AI Chip Export Rules Keep Nvidia’s Best Chips Banned In China

The US Commerce Department has rewritten AI Chip export controls at least five times since 2022, putting Nvidia and AMD’s ability to sell identical parts in China or Japan at stake.

A chip that is perfectly legal to sell in Japan can be a federal crime to sell in China, even though it is the exact same part with the same serial number format. That split is not about tariffs or sanctions in the traditional sense.

It is the result of a dense, overlapping set of AI Chip export controls that the US Commerce Department has rewritten at least five times since 2022, each time trying to close a gap that Nvidia (NVDA), AMD (AMD) or a foreign buyer had found in the previous version.

Understanding how these rules actually work, not just that they exist, is the only way to make sense of why some AI deals get approved in weeks and others get stuck for years.

How An AI Chip Becomes An Advanced Computing Item

The starting point for every AI Chip export control is a classification system, not a chip. BIS, the Commerce Department agency that administers the Export Administration Regulations, assigns every controlled item an Export Control Classification Number, or ECCN. In October 2022, BIS created new ECCNs specifically for advanced computing integrated circuits and the computers built around them, according to the Federal Register notice that launched the modern regime.

That 2022 rule did not name Nvidia or any specific chip. Instead, it set two technical trip wires. The first is total processing performance, a measure of raw computational throughput.

The second is performance density, which is processing power divided by die area, a proxy for how much compute is packed into a given chip footprint. Cross either threshold and the chip automatically falls into the controlled category, regardless of what the datasheet calls it or who designed it.

A chip becomes “controlled” the moment its processing performance or performance density crosses a numeric line BIS sets, not when a regulator decides to ban a product by name.

This is why Nvidia’s A100 and H100 chips became the first high-profile casualties of the rule. They were not targeted by name. They simply exceeded the thresholds BIS had drawn.

That forced Nvidia to design cut-down variants like the A800 and H800 specifically to land just under the line for China. BIS closed that loophole in October 2023 by tightening the density threshold and adding a second, lower performance band that catches those downgraded chips too, a change documented in the October 2023 rule.

For chip designers, those thresholds function like another physical limit alongside power budgets, thermal limits and available capacity. The A800 and H800 show that a regulatory number can reduce performance available per unit of inference even when demand for compute has not changed.

Also Read: Nvidia Eyes Reflection AI Stake To Lock In Chip Demand After $1B SpaceX Deal

Country Groups Decide Who Needs A License At All

Once a chip is classified as a controlled advanced computing item, the next question is where it is going. The Export Administration Regulations sort destination countries into groups, and the group that matters most here is Country Group D-5, which covers countries subject to a US arms embargo plus a handful of others flagged for diversion risk. China and Macau sit inside this group.

A license is required to export advanced computing items to entities headquartered in Country Group D-5 destinations, under rules BIS maintains and explains on its own site. That single sentence is the backbone of the entire China chip restriction story.

It does not say “no chips to China.” It says any company headquartered in a D-5 country needs a license before it can receive the controlled part, and BIS reviews those license applications individually.

For most other countries, including US allies in Europe and much of Asia, no license is needed at all for standard commercial sales below certain thresholds. That is the part people miss when they call this a “China ban.”

It is a licensing requirement layered on top of a classification system, and licensing requirements can be approved, denied or approved with conditions. In June 2026, BIS issued fresh guidance clarifying that this license requirement for D-5 and Macau-headquartered entities remains fully in force even after an earlier “AI diffusion” framework was shelved, a point confirmed in legal analysis of the guidance.

The underlying license rule never went away. Only the broader diffusion framework around it did.

For operators buying an AI Chip, the practical constraint is not simply whether silicon exists. It is whether approved supply can be allocated to a buyer, at a given performance level, before a policy change alters the cost and availability of that compute.

The Entity List Adds A Second, Sharper Filter

Country groups control geography. The Entity List controls specific organizations, regardless of where else in that country chips might legally flow. BIS maintains this list of named companies and institutions that face additional or absolute restrictions on receiving US-origin technology, often because of suspected military ties or past violations.

A chip buyer can be headquartered in a country with no blanket restriction at all and still be legally blocked from buying advanced computing items if that specific entity appears on the list. This is the mechanism that has been used against Chinese supercomputing centers, certain cloud providers and AI labs flagged for links to military end-users.

The Congressional Research Service traces this layering of country controls and entity-specific bans back to a policy shift that began in earnest in 2018 and accelerated sharply afterward, as its report on advanced semiconductor controls lays out. The practical effect is that a license granted for sales to “China” in general tells you almost nothing. The real decision happens at the level of the individual buyer.

A cloud company with no flagged ties might get a license, while a research institute one subsidiary removed from a flagged military university might not, even if it is in the same city.

> Two buyers in the same country, buying the identical chip, can get opposite answers from BIS because one of them appears on the Entity List and the other does not.

Also Read: KONST Bags $30M From ADATA To Expand AI Data Centers Across Asia

The Foreign-Produced Direct Product Rule Closes The Factory Loophole

Export controls only work if they cover more than direct shipments from US soil. Nvidia designs its chips in the US but manufactures almost all of them at TSMC in Taiwan using equipment that traces back to American and allied toolmakers.

Without a rule reaching into that supply chain, a US export ban on Nvidia would be nearly meaningless, since the actual silicon never physically leaves a factory on US territory. The Foreign-Produced Direct Product Rule solves that by extending US jurisdiction to any foreign-made chip that was produced using specified US-origin technology, software or equipment, even if no American company ever touched the finished product.

BIS expanded and refined this rule again in December 2024, adding new chip categories and tightening the scope of what counts as a “direct product” of controlled US technology, according to the Federal Register rule issued that month.

This is also the legal hook that lets Washington restrict semiconductor manufacturing equipment itself, not just finished chips. An April 2024 clarification built on the October 2023 interim final rules covering both the chips and the manufacturing tools used to make them, as detailed in that update.

The result is a control regime that follows the technology through the supply chain rather than stopping at the US border. That makes fab capacity and tool access part of the AI Chip control system, not merely background to it.

When Rules Reverse, The H200 And The Revenue-Share Deal

Export controls are not static law in the way a tariff schedule is. They are policy instruments that the executive branch can tighten or loosen within months, and 2025 and 2026 produced the clearest demonstration of that yet.

In August 2025, the US government allowed Nvidia and AMD to sell certain advanced AI chips to China on the condition that Washington receive 15% of the resulting China revenue, an arrangement described in coverage of the policy shift.

That reversal did not mean every chip became exportable. It meant a specific, negotiated carve-out opened for specific products under specific financial conditions, while the underlying licensing architecture described above stayed in place for everything else.

Months later, the administration had issued licenses covering roughly 750,000 H200 chips, but those licenses remained stalled in practice, a bottleneck examined in discussion of the stalled H200 deal.

A license being granted and a shipment actually leaving a warehouse are two separate events, and the gap between them can run into the hundreds of thousands of units. For AI Chip buyers, that gap is a capacity constraint. Planned compute does not lower inference costs or add serving capacity until the hardware arrives.

Then, in January 2026, BIS revised its license review policy again, this time specifically for advanced computing commodities headed to China and Macau, tightening the standard reviewers apply when a new application comes in, per the January 2026 rule. Companies that got comfortable with the 2025 opening had to recalibrate within months. This is the defining feature of the regime, durable mechanics, unstable settings.

The country groups, Entity List and compute thresholds are the permanent machinery. What flows through that machinery changes with the political weather.

A chip that was exportable under a August 2025 revenue-share arrangement can become un-exportable again under a January 2026 policy tightening, without a single word of the underlying classification rule changing.

How Nvidia And AMD Actually Disclose This Risk

Export control exposure is not an abstraction for either chipmaker. It shows up as a named, recurring risk factor in securities filings, because both companies’ China-exposed revenue depends on a regulatory decision neither of them controls.

Nvidia’s fiscal 2026 annual report states plainly that export controls increase the risk of investing in US advanced computing companies, a line carried in its Form 10-K filing for the fiscal year ended July 2026. AMD’s own 2025 annual report lists US export restrictions on semiconductors and semiconductor technology to China as a standing operational risk, language that appears in its Form 10-K covering the period ending Dec. 27, 2025.

Smaller companies downstream of the two chipmakers carry the same disclosure. A supercomputing hardware maker’s 2025 filing describes its own exposure to “restrictions imposed through the Bureau of Industry and Security” on AI chips and advanced computing technology sold to certain countries, as stated in its 10-K.

The pattern across every filing is the same, revenue guidance now comes with an explicit caveat that a BIS decision could redirect it overnight. This matters for anyone reading an earnings call or a revenue forecast involving either chipmaker. When Nvidia or AMD say China revenue is “uncertain,” they are not hedging about demand.

Demand in China for the best available AI chips is effectively unlimited. The uncertainty is entirely regulatory, and it is driven by the same licensing and entity mechanics described above, not by anything happening inside either company.

For investors, spending and shipment capacity deserve more weight than broad strategy language. A company cannot convert stated AI Chip demand into revenue or lower the cost per unit of inference if licenses and deliveries do not clear.

Who Actually Needs To Track These Rules Closely

Not every reader of this piece needs to monitor the Federal Register weekly, but several groups genuinely do, and the stakes differ sharply between them. Chip buyers and cloud operators headquartered in Country Group D-5 nations need to track both the country-level license requirement and their own status relative to the Entity List, since a change in either can shut off supply without warning.

Investors in Nvidia, AMD or any company with meaningful China-exposed AI hardware revenue need to treat every quarterly filing’s export control risk language as a live variable, not boilerplate, given how often the policy itself has shifted in the past two years.

Hardware engineers and chip architects at firms like Nvidia and AMD face a different problem entirely. Because the controls are defined by measurable performance and density thresholds rather than by product names, chip design teams now treat the BIS thresholds as a hard design constraint alongside power budget and thermal limits.

The A800 and H800 were not marketing decisions. They were engineering responses to a regulatory line drawn in teraflops per square millimeter.

Also Read: Samsung Locks 80% Of 2027 Chip Supply As AI Memory Demand Soars

Conclusion

Watch BIS’s next license-review guidance, Entity List changes and decisions on individual China and Macau applications, then compare those actions with actual shipments of the roughly 750,000 H200 chips already licensed.

The operative constraint is not Chinese demand but whether approved supply can leave the factory under the policy in force that week. For Nvidia and AMD, that is the variable that can redirect revenue, capacity and AI Chip availability overnight.

Read Next: Chai Discovery Wins GSK Deal After Passing Real Wet-Lab Drug Test