What AI chip export controls are — three years of moving the line

One sentence recurs in AI coverage.

"The US has allowed / blocked exports of this chip to China."

Allowed and blocked alternate on the same part within months. Here is what the regime is, and why the line keeps moving.

1. The object of control is a number, not a company

AI chip export controls regulate a performance figure, not a corporate name.

AI accelerators fall under the second. Thresholds are drawn on measures such as compute throughput and memory bandwidth; above the line, a licence is required.

Enforcement sits with the Commerce Department's Bureau of Industry and Security (BIS).

The reason for regulating capability rather than use is dual use. The same accelerator trains a commercial model and supports military or intelligence work. Since the application cannot be separated at the point of sale, the capability is regulated instead.

Which reduces the entire design problem to one question — where to draw the line.

2. Three years of the line moving

The line keeps moving because drawing it creates a market for chips designed to sit just underneath it.

The table describes a pursuit. A threshold is drawn; a part appears beneath it; the part turns out to be useful; the threshold comes down.

3. The H20 — an episode neither side won

The H20's ending is the most instructive part of this history.

Even after permission, Nvidia sold none. Chinese authorities told domestic AI companies not to buy, citing security concerns. Nvidia stopped manufacturing them.

Two lessons came out of that.

First, export control outcomes are not determined by the exporting state alone. An importing state can close the door itself, in the service of its own industrial policy.

Second, prolonged restriction erases the market rather than shaping it. Chinese AI firms moved to domestic accelerators during the interval. That is the situation Brookings summarised by saying the United States is out of the Chinese AI chip market.

4. The current rule — case-by-case plus end-use certification

What the January 13, 2026 rule changed was the review posture.

The certification requires the exporter to confirm the chips will not be used for military, intelligence or weapons-of-mass-destruction purposes.

Moving from presumption of denial to case-by-case reads as loosening, but the real effect depends on the approval rate. Case-by-case review is also a structure in which grounds for refusal can be constructed individually. This page could not obtain approval rates or case volumes.

Then came the June 2026 interpretation: the prohibition also reaches Chinese firms located outside China. The axis of the regime shifted.

Which country it ships to → who ends up using it

That closes transshipment routes. It also raises compliance costs sharply, because an exporter must now trace the ownership structure of the end user.

5. The second layer — blocs

Export control began as a bilateral instrument: one state declines to sell one thing to another.

In 2026 it has acquired a multilateral layer.

This page covered the case on August 24 in Pax Silica letter — the US tells 35 countries to pick one AI bloc. The US-led Pax Silica framework bundles AI models, semiconductors and critical minerals, and requires members to align their export controls with one another. American controls are replicated as members' controls.

That matters because of a structural weakness in any single-country regime.

China's counter takes a different form — distributing open-weight models to third countries. Chips can be controlled at the border; downloaded weight files are much harder to stop.

6. Where suppliers sit

Producer economies touch this regime from three directions at once.

They do not always point the same way. Interests as a supplier and obligations as an ally can diverge.

This page draws no conclusion about how HBM and memory products are specifically treated — the handling has changed with each revision of the rules.

7. What is not confirmed

• Threshold values — compute and memory-bandwidth limits change with each revision; the figures in force in August 2026 were not verified against the rule text.
• How case-by-case is working — approval rates and case volumes are not in public data.
• H200 shipments — actual deliveries to China and their scale could not be confirmed.
• Component treatment — HBM and memory handling has varied by rule; no conclusion is drawn.
• Rule texts — the Federal Register versions were not read in full; the timeline comes from government statements cross-checked against multiple reports.
• Next checkpoint — publication of case-by-case decisions. Once approval rates surface, it becomes possible to say whether January 2026 was a loosening at all.