AI Card Bao NAS.AICARDBAO.COM

Industry News

TSMC suspends taking orders for 16nm and below from Chinese companies and requires U.S. permission before supplying

TSMC has informed its customers that orders for advanced manufacturing processes using 16nm and below technologies in Chinese companies have been suspended.

The basis for this is the newly tightened regulations regarding U.S. semiconductor exports. Under these new rules, it is not possible to supply goods unless prior export permission or exemption has been obtained from the United States.

The change lies in the word “prior to the process”. Previously, TSMC could decide on its own whether a customer’s chip design met the exemption requirements; if it deemed it did, the company would accept the order. Now, customers must first apply for approval before placing their orders.

The list of companies that have been approved, as currently announced by the United States, includes Apple, AMD, and NVIDIA. None of these companies are Chinese.

The consequences are quite obvious. Domestic chip design companies can only use mature manufacturing processes such as 28nm and 40nm. The original text also mentions that high-bandwidth memory cannot be utilized either, yet it is an essential component for AI data centers.

The 16nm technology was developed quite late. It became commercially available around 2015, which means it is several generations behind today’s 3nm and 2nm technologies. Mobile phone processors and AI acceleration chips have long since moved on to more advanced processes, but a large number of automotive-grade chips, networking chips, and mid-range processors are still using 16nm and 12nm technologies. Therefore, not only companies that produce AI chips will be affected by this situation.

TSMC's response was simply one sentence: "We will comply with the new regulations."

Those affected are the domestic design companies that have been relying on TSMC’s manufacturing services. They either have to obtain licenses from the United States or move their production operations back to their own domestic facilities, which means they will have to start from scratch in terms of achieving satisfactory yield rates and increasing their production capacity.

Within the same week, there were two additional developments that aligned with these trends: The U.S. House of Representatives made public a list of 879,000 AI chips that had been approved for use by Chinese companies in the past six months, criticizing the Department of Commerce for its overly lenient approval process; ByteDance was also required to obtain additional permits regarding its transactions involving the rental of computing resources through overseas data centers. In short, all three approaches – purchasing ready-made products, renting computing power, and developing custom chips in-house – are facing increasing restrictions.

Related