The United States restricts advanced AI chip sales to China through export controls enforced by the Bureau of Industry and Security (BIS) under the Export Administration Regulations (EAR), aiming to slow China's AI and military computing capacity. Controls are set by compute-performance thresholds, cover specific chips from NVIDIA and AMD, and extend to equipment and named entities such as Huawei and SMIC.
Why does the US restrict AI chip sales to China?
The United States uses export controls to slow China's advances in AI and military computing, enforced by the Bureau of Industry and Security (BIS) under the Export Administration Regulations (EAR)CITE:E1. This legal and institutional structure — a named regulation (EAR) administered by a named agency (BIS) — is the foundation for every specific restriction described belowCITE:E1.
How do controls use compute performance as the threshold, and how have the rules evolved?
Controls set licensing requirements based on a chip's computing performance and interconnect bandwidth, with chips above certain thresholds facing a presumption of denialCITE:E2. Chipmakers previously produced downgraded, China-specific variants — NVIDIA's A800 and H800 — to fall below those thresholds, and regulators subsequently tightened the rules to close that pathCITE:E2.
Which AI chips are banned today, and which face case-by-case review?
The current chip-by-chip status spans outright bans, license requirements, and case-by-case review with added tariffs, based on EffectStory's own tracking dataCITE:E3.
| Chip | Status | Effective Date | Additional Measure |
|---|
| H100 / A100 | Banned | — | — |
| H20 (China-specific variant) | License required | 2025-04 | — |
| H200 | Case-by-case review | 2026-01 | +25% tariff |
| Blackwell (B200 / GB200) | Banned | Expands to overseas subsidiaries in 2026 | — |
| AMD MI series | Similar restrictions to NVIDIA equivalents | — | — |
H100 and A100 remain banned outright, while H20 — NVIDIA's China-specific variant — has required a license since 2025-04CITE:E3. H200 moved to case-by-case review starting 2026-01, alongside a 25% tariffCITE:E3. Blackwell-generation chips (B200/GB200) are banned, and that ban is set to expand to overseas subsidiaries in 2026CITE:E3. AMD's MI series faces similar restrictionsCITE:E3.
Why are "ban," "tariff," and "revenue-sharing" often confused?
Three distinct mechanisms get conflated: an already-effective ban/case-by-case regime, an uncodified revenue-sharing arrangement, and a rescinded rule that never took effectCITE:E4. First, there is the licensing and review regime already described — bans and case-by-case approvals that are currently in forceCITE:E4. Second, there is a 15% revenue-sharing arrangement tied to H20 and AMD's MI308, which NVIDIA has described in its own financial disclosures as not codified into law, with its constitutionality disputedCITE:E4. This is separate from the 25% tariff applied to H200-class chips, which was established through a January 2026 presidential proclamation and does have a legal basisCITE:E5. Third, the Biden-era AI Diffusion Rule was rescinded in 2025-05 and never actually went into effectCITE:E4. Keeping these three apart — codified tariff, uncodified revenue share, and a rule that was scrapped before implementation — is necessary to understand the current state of controlsCITE:E5.
What does export control cover beyond chips?
Control extends past chips to equipment and named companies, creating layered restrictions across the supply chainCITE:E6. The US simultaneously restricts export of advanced equipment such as EUV lithography systems to China, and has placed companies including Huawei and SMIC on the Entity ListCITE:E6. Together, chip-level, equipment-level, and entity-level restrictions form a multi-layer blockadeCITE:E6.
What does this mean?
The chip-by-chip record shows a pattern of escalation rather than a single fixed line: A800/H800 workarounds prompted tighter thresholdsCITE:E2, H20 moved from unrestricted to license-required in 2025-04CITE:E3, and H200 moved to case-by-case review with a 25% tariff in 2026-01CITE:E3, with Blackwell's ban set to reach overseas subsidiaries later in 2026CITE:E3. At the same time, the 15% revenue-sharing arrangement sits outside this codified track entirely — NVIDIA itself has flagged it as uncodified and constitutionally contested, distinct from the tariff that does have a presidential-proclamation basisCITE:E4CITE:E5. The equipment and Entity List measures extending to Huawei and SMIC show the same control logic applied beyond chips aloneCITE:E6.
Author's Take・林紀旭 James Lin
The clearest signal in this record is the split between the two percentage-based measures: the 25% tariff on H200-class chips traces to a January 2026 presidential proclamation with a legal basis, while the 15% revenue share on H20 and MI308 remains uncodified and constitutionally contested by NVIDIA's own account. That gap matters more than either number alone — one mechanism sits inside the regulatory system, the other sits outside it on unresolved legal footing. The chip-by-chip timeline (H20 licensed in 2025-04, H200 shifted to case-by-case with a tariff in 2026-01, Blackwell's ban set to reach overseas subsidiaries later in 2026) shows the codified track tightening steadily. The one to watch is whether the 15% revenue-sharing arrangement ever gets folded into that same codified structure, or stays a standalone, disputed arrangement running in parallel to it.