of its intent.
Restricting hardware created enormous pressure to do more with less, and the efficiency work that followed has been genuinely impressive. DeepSeek’s R1 results came from reinforcement learning applied at scale rather than from a bigger cluster.
Open weights compound that. A restricted lab can build on models released publicly by others, and capable open models keep arriving under permissive licences with no export control attached to a download.
Chips are also fungible in ways policy struggles with. Cloud capacity rented outside the restricted jurisdiction achieves much of what buying the hardware would, which is the enforcement gap the May rule was written to narrow.
The case for keeping them
The counter-argument is that a delay is worth having even if it isn’t a wall.
Training frontier models is compute-bound, and a lab running a generation-old fleet pays more and moves slower than one that isn’t. Compounded across years, that’s a real gap rather than a rounding error.
The efficiency response also has limits. Better training methods help everyone, including the labs with unrestricted hardware, so the relative position may not have moved as much as the headlines about efficient training suggest.
What this means if you’re buying
Most companies reading this are not exporting chips, and the rules still reach procurement decisions.
Corporate structure now matters more than location, so a subsidiary of a China-headquartered parent faces the restriction wherever it operates. Enterprise guidance on chip rules and model choice works through the practical checks.
And build for portability. Two reversals in five months is a strong argument against architecture that assumes any specific hardware will remain available to you.
Cloud capacity is the practical hedge for most teams, since a provider absorbs the procurement risk and you rent whatever is currently legal for it to operate.
The same instability runs through domestic AI rules, where companies are planning against a state patchwork that could be preempted at any point. Uncertainty is the common cost.
What that hardware buys is the other half of the story, and our piece on what a frontier training run actually costs covers why compute access translates so directly into capability.
Watch whether the May rule holds through the year. Policy that swings this fast tends to swing again, and the January reversal came with as much confidence as this one did.
Get the daily rundown
One email each weekday with the AI news that matters, every claim linked to its primary source.
Free, one email each weekday, unsubscribe in one click. We never sell or share your address.
