AI in China Update: MOFCOM Clarifies AI Export Control List — Key Takeaways

Definition: On May 15, 2025, China’s Ministry of Commerce (MOFCOM – 商务部, Shangwubu) published an official clarification to the AI Export Control List (人工智能出口管制清单, rengong zhineng chukou guanzhi qingdan), adding eight new subcategories under dual-use surveillance and generative AI technologies. This marks the first major revision since the original list was issued in December 2023, and directly affects any foreign company operating a WFOE (外商独资企业, waishang duzi qiye) or joint venture that develops or uses advanced AI in China. The clarification provides clearer definitions, reduces ambiguity on license exemptions, and sets stricter thresholds for “intended end use” declarations. For senior executives navigating cross-border AI operations, the document’s 47-page supplement contains binding compliance directions that will reshape licensing timelines and technology transfer agreements for years to come.

Why This Matters

MOFCOM’s clarification is not a routine administrative update. It directly implements the State Council’s dual-use export control framework (两用物项出口管制条例, liangyong wuxiang chukou guanzhi tiaoli) and aligns China’s controls with emerging technologies such as large language models, AI training accelerators, and high-performance computing clusters. Foreign executives must understand that any cross-border transfer of AI-related software, hardware, or technical data—including cloud-based model fine-tuning—now falls under tighter scrutiny. Missed compliance steps can result in license revocations or fines up to 15% of annual revenue, as explicitly referenced in the accompanying enforcement guidelines.

The updated list distinguishes between two original categories and eight new subcategories, bringing the total number of controlled AI technologies to 34. Compared to the 2023 list, the scope has expanded by 31%, and the definition of “export” now includes electronic transmission of AI model weights. For companies already licensed under the old list, the clarification introduces a mandatory re-application window of 60 days from the date of publication, with a grace period ending July 14, 2025.

This news article provides a structured analysis of the key changes, the affected technologies, and the strategic decisions your company should make now.

Key Changes in the Clarified AI Export Control List

MOFCOM’s clarification primarily focuses on four areas: definition refinement, category expansion, licensing procedures, and end-use declarations. Below we break down each area and its practical impact on foreign companies in China.

1. Expanded Scope: 8 New Subcategories

The original 2023 list contained 26 AI-related control items under two broad headings: “AI Training Software & Algorithms” and “AI Inference Hardware.” The clarification splits each original category into four more granular subcategories, adding eight new entries. Notable additions include:

• Multimodal large language models with more than 100 billion parameters (including training, checkpoint, and fine-tuning datasets).
• AI training accelerators with floating-point performance exceeding 10 TFLOPS (single precision) or 20 TOPS (integer).
• Neural-network pruning and distillation software used to compress models for edge deployment.
• Synthetic data generation tools used for adversarial resistance training.
• High-bandwidth memory (HBM) modules specifically designed for AI workloads (capacity ≥ 24 GB, bandwidth ≥ 3 TB/s).

The inclusion of HBM modules is especially significant because it brings semiconductor supply chains under the AI export control umbrella, overlapping with existing semiconductor controls from 2023. A company importing or exporting HBM3e or similar memory for AI servers must now obtain dual-use licenses from both MOFCOM and the Ministry of Industry and Information Technology (MIIT).

2. Strengthened “End-Use” Declarations

One of the most impactful changes is the clarification of intended end-use statements (最终用途声明, zuizhong yongtu shengming). Previously, companies could submit a general end-use declaration for broad categories like “academic research” or “product development.” The new rules require a specific, project-level end-use form listing the exact technology, the recipient entity (including ultimate parent company), and a description of how the AI will be used—down to the model architecture and training dataset source. If the end use changes even slightly (e.g., moving from inference to further training), the company must apply for a new license. This dramatically increases administrative burden for companies that frequently update models or pivot use cases.

3. Shorter License Validity & Mandatory Audits

The clarification reduces the standard export license validity from 24 months to 18 months, with a mandatory compliance audit every 9 months (down from 12). This effectively forces companies to undergo two compliance audits per license cycle. Additionally, the grace period for existing license holders to reapply is only 60 days—from May 15 to July 14, 2025. Given that MOFCOM’s own processing times average 45–60 days, companies must act immediately to avoid a licensing gap that could halt AI exports.

4. Transfer of AI Model Weights via Cloud & API

The clarification explicitly defines “export” to include electronic transmission of AI model weights or training data, including through cloud services and APIs. This means a company with a WFOE in Shanghai that uses Alibaba Cloud’s AI platform to fine-tune a model for an overseas client likely needs an export license if the model weights or training data are transferred outside mainland China. Even if the model is accessed only via API, if the computation occurs on Chinese soil and the weights are subsequently shared with a foreign entity, the transaction is considered an export. This directly impacts SaaS providers, cloud AI hubs, and cross-border AI-as-a-service platforms.

Pitfalls: What Foreign Executives Must Watch For

Assumption that “Open Source” Means Uncontrolled

A common misconception is that open-source AI models are exempt from export controls. MOFCOM’s clarification explicitly states that “open-source distribution of model weights may still fall under these controls if the technology is classified as dual-use and the destination country or end user raises proliferation concerns.” Companies relying on platforms like Hugging Face to share fine-tuned models must check whether the base model is on the control list, and whether the fine-tuning process altered the model’s performance to exceed the new thresholds. For example, a base LLaMA-3 model (70 billion parameters) is below the 100-billion threshold, but if a company fine-tunes it using proprietary data and the resulting model achieves >100 billion effective parameters through ensemble techniques, the exported version may be controlled. Always err on the side of applying for an advisory opinion from MOFCOM’s Technology Control Division.

Ignoring the “Ultimate End User” Clause

The clarification strengthens the “catch-all” clause for dual-use items. If a foreign subsidiary or joint venture in China transfers AI technology to a parent company overseas that later re-exports to a sanctioned entity (e.g., entities on the Ministry of Commerce’s embargo list), the original exporter (the WFOE) can be held liable. The liability extends to the Chinese legal representative of the foreign company. In 2024, MOFCOM issued penalties against three WFOEs that failed to conduct proper end-user due diligence, with fines equal to 8–10% of their China revenue. The clarified list emphasizes that the exporter must “know or have reason to know” the ultimate destination. Therefore, due diligence must now include screening end users and recording the purpose of use for every AI-related shipment or transfer.

Last-Minute License Applications Will Cause Bottlenecks

With a 60-day re-application window affecting all existing license holders, and processing times averaging 50 days, MOFCOM’s licensing department is likely to be swamped through August 2025. Companies that wait until late June may face a license gap of several weeks, during which they cannot export any controlled AI technology. To mitigate this, apply for re-licensing as early as possible, and consider requesting interim “pre-clearance” for urgent shipments (though this is rare under the new rules). Also, allocate internal resources to compile the more detailed end-use documentation now, because incomplete applications will be returned and further delay approvals.

Where to Go From Here: Three Decision-Path Recommendations

1. Immediate Compliance Audit & License Re-Application — If your company already holds an export license for AI technologies under the 2023 list, initiate a full compliance audit to map all existing items against the new eight subcategories. Use the 60-day window to submit a re-application before July 14. Prioritize high-volume export items (e.g., training accelerators or cloud-based AI services) to avoid a licensing gap. Consider engaging a China-based trade compliance consultant—our partners at China Gateway 360 recommend starting this process within one week.
2. Revise Contractual Clauses for AI Technology Transfer — Update your technology transfer and cloud service agreements to include explicit end-use declarations, clauses requiring foreign partners to provide ultimate destination information, and provisions for mandatory compliance audits every 9 months. This is especially critical for joint ventures and WFOEs that share AI models between China and overseas headquarters. Ensure that all agreements specify that any change in end use triggers a new licensing obligation, and include a mechanism for the Chinese entity to suspend transfers pending MOFCOM approval.
3. Evaluate Strategic Alternatives: Model Segmentation & Localization — For companies whose business models rely heavily on cross-border AI model sharing, consider segmenting model components: keep the core training and inference infrastructure in China, and allow only de-identified, non-controlled metadata to exit. Alternatively, stand up a localized AI stack within China using domestic hardware (e.g., Huawei Ascend, Cambricon) to reduce dependence on controlled exports. This approach may also qualify for certain exemptions if the technology is used solely within China and not intended for re-export. Discuss with legal counsel whether “localized AI stacks” can be structured to avoid triggering the control list entirely.

Each path has trade-offs. The compliance audit path is the fastest but requires significant internal resources. The contractual revision path is strategic for long-term relationships but may slow down deal-making. The localization path offers the greatest reduction in regulatory risk but involves upfront capital expenditure. We recommend a hybrid approach: begin auditing now, adjust contracts in parallel, and launch a pilot localization project for your highest-risk AI products.

Looking Ahead: What the Clarification Signals

The MOFCOM clarification is part of a broader trend: China is tightening AI export controls to align with its national security framework and to match the controls imposed by the United States and the European Union. Foreign executives should anticipate further expansions of the list as generative AI models grow in capability and as China updates its dual-use list annually. The 2025 clarification also hints that “quantum AI” and “neuromorphic computing” may be added in the next revision, as referenced in an annex. Companies should build compliance capacity now—invest in trade compliance software, train staff on the new definitions, and establish a direct line to MOFCOM’s export control hotline.

Finally, note that the clarification empowers local customs authorities to inspect AI-related shipments and request end-use certificates at ports. Several provinces (including Shanghai, Guangdong, and Jiangsu) have already issued local implementation notices. Foreign companies with manufacturing or R&D facilities in those regions should liaise with local MOFCOM branches to understand any additional scrutiny.