U.S. ordered TSMC to halt shipments of advanced AI processors to China: Report

TSMC may not have voluntarily cut shipments of advanced AI processors to Chinese entities, as reported last week. Rather, it was ordered to do so by the U.S. government, according to a South China Morning Post report published earlier today.

The U.S. Department of Commerce issued a notice to TSMC imposing export controls on advanced AI processors and GPUs made on technologies that are produced on a 7nm-class or more advanced fabrication process. The regulation does not cover chips for automotive and consumer devices.

This action was prompted by a recent discovery: a TSMC-manufactured chiplet was found in an AI processor developed by Huawei, which suggested a breach of existing export regulations. Indeed, it was promptly discovered that Huawei used a proxy to plant an order at TSMC.

Last month, TSMC notified the U.S. Commerce Department after research company TechInsights disassembled a Huawei AI server and identified HiSilicon’s Ascend 910 processor with a TSMC-made chiplet in it. Since Huawei is on a U.S. trade restriction list, companies are required to obtain a U.S. export license to supply any product containing 25% of American technology to it. It just so happens that almost any advanced product is designed or made using well over 25% of technologies developed in America.

With the latest U.S. Commerce Department directive, more companies are subject to scrutiny to prevent potential rerouting of advanced processors to Huawei. While companies like Alibaba and Baidu are to some degree competitors to Huawei, the U.S. government considers them allies due to their origin.

When the news about TSMC cutting the supply of advanced AI processors to Chinese entities broke a few days ago, we were a bit surprised. Typically foundries do not voluntarily halt shipments, as it hurts their revenue and margins. But apparently, TSMC’s action was not exactly voluntary.

Both TSMC and the U.S. Commerce Department have largely declined to provide additional comments, with TSMC stating…