TSMC spends big but builds small in deal with Uncle Sam

At the end of last week, the world’s largest semiconductor foundry said it would build a leading-edge fab on US soil. This will not be TSMC's first operation in the US: one already exists in Washington state. That 20-year-old plant also was created at a time of economic shocks, although a much more localised one than that caused by today’s pandemic. 

WaferTech, as the Washington fab is known, was the result of a massive demand spike, fed partly by the rise of the public internet that made chip companies fear they would not be able to get manufacturing capacity. Companies were suddenly finding that capacity they had booked had in fact been sold on to a higher bidder. Banks were encouraging them to get ahead of the problem, so some of them formed groups that would share the burden of building expensive fabs that could keep the supplies going. The plant in Camas was funded initially by Altera, Analog Devices, Integrated Silicon Solutions and TSMC. 

Pretty soon, though, circumstances changed. A demand spike turned into a supply glut and chip prices crashed. Despite interruptions from the odd mini-boom, prices kept on falling for the most part for the next 20 years and only recently began to look as though they had turned the corner - just in time for Covid-19 to sweep around the world. Foundries like TSMC wound up acquiring sole ownership of the few consortium fabs that got the green light before the first internet hangover. Now a set of circumstances is taking hold, although TSMC’s fab is not quite the major move it might at first seem. 

TSMC partitions its existing fabs into three groups: mini, mega and giga. Based on older 200mm wafer lines, you can look at Camas as being a mini. At 20kwafer/mo, running today’s more cost-effective 300mm wafers, the Arizona plant falls into the mega category. This is some way short of the 100kwafer/mo of which the gigafabs are capable. Even at a cost of $12bn, the Arizona plant will clearly be dwarfed by TSMC’s 5nm-capable Fab 18 in Tainan, which has 160,000 square metres of cleanroom space. That in itself is a good third bigger than the original floorpan of the not-exactly-small Fab 15, also on the island of Taiwan, which began construction a decade ago.

Arizona itself may at first look an odd location for a US fab, but that part makes sense. It’s long been home to factories operated by Intel and the former Motorola chipmaking operation that is now part of former Philips subsidiary NXP Semiconductors. That gives TSMC a pool from which to recruit, especially as Intel pressed pause on its own Fab 42 project - originally scheduled to go onstream this year - after running into problems with the 10nm process technology. 

If you include Fab 42, Intel’s largest grouping of fabs outside the R&D and early-production cluster in Hillsboro, Oregon – Intel’s operational home after moving most of it out of Silicon Valley – is in the Chandler area of Arizona. The other main wafer-producing sites are in Kiryat Gal, Israel and Leixlip, Ireland. 

An advanced TSMC fab in Arizona looks like a win for the Trump administration in that it’s “bringing jobs home”. Yet the cost structure, with a high to eye-watering capital-cost to output ratio, seems unlikely to appeal to TSMC’s biggest customers. Apple could consume the Arizona fab’s capacity several times over but appears to be happy enough with having chips made in Taiwan – especially as they will, with the exception of the Mac Pros, be assembled into consumer products in neighbouring China. 

To that end, it looks more a decision made of political expediency rather than one based on an analysis of supply-chain resilience or economic development. It is way less ambitious than the plan the European Commission followed for a while to create an 'Airbus of chips' in the first half of the past decade. This would have created a gigafab able to use a larger size of wafer than those used today to bring costs down further. 

At the time, the big suppliers were keen on the shift to 450mm. They had the cost structure to support such a transition, especially as it would put even more distance between them and smaller, less cash-rich competitors. However, fears among the equipment suppliers that they would be forced to cut prices to an even smaller customer base than the one they have now, coupled with a lack of confidence in growth, meant the 450mm transition simply fizzled out. 

For the 'Airbus of chips', even the issues with the 450mm transition took second place. The local chipmakers had little interest in the fab itself: they were happy enough sending designs to the Far East for manufacture as it kept their capital costs down. In the event, a lack of political will allowed the project to collapse. 

Even in the current environment, it seems almost inconceivable that a second plan could get off the ground in Europe. Supplies from the fabs have proved to be pretty reliable even under lockdown conditions. They are, after all, highly automated and amenable to social-distancing rules.

There are risks to having a lot of advanced fabs in one region and it does make the supply chain look fragile, especially when you consider that they are in earthquake zones. Even if you distribute fabs around the world, that does not help most customers. Unless those fabs run identical recipes, there is almost no way to split production among them. Also, the number of wafers you need to run to have split production make sense economically means only the highest-volume customers would be interested. The evidence so far is that they are not interested.

The US military will no doubt be happier to have an advanced fab in the country. Pretty much everyone else will still be looking for the cheapest option.