Loading…
Loading…
Base rate computed from analyst-asserted responds_to: edges in the reverse direction (target-country → issuer-country) for prior issuer-actions on the same target. Modal type + lag percentiles only — not a model output. Treat as a historical anchor for sizing counter-response scenarios, not a forecast in itself.
The rule amends EAR Supplement No. 1 to Part 742 (CCL-based country-chart license review policies) for ECCN 3A090.a and 3A090.b advanced-computing ICs destined to China and Macau. Where the prior policy (set by the 7 October 2022 rule and expanded by the 17 October 2023 rule) defaulted such items to a presumption of denial, the new rule allows case-by-case review when the chip falls within a defined performance band:
This band is explicitly named in the preamble as covering the NVIDIA H200 (Hopper-class HBM3e SKU) and AMD MI325X (CDNA-3 SKU) — i.e. the previous-generation flagship AI accelerators, not Blackwell B200 / B300 / GB200 NVL72 or MI355X-class chips, which remain under presumption of denial.
To clear case-by-case review, four exporter certifications are required:
1. US supply sufficiency — US-based end-users will not be delayed and the export will not divert global foundry capacity from US-bound orders. 2. 50% volume cap — aggregate TPP of advanced-node ICs exported to China/Macau under this pathway will not exceed 50% of domestic shipments by the same exporter. 3. End-use / end-user compliance — confirmation that items will not be diverted to military end-use, military intelligence, prohibited Entity-List recipients, or to chip design / production facilities in China. 4. Independent third-party testing — prior to export, a qualified US-headquartered lab must confirm the technical capabilities and functions of the AI commodities, either per-unit or via representative batch sampling.
This is the first structural reversal — not just an enforcement adjustment — of the four-year US advanced-computing perimeter. It re-opens the legal sale of last-generation flagship AI accelerators to PRC hyperscalers (Alibaba Cloud, Tencent Cloud, Baidu, ByteDance, Huawei Cloud) under license, conditional on geopolitical reciprocity. Severity is rated 4 rather than 5 because: (i) cutting-edge Blackwell-class silicon stays under denial, (ii) the 50% volume cap and US-supply-sufficiency certification reserve a material share of capacity for US-domestic AI buildout, and (iii) the third-party testing gate creates a friction layer that the December 2024 HBM package and the rescinded January 2025 AI Diffusion Framework did not have.
were stranded by the December 2024 HBM package and Oct-2023 performance-density metric become exportable to China under license — partial recovery of the FY2024 China data-center revenue lost to controls.
for Alibaba / Tencent / Baidu / ByteDance AI training clusters at the H200 tier, partially relieving demand pressure on Huawei Ascend 910C / 910D.
domestic advanced-AI silicon by re-introducing a legal H200 supply, marginally negative for SMIC / CXMT capex acceleration.
unilaterally from the Japan METI / Netherlands ASML positions, which retain pre-existing controls — first crack in the trilateral chip-equipment perimeter built 2022-2024.
Biden AI Diffusion Framework on 13 May 2025; this rule fills part of the policy vacuum left by that rescission with a bilateral-leverage architecture rather than a tier-1/2/3 global perimeter.
Taiwan) align their own license-review policies with the new US case-by-case posture, or will the trilateral perimeter fragment?
statistics or per-license disclosure? The rule is silent.
the corporate group, or per legal entity per ECCN? Compliance practice will need BIS clarification.
follow into a similar case-by-case band once a successor generation (Rubin / MI400) is in volume — the rule's structure is generation-following, not absolute.