ECCN 4E091: New U.S. Controls on Artificial Intelligence Model Weights

The Bureau of Industry and Security (BIS) has introduced a new classification, ECCN 4E091, targeting the export of certain artificial intelligence model weights. These measures reflect a growing strategic priority for the United States: ensuring that the most advanced AI capabilities do not end up in environments where they could undermine national security or technological leadership.

Under this new rule, a license is now required for exporting the weights of AI models that meet two specific criteria. First, the model must have been trained using computing power exceeding 10²⁶ operations. Second, the training must have relied on high-performance chips already controlled by the United States, including items classified under ECCN 3A090.a or 4A090.a. Importantly, these controls apply even when the training process occurs entirely outside U.S. territory, using foreign infrastructure or resources. This extraterritorial scope marks a significant evolution in U.S. export control policy.

The regulatory framework distinguishes three levels of control, depending on the destination country. For allied nations and strategic partners, access remains possible but contingent upon defined conditions that aim to preserve responsible technology transfer. Countries subject to enhanced scrutiny face limitations related to computing power thresholds, though waivers may be granted following an individual assessment. For destinations considered most sensitive, access is tightly restricted and applications are generally met with a presumption of denial.

Compliance with this rule becomes mandatory in May 2025, with further phases planned for 2026 that will expand or refine the scope. These new requirements reflect the broader U.S. strategy to prevent the proliferation of AI systems capable of enabling offensive cyber operations, intelligence manipulation, or strategic military advances.

For companies working with high-performance AI models, the introduction of ECCN 4E091 demands a rigorous reassessment of their compliance procedures. Understanding training resources, documenting the hardware used, and classifying models accurately will become essential steps to avoid regulatory breaches. Beyond the immediate operational impact, these controls underline a deeper geopolitical message: the United States intends to shape the future of AI governance through export policy.