Short answer
The US Commerce Department requires a licence to export chips above set performance limits to China and certain other countries, and it has rewritten those rules several times since October 2022. Since January 2026, chips at the level of Nvidia's H200 can get case-by-case review for China under strict conditions, while Beijing's own restrictions have kept actual sales small.
Follow-up question Do these rules affect buying AI chips or compute in Canada?
An export control is a government rule that says certain goods cannot leave the country, or go to certain places or people, without permission. For AI chips, the rule-maker is the Bureau of Industry and Security (BIS), part of the US Department of Commerce. The rules can also cover re-exports from other countries and buyers whose parent company is based in a restricted country.
The stated aim is national security: to limit the computing power available to countries of concern for military and other sensitive uses. Critics, and Nvidia in its filings, argue that the rules also hand business to foreign competitors and encourage rivals to build their own alternatives.
How a chip gets classified
BIS does not ban chips by name. It sets technical limits, so that new chips are caught automatically. The first version, published in October 2022, used two tests together: how fast a chip can send data to other chips, and how much processing power it has.
The second test is a measure called total processing performance, or TPP. Roughly, it is the number of operations per second, in trillions, multiplied by the size of each number in bits. Here is a worked example using the 2022 limits, with a made-up chip:
| Question | Illustrative chip | 2022 limit |
|---|---|---|
| Processing speed | 1,000 trillion operations per second on 8-bit numbers | |
| TPP | 1,000 × 8 = 8,000 | 4,800 or more |
| Chip-to-chip data rate | 700 GB per second | 600 GB per second or more |
| Result | Needs a licence for restricted destinations |
Chipmakers responded by building versions just under the line, and regulators responded by moving the line. Nvidia's latest filing says the current China rules look at total processing performance, performance density (power relative to chip size), interconnect bandwidth and memory bandwidth, and that it cannot now offer a competitive data centre product across China with approval from both governments.
The rules so far
| Date | What changed |
|---|---|
| October 2022 | First BIS limits on advanced computing chips going to China |
| October 2023 | Requirements widened to more countries in the Middle East and Asia, and to companies headquartered in restricted countries |
| January 2025 | The AI Diffusion rule proposes a worldwide licence system with three tiers of countries |
| April 2025 | Nvidia is told it needs a licence for its H20 chip, and records a $4.5 billion charge |
| May 2025 | Commerce announces it will rescind the Diffusion rule, with a replacement to follow |
| December 8, 2025 | President Trump says the H200 and similar chips may go to approved customers in China |
| January 2026 | BIS moves those chips from a presumption of denial to case-by-case review, and a 25% tariff applies to certain advanced chips |
| February 2026 onward | The US grants licences for small amounts of H200 |
Nvidia's August 2026 filing says the replacement for the Diffusion rule has not yet appeared, and its scope and timing remain uncertain.
The H200 rule in detail
"Presumption of denial" means an application is expected to be refused. "Case-by-case" means it gets weighed on its merits. Under the January 2026 rule, the new treatment covers chips with a TPP below 21,000 and total memory bandwidth below 6,500 GB per second, such as the Nvidia H200 and AMD MI325X. The applicant must certify, among other things, that:
- enough supply exists in the US that American orders will not be delayed;
- foundry capacity is not being diverted from similar or better chips for US users;
- combined exports to China and Macau stay at or under 50% of what US customers received;
- the buyer screens its customers to prevent unauthorized remote access;
- every shipment passes through a third-party testing lab in the US first.
Because the chips must go through the US for inspection, Nvidia says they are subject to the 25% tariff when they enter the country, and that it has not been able to pass that cost to customers.
The other government in the room
The licence is only half of the sale. Nvidia's filing says Chinese government restrictions have limited both its H20 and H200 sales. It took a $0.4 billion H200 charge in the first half of fiscal 2027, has made only a fraction of the shipments allowed, and those shipments were under 1% of data centre revenue last quarter. Its forecast for the current quarter assumes no data centre compute revenue from China.
That points to the core debate. Supporters of tight controls argue that they slow China's access to the most capable systems. BIS defended the looser H200 policy as permitting sales under controlled conditions that strengthen the American technology base. Skeptics note that China is developing its own chips, including Huawei's Ascend, and that DeepSeek reported training a strong model on 2,048 H800 chips, a reduced-capability Nvidia part it could still buy under the earlier limits. Both things can be true: limits on compute raise costs and slow scale, and they also push a large country to build around them. Congress is active as well: the Senate passed the GAIN AI Act as part of a defence bill in October 2025, and Nvidia's filing lists that among the possible sources of new restrictions.
Try it: Pick a chip and find its published speed and bit width. Multiply them as in the table above, and compare the result with the 21,000 line. Treat your answer as an estimate, because the official calculation includes technical details this simple version leaves out.
Do these rules affect buying AI chips or compute in Canada?
Not directly, for most Canadian organizations today. Nvidia's description of the current country rules centres on China and country groups that take in places such as Saudi Arabia, the UAE and Vietnam. When the 2025 Diffusion rule tried to sort the whole world, its text put Canada on the short list of favoured destinations, along with Australia, Japan, the UK and others, but that rule was rescinded.
Three indirect effects remain. A Canadian subsidiary of a company headquartered in a restricted country can face the same limits as its parent. Reselling or renting out compute to restricted parties can raise compliance questions, so cloud providers should get legal advice. And Canada's access rests on US policy that has changed often. The replacement rule is still pending, which is part of why Ottawa talks about sovereign compute at all, even though the chips in a Canadian data centre are usually American designs. This article is general information, not legal advice.
Sources
- Revision to License Review Policy for Advanced Computing Commodities — US Bureau of Industry and Security, January 13, 2026
- NVIDIA Form 10-Q, quarter ended July 26, 2026 — Nvidia (SEC), August 26, 2026
- Implementation of Additional Export Controls: Certain Advanced Computing and Semiconductor Manufacturing Items — Federal Register (BIS), October 13, 2022
From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0: credit the guide, link to https://guidetoartificialintelligence.com/en/learn/ai-chip-export-controls/, no commercial use without permission, and keep the same licence.