The US authorises Blackwell-class GPU sales to China

Where things stand (17 Oct 2026)

Nothing qualifies yet. I found no licence, rule or Nvidia filing since 16 Sep 2026 that authorises a Blackwell-or-newer datacentre GPU for China.

• Nvidia's latest 10-Q (filed 26 Aug 2026) says US licences cover only "small amounts of H200 products." It also says: "Under the current rules and geopolitical landscape, we are unable to create and deliver a competitive product for wide distribution in China's data center market with the approval from both the USG and the Chinese government." Its China sales are "uncontrolled products… gaming and workstation GPUs." https://www.sec.gov/Archives/edgar/data/1045810/000104581026000075/nvda-20260726.htm
• The US line is still "no Blackwell."
  • Nov 2025: the White House blocked the cut-down B30A, and Trump said "we don't give [Blackwell] to other people."
  • Jan 2026: the H200 approval explicitly excluded Blackwell and Rubin.
  • Jul 2026: Bureau of Industry and Security (BIS) chief Kessler testified that Blackwell remains barred from direct export to China. On 31 May BIS had closed the loophole for Chinese firms' overseas subsidiaries (TechTimes, 15 Jul 2026).
  • Sep 2026 Xi state visit: trade representative Greer said national-security export controls were "taken off the table" and called chips "the crown jewels… to lead in the race for superintelligence" (Chosunbiz, 26 Sep 2026).
  • Rules to curb Chinese remote access to chips are being drafted. David Sacks, the main in-house advocate for selling chips to China, has left the AI-czar role; the new czar is the intelligence chief, Jay Clayton (NPR, 4 Oct 2026).
• The RTX PRO 5500 does not count. Beijing's reported willingness (27 Sep, The Information) to let ByteDance and Alibaba buy it is irrelevant here: it is a desktop-workstation card and apparently needs no US licence.
• The January 2026 rule (FR 2026-00789) only changed review policy for chips below 21,000 TPP (a compute-performance score) and 6,500 GB/s memory bandwidth. That is not enough on its own to resolve this question. In principle it would allow case-by-case licences for low-end Blackwell server cards such as the RTX PRO 6000 Server Edition, but no such application or grant has been reported.
• Congress may write the ban into law.
  • The AI OVERWATCH Act passed the House Foreign Affairs Committee 42–2 and is in the Senate's FY2027 defence-bill managers' package. The House text would require denial of any chip at or above 21,000 TPP, and of any chip first marketed after 1 Jan 2026 at or above 4,800 TPP. The Senate version would codify the top-chip ban for 18 months.
  • The defence bill is stalled until after the 3 Nov midterms. Nvidia and AMD are lobbying the White House to strip these provisions (Roll Call, 23 Sep; Politico, 29 Sep; House text at docs.house.gov). A similar chip bill, the GAIN AI Act, was dropped from last year's final defence bill. I put roughly 40% on some statutory ban being enacted.

Reference class and model

• Past relaxations. The only qualifying-type relaxations happened under Trump: H20 licences (Aug 2025) and H200 licences (Feb 2026). For Hopper, that was about 3–3.5 years after the architecture launched. The same gap would put a Blackwell-variant approval around 2027–28, when Rubin and then Feynman lead Nvidia's line-up. That fits the administration's "sell them chips a generation or more behind" logic and Trump's own Aug 2025 openness to a Blackwell cut by 30–50%.
• Disclosure is reliable but lags. Nvidia disclosed both earlier grants in its next filing: about 6 weeks after the decision for H20, and about 11 weeks for H200 (10-K filed 25 Feb 2026).
• How I computed the series. I set a hazard for a US decision each quarter and assumed 60% of grants are disclosed in the same calendar quarter and 40% in the next (from Nvidia's late-Feb/May/Aug/Nov filing schedule), then accumulated the results (Python; quarterly hazards listed below).

Period	Decision hazard per quarter	Why
Q4 2026	1.1%	Shenzhen APEC summit (18 Nov) and Miami G20 (14 Dec), but chips are off the table
2027	1.8–2.8%	Truce deadline 10 Jan; Rubin ramps up; possible statutory ban
2028	2.4–2.8%	Blackwell is two generations old; election year
2029 onward	0.9%	New administration, probably more hawkish

Result: about 19% chance of a decision before Trump leaves office, and about 28% cumulative chance of a qualifying publication by end-2031.

Paths and the counter-case

• Paths to YES: a deal between Trump and Xi or a Huang-brokered Blackwell variant; quiet case-by-case licences for low-end Blackwell server cards; or a later administration raising the thresholds as Blackwell ages. Any China-specific or cut-down variant counts, and so does any later architecture.
• Paths to NO: a statutory ban; the "superintelligence race" framing hardening further; Beijing's self-reliance push removing the commercial case (H200 shipments were "trivial"); a Taiwan or Iran-related crisis; uncontrolled products (which count only with an explicit US determination) absorbing Nvidia's China business.
• Strongest case for higher: Trump has reversed himself twice on China chips, and over five years a policy of "sell the previous generation" would naturally reach Blackwell.
• Strongest case for lower: in 2026 the policy line has hardened, not softened, and opposition in Congress is bipartisan. I weight both, which leaves the near term very low and the cumulative five-year figure moderate.

Question. By the last day of each quarter, has the US government authorised sales to China of NVIDIA Blackwell-generation or newer datacentre GPUs?

Definitions.

• Qualifying GPUs: NVIDIA GPUs that NVIDIA identifies as using Blackwell or a successor GPU architecture, and markets for datacentre or server computing. Blackwell Ultra and subsequent architectures count. China-specific, reduced-performance and inference-focused variants count, without a minimum performance threshold. Server-edition RTX products count; consumer gaming and desktop-workstation products do not count merely because someone installs them in a datacentre. GPUs sold within servers, accelerator modules or rack-scale systems count. NVIDIA product documentation establishes architecture and intended market.
• China: mainland China, Hong Kong or Macau; not Taiwan. The authorisation must permit physical delivery for use in one of these destinations. Sales to Chinese-owned entities solely for use elsewhere, and remote access to GPUs located elsewhere, do not count.
• Authorise sales: grant at least one export or re-export licence permitting a commercial sale of qualifying GPUs, or adopt a final rule, licence exception, waiver or other legally operative measure permitting such sales without an individual licence. There is no minimum number of GPUs, buyers or dollar value. Buyer-specific approvals and approvals with quantity limits, reporting requirements or end-use conditions count. A final measure with a specified future effective date counts on publication if no further discretionary US approval is needed for the permitted transactions.

Resolves YES if a qualifying source confirms that a licence has been granted or a final authorising measure adopted, identifying the qualifying product or a product class unambiguously encompassing it and the permitted Chinese destination. Actual shipment, customer acceptance and Chinese import approval are not required. An explicit US government determination that a qualifying product may be sold to China without a licence also counts.

Resolves NO / Exclusions.

• Proposals, negotiations, promises to approve exports, licence applications, or announcements that applications will be reviewed case by case do not count without an actual licence grant or final measure authorising transactions.
• Authorisation limited to Hopper-generation products, including H20 or H200, or other pre-Blackwell GPUs does not count, even if announced after Blackwell's launch.
• Permission solely for manufacturing, testing, repair, temporary demonstration or return shipment, without a commercial sale for use in China, does not count.
• Smuggling, suspected diversion, product availability, or a seller's unsupported claim that a product complies with export controls does not establish US authorisation.
• Resolve NO for a horizon if no qualifying publication exists by that horizon. Subsequent revocation does not undo a YES under the Resolution window convention.

Resolving source. Official US Department of Commerce/Bureau of Industry and Security rules, licence announcements or written determinations; final measures published in the Federal Register or on an official US government domain; or NVIDIA-authored SEC filings (Forms 10-K, 10-Q or 8-K) and formal announcements on nvidia.com or investor.nvidia.com explicitly confirming granted US authorisation. These are the narrower resolving sources for Public confirmation (A2); executive interviews, leaks and press reports alone do not count. Relevant baseline sources are BIS's Department of Commerce Revises License Review Policy for Semiconductors Exported to China (January 13, 2026) and Revision to License Review Policy for Advanced Computing Commodities (Federal Register, January 15, 2026, document 2026-00789); a change in review policy alone is insufficient. Apply the Resolution window convention and publication dating under Resolution date (A3): only qualifying publications on or after 2026-09-16 count, including a first qualifying confirmation of an earlier licence grant.