The US drops export licences for advanced chips to China

How I read the question

To resolve YES, a binding rule, statute or presidential measure must remove individual licensing to China for every chip in the 3A090.a/.b class as it stood on 16 Sep 2026. That class covers everything with TPP ≥ 4,800, so it includes Blackwell, Rubin and any future top chip. Approving licences, presumptions of approval, notification-based exceptions, buyer-specific authorisations and threshold increases all fail. In practice, YES means the US gives up licensing AI chips to China altogether.

Where things stand (as of 12 Oct 2026)

• No qualifying publication. I queried the Federal Register API for BIS documents since Dec 2025 and also searched by keyword ("advanced computing", "3A090"). The only advanced-computing rule I found is 2026-00789 (91 FR 1684, 15 Jan 2026). Nothing relevant has been published since 16 Sep. The one AI-related presidential order since then ("Inaugurating the Era of Super Intelligence", 2 Oct) only changes terminology. That January rule kept licensing. It moved chips below TPP 21,000 / 6,500 GB/s from presumption of denial to case-by-case review, with heavy conditions, and left a presumption of denial above that level.
• The administration is tightening, not loosening.
  • On 31 May 2026 Commerce closed a loophole for Chinese-headquartered firms buying abroad (CNBC).
  • It is drafting a rule on remote access to chips (Tom's Hardware/The Information, Aug 2026).
  • It is using aircraft-part licences to COMAC as leverage in trade talks (Reuters/Asia Times, Oct 2026).
  • Smuggling prosecutions are continuing.
• Trump's stated position. He has said the most advanced chips are reserved for the US (Reuters, 2 Nov 2025). USTR Greer said national-security export controls were "taken off the table" at the September summit. None of the eight summit deliverables mentions chips (Geopolitechs readout; The Information via Investing.com).
• China is not pressing for this. Beijing has throttled even approved H200 purchases (Brookings, Jun 2026; Reuters/CNBC, Jul 2026).
• Congress is moving the other way. The Senate NDAA manager's amendment includes the AI OVERWATCH Act. It would write a licence requirement for these chips to China into law and bar general licences (H.R. 6875 text; Roll Call, 23 Sep 2026). The NDAA is stalled until after the midterms. Even Nvidia and AMD are lobbying only to change these provisions, not to roll back China controls broadly (Politico via Stocktwits, 29 Sep 2026).

Reference class and base rate

The comparison is full, class-wide decontrol of a frontier strategic technology for a principal rival. That has historically needed a geopolitical realignment, such as the end of the Cold War or 1980s US–China alignment. Otherwise controls were loosened by raising thresholds while the top end stayed licensed. Examples are supercomputer controls since the 1990s, post-Tiananmen controls and satellites. On that basis I put the chance at a few percent at most over five years.

Paths to YES and NO

• YES:
  • A Trump–Xi grand bargain around the Jan 10, 2027 truce expiry, APEC or a later summit that explicitly lifts chip controls.
  • Trump, influenced by Huang, deciding the controls have failed.
  • A future administration abandoning the controls.
• NO:
  • Licensing is how the administration collects its revenue share, sets conditions and keeps leverage.
  • There is bipartisan hawkishness, and a statute may lock licensing in.
  • China's self-reliance policy means it is not asking for the change.
  • Every past relaxation kept licences in place.

Rough decomposition of the five-year chance:

• Grand bargain: about 0.5–0.75%.
• Unilateral Trump decontrol: about 0.4–0.8%.
• Post-2029 reversal: about 0.3–0.6%.
• Other: about 0.2%.

That sums to roughly 1.5–2.5%.

Hazard series

I computed the series in Python as one minus the cumulative product of survival.

• Q4 2026: 0.10%.
• Q1 2027: 0.15%, because of the truce deadline.
• Through the end of Trump's term: 0.11% per quarter.
• 2029–31: 0.08–0.09% per quarter.

Cumulative: 0.1% by end-2026, about 1.0% by end-2028, about 2.0% by end-2031.

Strongest case against

Trump reversed the H20 ban within months and approved the H200 over strong objections, and Bessent says he is "completely aligned" with Huang. Five years is long, and AI may change the strategic picture. This is why I don't go below about 1%. Still, even the most dovish moves so far kept licensing, so I don't think a materially higher number is justified.

Question. By the last day of each quarter, has the US government removed the requirement for an export licence to sell advanced AI chips to China?

Definitions.

• Advanced AI chips: integrated circuits within ECCN 3A090.a or 3A090.b, or the corresponding advanced-computing integrated circuits classified under other ECCNs’ .z paragraphs, using the Export Administration Regulations (EAR), 15 CFR part 774, supplement no. 1, as they stood on 2026-09-16. This fixes the covered technical class even if the classifications are subsequently renamed or removed.
• China: mainland People's Republic of China; changes confined to Hong Kong or Macau do not count.
• Sell to China: commercial exports from the United States of chips subject to the EAR for delivery to and civilian use by customers in mainland China, including Chinese-headquartered customers. A contract to sell, offshore cloud access, and sales for use outside China are not exports to China for this question.
• Remove the requirement: adopt a legally binding measure eliminating the generally applicable requirement for individual export licences across the entire defined chip class for ordinary civilian commercial sales to China. Repeal, removal of China from the relevant destination controls, or a generally available licence exception requiring no case-by-case government permission counts. Recordkeeping, reporting, and automatic registration requirements may remain. Restrictions on specifically listed entities, military or intelligence users, and prohibited weapons end uses may remain.

Resolves YES if an officially published final or interim final rule, enacted statute, or binding executive measure makes the removal defined above legally operative or unconditionally provides for it on a specified effective date. The measure must cover the whole defined chip class, not merely particular models, performance bands, suppliers, or approved buyers. Equivalent licensing requirements imposed elsewhere in US law must not continue to require individual permission for those ordinary civilian commercial exports. An explicitly time-limited general suspension counts if it meets these conditions; a mere enforcement moratorium does not.

Resolves NO / Exclusions.

• No qualifying measure has been published by the quarter-end.
• Proposed rules, political commitments, trade-deal announcements requiring implementing action, guidance promising non-enforcement, or relaxation of penalties do not count.
• Approving licences, adopting a presumption of approval, or changing review from a presumption of denial to case-by-case review does not count. In particular, BIS’s January 15, 2026 final rule, Revision to License Review Policy for Advanced Computing Commodities (91 FR 1684; document 2026-00789), does not remove the licence requirement for H200 and equivalent chips.
• Raising the control threshold while retaining licensing for higher-performance chips within the defined class does not count. Buyer-specific authorisations, temporary general licences limited to selected transactions, and exceptions under which the government must first approve a shipment do not count.
• Replacing licensing with an outright export prohibition does not count.

Resolving source. The operative legal text published in the Federal Register by the Department of Commerce’s Bureau of Industry and Security (BIS), or an enacted statute published by Congress/GPO on Congress.gov or GovInfo, or a legally operative presidential instrument published in the Federal Register. EAR provisions to examine include 15 CFR §§ 742.6 and 744.23, part 740, and part 774, supplement no. 1, together with successor provisions. These are narrower resolving sources than Public confirmation (A2); press reporting and agency press releases alone cannot resolve. Apply the Resolution window and Resolution date (A3): use official publication of the qualifying legal measure, not an earlier announcement or its later effective date. Once YES, subsequent reinstatement does not reverse it.