TSMC has announced it will build a cutting-edge semiconductor foundry in Arizona, with plans to bring the fab online at the 5nm node. There are questions about just how big the factory will be and how central it is to the company’s plans, but choosing to site a foundry in the US is a significant step in the economic relationship between the United States and the Taiwanese foundry company.
The details of the plan, however, are a little surprising. According to TechCrunch, the new fab will target 5nm production, won’t be ready until late 2023 / early 2024 at the earliest, and is apparently targeting just 20,000 wafer starts per month. That puts it at less than half the size of the GlobalFoundries New York factory, which can handle about 50,000 wafer starts per month.
It’s a little surprising to hear that TSMC is targeting the 5nm node for this foundry, though that could change over time. With foundries typically taking 3-4 years to build, companies typically do not target them for current nodes. If TSMC intended to make the US facility a leading-edge node, it would already be targeting 3nm for the factory.
It’s possible, however, that TSMC will either change its node target, or that we’re seeing the impact of fewer and fewer companies making the transition to leading-edge nodes. The stated reason why GlobalFoundries got out of the game several years ago is that the company didn’t believe there would be enough customers available at 5nm in order to make continuing to challenge the leading-edge worth it. As the number of companies eager to advance to the leading edge shrinks, it means (logically) there are going to be more companies left behind on older nodes. TSMC may also hope to entice companies down to nodes they aren’t currently using if it can continue to improve those processes and lower the cost of adoption, which could make 5nm a long-term node for the manufacturer.
Overall, this announcement comes at a complex time for TSMC. Both mainland China and the United States are critically important customers for the company, and the increasing tension between them can’t have made things easy for the foundry right now. The surprising thing about the announcement is that there’s no mention of Intel at all. This could reflect the fact that Intel is currently focused on ramping up production at Fab 42 and other facilities, or that the company didn’t find it advantageous to consider a foundry investment right now, but it’s a surprising outcome. Given President Trump’s heavy emphasis on “America first” and the fact that both companies were in talks with the Administration, I and a lot of other people unconsciously expected an Intel announcement, I think.
This isn’t TSMC’s first US foundry, but it’ll be much closer to leading-edge than the company’s other facility. The other fab, in Camas, Washington, builds primarily flash memory on the 350nm – 160nm node. Even if 5nm isn’t cutting-edge by the time this new facility opens, it’ll be much closer than anything we build for TSMC in the United States currently.
There are some other strange angles to this deal. As Junko Yoshida points out, TSMC’s total capacity is 13 million wafer starts per year, or over one million per month. No 20,000-wafer facility built on 5nm is going to make a dramatic difference there. Also, there’s no mention of where in Arizona this fab is actually going to be built — just an announcement that TSMC is building one. She thinks TSMC’s willingness to build a fab in the US may be part of why the DoD recently shifted its position on enforcing further sanctions against Huawei. This announcement may have precious little to do with actually building a foundry.
The facility is expected to employ roughly 1,600 people and is scheduled to break ground in 2021.
- Intel, TSMC Reportedly in Talks to Build New US Foundries
- The US May Make It Illegal for TSMC to Sell Parts to Huawei Without a License
- TSMC Starts Development on 2nm Process Node, but What Technologies Will It Use?