AI Diffusion Rule Withdrawal and What Replaced It — Five Checks Japanese Companies Should Make Now

Hello, this is Ryuta Hamamoto from TIMEWELL.

The "AI Diffusion Rule," which was meant to systematize U.S. AI export controls in one stroke, was rescinded on May 13, 2025, only two days before it would have taken effect. As of May 2026, roughly one year later, no successor rule has been published.

Over the past several months I have received a visible uptick in questions about what this means for a company's AI development and export-control program. I wrote this for people who are new to export control: original structure, path to withdrawal, substitute measures, and what Japanese companies should do in practice.

What this piece covers

• What the AI Diffusion Rule was designed to regulate, with the key terms explained
• The timeline through the May 13, 2025 rescission and the three reasons BIS gave
• The meaning of the three substitute guidance documents issued the same day (GP10, AI training end-use controls, and the diversion red-flag set)
• Why the phrase "15% fee on exports to friendly countries" is a misunderstanding, and why the correct target is China-bound H20 / MI308
• Five practical issues Japanese companies should check now

Three terms to learn first

Before the main discussion, three minimum terms.

(1) AI Diffusion Rule (Framework for Artificial Intelligence Diffusion). The export-control rule for AI semiconductors and AI model weights published by the U.S. Department of Commerce's Bureau of Industry and Security (BIS) on January 15, 2025. BIS administers the Export Administration Regulations (EAR). In Japanese terms it is closest to METI's Security Trade Control Division.

(2) Tier classification. The core of the AI Diffusion Rule was a three-tier structure: Tier 1 for trusted partners (Japan, the UK, Germany, France, and 19 countries/regions in total), Tier 2 for most other countries with quantitative caps, and Tier 3 for China, Russia, and other destinations subject to a near-total ban.

(3) FLOPs threshold. FLOPs stands for Floating-Point Operations, a unit of computational volume used to train an AI model. The AI Diffusion Rule would have controlled the weights of closed-weight AI models trained at 10^26 FLOPs or more. That threshold sits above GPT-4-class systems and would have captured only a handful of developers worldwide. Closed-weight means model parameters are not released externally (ChatGPT, Claude, Gemini, and similar). Open-weight models (Llama, DeepSeek-R1, and similar) were out of scope.

The original structure of the AI Diffusion Rule

The rule has been rescinded, but understanding its structure still matters for reading any successor framework.

Publication history and form

The formal title was "Framework for Artificial Intelligence Diffusion." Publication date: January 15, 2025. Scheduled compliance date: May 15, 2025 (Federal Register citation 90 FR 4544). The form was an Interim Final Rule (IFR), a fast-track device that allows a rule to take effect while comments are still being received, without a prior notice-and-comment cycle. The choice of form itself signaled urgency around AI controls.

The three-tier structure

The core design divided the world into three tiers.

Tier	Character	Main destinations
Tier 1	Trusted partners; largely unrestricted	Japan, the United States, the UK, Germany, France, Korea, Taiwan, Australia, and others — 19 countries/regions in total
Tier 2	Quantitative caps; case-by-case review	India, Singapore, the UAE, Saudi Arabia, and most other countries
Tier 3	Near-total ban	China, Russia, North Korea, Iran, Cuba, Belarus, and others

Singapore's placement in Tier 2 drew attention at the time, and the split of EU member states between Tier 1 and Tier 2 produced a wave of diplomatic concern.

Technologies in scope

Two technology clusters were in view. (a) High-performance AI semiconductors such as NVIDIA H100, H200, Blackwell, and AMD MI300X (ECCN 3A090.a and related). (b) Closed-weight AI model weights — for the first time, AI models themselves would have been controlled under a new ECCN 4E091. The threshold was closed-weight models trained at 10^26 FLOPs or more. Open-weight models were out of scope. It was a symbolic moment in which controls moved from hardware into software.

Industry pushback

NVIDIA, Oracle, Microsoft, and governments including Germany, Spain, and Poland publicly opposed the rule. The main arguments were three: (1) license exceptions such as UVEU, NVEU, LPP, and NAC were too layered and compliance cost was excessive; (2) placing much of the EU in Tier 2 damaged alliance relationships; and (3) if Tier 2 countries could not obtain U.S. AI semiconductors, demand would shift toward Chinese alternatives such as Huawei, ultimately benefiting China's AI ecosystem.

Path to withdrawal

The AI Diffusion Rule was rescinded just short of its May 15, 2025 compliance date, after a change of administration.

The announcement date was May 13, 2025, only two days before compliance would have begun. Internal instructions went to BIS enforcement not to enforce the rule, followed by a public rescission announcement (the formal Federal Register notice followed weeks later).

BIS listed three reasons in its press release:

1. Overly bureaucratic — the compliance burden on companies was too heavy
2. Stifled American innovation — the rule would slow AI development and exports that the United States should lead
3. Damaged relations with allies — treating dozens of countries as "second tier" eroded trust

The language of the rescission closely tracked the industry critiques. Commerce Secretary Howard Lutnick stated that U.S. allies can purchase AI semiconductors, provided certified U.S. data-center operators and certified cloud operators are involved. My reading: a policy switch driven by industry and ally pushback, not merely a post-election housecleaning.

Substitute measures: three guidance documents

On the same day as the rescission, May 13, 2025, BIS published three guidance documents. Those documents are the current operational baseline.

(1) GP10 guidance on Huawei Ascend chips

The guidance warns that using, selling, transferring, financing, or providing services for Huawei AI chips such as Ascend 910B / 910C / 910D carries a high risk of violating U.S. export controls.

The key mechanism is General Prohibition 10 (GP10), the tenth of the EAR's general prohibitions. It bars the sale, transfer, or service of an item when a party knows that an EAR violation has occurred, is about to occur, or is intended. Two points matter: (1) GP10 applies to non-U.S. persons as well as U.S. persons (including Japanese companies handling items inside Japan); and (2) the standard is "knew or had reason to know," so even without deliberate intent, overlooking objective red flags can create violation risk. When Huawei AI chips are used in operations, the company itself bears a duty to confirm they were not produced in violation of U.S. export controls.

(2) Guidance on items for AI training

When semiconductors or related items are exported, reexported, or transferred for advanced AI model training, or when a U.S. person supports such training, a violation may arise if the party knows or has reason to know the end use is for China, Russia, or other countries of concern. The scope covers not only semiconductors and server hardware, but also engineering support, dataset provision, and preprocessing services under an end-use lens. The important point is that controls reach "services," not only "goods."

(3) Industry counter-diversion red-flag set

The third document functions as a due-diligence playbook. It flags indicators such as "final delivery or installation address not specified," "data-center operator cannot confirm infrastructure requirements," "abnormally large volumes of high-performance GPUs ordered on short notice by operators with no track record," and "payment routed through multiple third countries," and asks companies that see these signals to investigate further or walk away. It is a practical document that can be built into routine screening.

Important correction: "15% fee on friendly countries" is a misunderstanding

In the context of the AI Diffusion Rule withdrawal, the "15% fee" is often misstated. Clear the record.

Media and social posts sometimes claim that "the United States now charges a 15% fee on AI chip exports to friendly countries." That is not correct. What actually happened is a China-bound arrangement, announced on August 10, 2025, under which licenses to export NVIDIA H20 and AMD MI308 to China require payment of 15% of sales to the U.S. government. President Trump initially sought 20%; NVIDIA CEO Jensen Huang proposed 15%, and the parties agreed. In December 2025 an extension toward H200 and related products at 25% was indicated, and it was formalized by presidential proclamation on January 14, 2026.

In other words, the deal is China-bound, not friendly-country-bound. Internal materials must state the destination country explicitly. Separately, the U.S. Constitution bars export taxes and federal law restricts fees on export licenses, so constitutional debate continues.

Alongside the rescission, large "deal-by-deal" arrangements also advanced. During President Trump's May 2025 Middle East visit, the UAE was cleared to import up to 500,000 units per year of leading-edge NVIDIA AI semiconductors for 2025–2027, and Saudi Arabia's Humain received an immediate shipment of 18,000 NVIDIA Blackwell (GB300) units. Emblematic cases of deal-based administration.

Five impacts on Japanese companies

Impact 1: The Tier 1 auto-approval premise disappeared. The expectation that Japan-based users could freely procure AI semiconductors as Tier 1 parties is gone; the ordinary EAR individual-license regime is back. Predictability is lower, but the Tier 1 benefit never actually took effect, so practical harm is limited.

Impact 2: Comprehensive controls on closed-weight AI model exports are absent. The license requirement for closed-weight AI model weights at 10^26 FLOPs or more, including ECCN 4E091, was rescinded. Japanese frontier-model developers are currently outside a comprehensive framework, but a similar structure could return in a successor rule.

Impact 3: New GP10 risk on Huawei Ascend use. If a Japanese company deploys Chinese AI semiconductors (Huawei 910B/C/D and similar) while knowing they were produced in violation of U.S. export controls, GP10 risk can attach. Typical scenes include a Chinese subsidiary deploying Huawei Ascend, AI development support for Chinese customers on Huawei environments, and data centers for the company's own AI services that depend on Huawei. BIS has imposed substantial civil penalties for GP10 violations; pre-clearance is essential.

Impact 4: New end-use management on AI training support contracts. Supplying semiconductors, items, or support services for AI training to customers in China, Russia, or related entities falls under the new guidance. Cloud API access, joint research, and engineer dispatch can all be "services that contribute to AI training," regardless of form.

Impact 5: Japan's own FEFTA controls are unchanged. METI issued a notice in May 2023 on advanced semiconductor manufacturing equipment. That is an independent Japanese regime, not a derivative of the U.S. AI Diffusion Rule. The withdrawal does not change Japanese classification or licensing workflows under the Foreign Exchange and Foreign Trade Act.

Honestly, Impact 3 and 4 are where I see Japanese teams under-prepared. Tier 1 nostalgia is less dangerous than silent GP10 exposure.

Outlook for a successor rule (unpublished as of May 2026)

BIS said at the time of rescission that it would publish a successor rule "in the future," but as of this writing (May 2026) no schedule has been set. The working assumptions for practitioners are: (1) for now, the guidance and red-flag set are the operational standard; (2) administration will combine individual licensing with large deals (UAE, Saudi Arabia, and others); and (3) a successor rule may be simpler and more ally-respecting than the original Tier design, while still controlling AI semiconductors and model weights in broad terms. The present state is best summarized as "we do not know when it will appear, but when it does, everything can change at once."

Five practical steps

For export-control officers at Japanese companies, five steps to start now.

Step 1: Strengthen counterparty location and bona fide demand checks. Build the BIS red-flag set into internal due diligence. Add final delivery address, consistency between declared end use and order specifications, and payment-path complexity to the screening sheet for new transactions.

Step 2: Inventory Chinese AI semiconductor handling. Confirm whether the company, subsidiaries, or counterparties use Huawei Ascend 910B/C/D or similar. Companies with Chinese sites especially need visibility into actual server procurement. If you cannot immediately answer "we do not use them," inventory introductions by semiconductor vendor once.

Step 3: Screen AI training support contracts. Check whether end users are Chinese or Russian companies or related entities. Treat cloud APIs, dataset provision, and engineering support as "services," and make end-user collection and verification a required item at classification time.

Step 4: Internal education on GP10 exposure. Because GP10 turns on "knew or had reason to know," front-line staff need to recognize risk signals. GP10 overview training, a clear escalation path for "stop when you notice something," and internal publication of a translated red-flag summary, run once over a six-to-twelve-month cycle, meaningfully raise organizational sensitivity.

Step 5: Monitor BIS successor-rule developments. Track BIS press releases, the Federal Register, and major law-firm alerts on a regular cadence. The point is to be ready to revise internal procedures the day a successor rule appears.

If I had to pick one checklist item this week, it would be Step 2. Most teams still cannot answer semiconductor inventory questions without a multi-week hunt.

Common misconceptions / FAQ

Q1. Was the AI Diffusion Rule ever enforced? No. It was rescinded on May 13, 2025, two days before the compliance start date.

Q2. Have China-bound exports been relaxed? Not in a general sense. China-bound exports of NVIDIA H20 and AMD MI308 resumed in August 2025, but only under the 15%-of-sales payment condition to the U.S. government. Controls on higher-performance chips (H100, H200, Blackwell, and similar) toward China remain. The GP10 warning on Huawei Ascend was also added.

Q3. What is the problem if a Japanese company buys Chinese AI semiconductors? Purchase itself is not banned under U.S. law. But using, reselling, or servicing those chips while knowing they were produced or acquired in violation of U.S. export controls creates GP10 risk.

Q4. When will a successor rule appear? BIS has said it will publish one "in the future," but as of this writing (May 2026) no schedule has been announced.

Q5. Are open-weight AI models (DeepSeek-R1 and similar) controlled? Neither under the AI Diffusion Rule nor under post-rescission practice are open-weight model weights themselves a direct export-control object. Separately, the semiconductors and data-center operations that run those models may still be controlled.

Latest developments as of July 2026

There has been no major new publication on U.S. AI export controls themselves in the immediate period, but related economic-security moves have accelerated. The 16th Japan–India annual summit on July 2, 2026 produced a joint declaration on cooperation in five fields — semiconductors, critical minerals (rare earths), clean energy, ICT (subsea cables), and pharmaceuticals — with investment on the order of about 2 trillion yen (Japan–India summit joint press conference (Prime Minister's Office of Japan, July 2, 2026)). Bundling semiconductor and critical-mineral supply chains among allies and partners aligns with the U.S. "deal-by-deal + end-use management" approach discussed above. My reading is that as AI semiconductor procurement networks are reorganized under intergovernmental frameworks, end-user screening at Japanese companies becomes more, not less, important. Related developments are summarized in the 2026 Japan–India summit and economic security.

If you are interested in improving export-control operations or streamlining classification work, review the functional overview in the TRAFEED service catalog (PDF) or contact us.

Summary

Do not read "the rule was rescinded" as "controls disappeared." Shifting the center of gravity to end-use management and GP10 readiness is the most rational posture at present.

• The AI Diffusion Rule was rescinded on May 13, 2025, two days before it would have taken effect
• As substitutes, BIS published three documents: GP10 guidance, AI training end-use controls, and a counter-diversion red-flag set
• The "15% fee" targets China-bound H20 / MI308, not friendly countries
• For Japanese companies, the Tier 1 auto-approval premise disappeared, while GP10 risk around Huawei Ascend became a new issue
• As of this writing (May 2026), no successor rule has been published; guidance is the operational baseline for now

The withdrawal was a turning point at which U.S. AI export control shifted from a comprehensive framework toward case-by-case licensing plus end-use management. Monday morning: confirm whether any Huawei Ascend (or similar) sits in your estate, then map AI training support contracts to end users before the next board asks "are we clean?"

Reducing the burden of AI-related export control with TRAFEED

Export control around AI semiconductors, models, and services has become too information-dense for pure manual coverage: ECCN classification, end-use checks, red-flag assessment, and more. TIMEWELL's TRAFEED (formerly ZEROCK ExCHECK) is an export-control AI agent that assists classification, automatically screens against the BIS Entity List and Country Group D:5, checks counterparties for red flags, and provides METI-aligned multilingual support. A 30-minute online consultation walks through how the product applies to your company.

Related articles

• The AI Diffusion Rule Withdrawal and H200 China Release — Three Axes of Export Control in the Generative-AI Era
• China–Japan Export Control Trends in 2026
• Complete Guide to the BIS Affiliates Rule (50% Rule)
• China Export Risk 2026

References

• BIS, "Rescission of Biden-Era Artificial Intelligence Diffusion Rule" (May 13, 2025): https://www.bis.gov/press-release/department-commerce-announces-rescission-biden-era-artificial-intelligence-diffusion-rule-strengthens
• BIS, "General Prohibition 10 Guidance" (May 13, 2025): https://www.bis.gov/media/documents/general-prohibition-10-guidance-may-13-2025.pdf
• BIS, "AI Counter-Diversion Industry Guidance" (May 13, 2025): https://www.bis.gov/media/documents/ai-counter-diversion-industry-guidance-may-13-2025.pdf
• Federal Register, "Framework for Artificial Intelligence Diffusion" (January 15, 2025): https://www.federalregister.gov/documents/2025/01/15/2025-00636/framework-for-artificial-intelligence-diffusion
• Federal Register, "Promoting the Export of the American AI Technology Stack" (EO 14320): https://www.federalregister.gov/documents/2025/07/28/2025-14218/promoting-the-export-of-the-american-ai-technology-stack
• Morrison Foerster, "AI Diffusion Rule Out, But BIS Increases Compliance Expectations": https://www.mofo.com/resources/insights/250617-ai-diffusion-rule-out-but-bis-increases-compliance
• Akin Gump, "BIS Rescinds its AI Diffusion Rule and Issues Compliance Guidance": https://www.akingump.com/en/insights/alerts/bis-rescinds-its-ai-diffusion-rule-and-issues-compliance-guidance-regarding-advanced-computing-items
• CBS News, "Nvidia, AMD chip sales to China subject to 15% fee": https://www.cbsnews.com/news/nvidia-amd-chip-sales-china-15-percent-h20-mi308/
• RAND (L. Heim), "Understanding the AI Diffusion Framework": https://www.rand.org/pubs/perspectives/PEA3776-1.html