NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Case by Case: How the H200 License Window Was Designed — and Where It Stalled

A January 2026 rule moved Nvidia H200 and AMD MI325X exports to China from presumption of denial to case-by-case review, with a 50 percent shipment cap, KYC requirements and third-party testing — yet nearly a year later, commercial deliveries remain stuck between two governments.

2026-11-18 · 941 words · NeuroAI
Case by Case: How the H200 License Window Was Designed — and Where It Stalled

On paper, the door opened in January. The US Commerce Department's Bureau of Industry and Security (BIS) revised its license review policy for advanced computing commodities, moving exports of certain chips to China — including Hong Kong and Macau — from "presumption of denial" to "case-by-case review." The rule took effect upon publication in mid-January, and it is the framework that still governs every H200 question raised on Capitol Hill today.

Nearly a year later, the arithmetic of that framework is worth rereading, because it explains why the window exists, what qualifies, and why the pipeline has mostly stayed empty.

What Qualifies, and What the Conditions Are

The policy applies to chips below two hard thresholds: total processing performance (TPP) under 21,000 and total DRAM bandwidth under 6,500 GB/s. Nvidia's H200 and AMD's MI325X are the named examples. Anything above those lines remains locked out.

Qualifying is not automatic. An exporter must certify a bundle of conditions, several of which are unusually explicit:

  • US supply first. The product must be in adequate commercial supply in the United States, and exports to China must not delay existing or new orders from US end customers, nor divert global foundry capacity that would otherwise serve US customers for comparable or more advanced products.
  • The 50 percent cap. Total shipments to China and Macau may not exceed 50 percent of the same product's shipments to US domestic end customers. The rule hard-codes a ceiling into the license itself: China can never be the bigger market for the chips in question.
  • KYC at the final consignee. The receiving party must implement know-your-customer screening and security safeguards designed to prevent unauthorized parties from obtaining remote access — the clause aimed directly at datacenter rental workarounds.
  • Per-shipment testing. Every batch must be reviewed by a qualified, independent third-party testing laboratory in the United States to verify the performance specifications described in the license application.
  • End-user disclosure. Applicants must list remote end users located in — or headquartered in — Belarus, China, Cuba, Iran, Macau, North Korea, Russia and Venezuela.

The exclusions are equally telling. Re-exports and in-country transfers to Macau or D:5 country-group destinations remain under presumption of denial, as do cases where the ultimate parent is headquartered in those jurisdictions. When an application could fall under multiple policies, the stricter one wins.

The Timeline Behind the Rule

The sequence, as documented in Taiwanese and Chinese trade press citing Reuters and official publications: in mid-December 2025, the Trump administration launched an interagency review of the H200 sales proposal; on January 13, 2026, the approval was formalized and BIS published the revised policy; the rule entered force on January 15.

Chinese state media, including Beijing Daily's Chang'an Jie Zhi Shi (长安街知事) commentary account, reported that the approval terms included Nvidia handing the US government a quarter of sales revenue; a companion January presidential proclamation separately imposed a 25 percent tariff on certain semiconductor imports, and analysts read the two as linked. That framing comes from Chinese official accounts and a law-firm review of the proclamation, not from a US government summary — treat the exact mechanism as contested.

The Window That Mostly Stayed Shut

What happened next is the part that matters for anyone parsing current congressional debate. In late January, Reuters reported that China had cleared ByteDance, Alibaba and Tencent to purchase more than 400,000 H200s combined, but that final conditions were still being confirmed and none of the approvals had converted into orders. By March, Nvidia had stopped producing H200s for the Chinese market and redirected capacity toward its next-generation Vera Rubin platform, per Reuters. Through the spring and summer, terms remained deadlocked between the two governments, and Commerce Secretary Howard Lutnick said no H200s had actually been sold to Chinese firms, citing their difficulty obtaining Beijing's import approval.

Two licensing regimes — one American, one Chinese — now sit on either side of the same chips. The US side caps volume and demands verification; the Chinese side, per reporting, pressured domestic buyers to assess demand cautiously and weighed conditions of its own. Export control and import control met in the middle and produced a stalemate.

That stalemate is the context for November's legislative push. Congressional hawks who call the January framework too loose are negotiating against a baseline in which the framework, however permissive on paper, has produced almost no commercial flow. Critics in Congress, meanwhile, want to tighten remote-access rules precisely because the case-by-case regime leaves datacenter rental channels open — the subject of a separate interim rule that BIS moved on in mid-November.

Honest limitations

  • The 25 percent "revenue share" condition rests on Chinese state-media framing and a related presidential proclamation; the US government's own summary of the license policy does not state it in those terms, so treat the mechanism as reported rather than confirmed.
  • Shipment-volume figures, including the 400,000-chip authorization, come from Reuters reporting via Taiwanese press; no party has published order-level data.
  • Compliance guidance here is descriptive, not legal advice; the Federal Register text and qualified counsel govern any actual transaction.
  • The policy landscape shifted repeatedly through 2026, including November's remote-access rule; any snapshot ages quickly.

What readers can do now

  • Read the actual Federal Register rule — the TPP and DRAM-bandwidth thresholds, the 50 percent cap and the KYC clauses are the four numbers that determine whether any specific chip qualifies.
  • If you track Nvidia's China exposure, separate announced policy from converted orders; the gap between them has defined the entire year.
  • Watch whether Congress folds the Remote Access Security Act into the NDAA — that would change the rules this license framework operates under, from shipment-based to access-based control.

Related coverage

Cite this article

NeuroAI (脑机网), "Case by Case: How the H200 License Window Was Designed — and Where It Stalled," 18 November 2026, https://neuroai.site/a/na-chips-nvidia-h200-casebycase-50cap

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Chips & Compute” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories