The export-control arc: October 2022 to the Diffusion Rule and back

Round 1 — October 2022: the performance thresholds

BIS drew the line with two technical parameters — chip-to-chip interconnect bandwidth and total processing performance. NVIDIA’s response arrived within weeks: the A800 and H800, chips engineered to sit just under the interconnect threshold while keeping most of the training performance. Entirely legal, openly sold, and a masterclass in why parameter-based controls invite parameter-targeted engineering.

Round 2 — October 2023: closing the loophole, adding density

BIS rewrote the rules to catch the A800/H800, replaced the interconnect test with total processing performance and performance density metrics, added a gray-zone notification tier, and extended controls to dozens of countries seen as transshipment risks. NVIDIA responded again: the H20, a China-market chip under the new lines — weaker for training but genuinely strong for inference. The cat-and-mouse pattern was now structural.

Round 3 — December 2024: HBM and the toolmakers

The third round targeted high-bandwidth memory — the component modern accelerators are starved without — plus additional semiconductor-manufacturing equipment and some 140 entity-list additions. The logic had shifted visibly upstream: if chips leak, control the ingredients of chips.

Round 4 — January 2025: the AI Diffusion Rule

In its final week, the Biden administration published the Framework for Artificial Intelligence Diffusion — the most ambitious compute regulation ever attempted. It sorted the world into three tiers: ~18 close allies with near-unrestricted access; a broad middle tier subject to country-level compute caps and validated-end-user licensing; and arms-embargoed states (China, Russia, and others) essentially excluded. For the first time it also controlled the export of certain closed model weights — treating the trained model itself, above roughly 10²⁶ FLOPs of training compute, as a controlled item.

Round 5 — May 2025: rescission and improvisation

Days before its compliance date, the Trump administration rescinded the Diffusion Rule, calling the tiering unworkable and insulting to middle-tier partners, and promised a replacement built on bilateral deals. What followed instead was policy by episode: guidance warning that using Huawei Ascend chips could violate US export law; an April 2025 license requirement that halted NVIDIA H20 sales to China (a multi-billion-dollar write-down); then a mid-2025 reversal reportedly permitting H20 sales in exchange for a share of China revenue paid to the US government — an arrangement with no clear precedent in export-control law.

The frontier-policy years

• 2022-10-07 — US advanced-chip export controls on China:

  Sweeping controls on advanced semiconductors and manufacturing equipment — compute becomes an explicit instrument of AI policy.

• 2023-10-17 — US tightens chip export controls:

  The 2022 rules are extended to close performance-density loopholes and cover more countries — the compute-control net widens.

• 2025-01-13 — AI Diffusion Rule issued — then rescinded:

  A three-tier global framework for AI chip and model-weight exports, issued in the administration’s final week and rescinded by its successor in May 2025 — export policy whiplash.

• 2025-02-01 — Frontier safety frameworks become table stakes:

  Following the Seoul commitments, major labs publish or update frontier safety policies (capability thresholds, evaluation gates, deployment mitigations) ahead of the Paris summit.

• 2026-06-01 — EO 14409 on frontier-model cybersecurity:

  Federal attention narrows to security of frontier models — weights, infrastructure, adversarial threats. Check current status for implementing rules.