How Many New Fabs Are Under Construction? Ask Christian Dieseldorff

Everyone must have heard that there is a semiconductor shortage by now. The story that we've all heard is two-fold. First, automotive seems to be a special case. Their sales cratered at the start of lockdown, so they canceled all their semiconductor orders. Then when their sales recovered much faster than expected, they tried to reinstate them. But there were two problems. One is that chips take a long time to build and so you can't just turn orders on and off like a tap, so they might well have been in trouble anyway. But by the time they tried to place orders, there was no capacity.

The second part of the story is that during lockdown we all needed more electronics, everything from video game consoles for play, new laptops and monitors for work from home, upgraded routers, and so on. And while NVIDIA says it's not true, the received wisdom is that bitcoin miners hoovered up all the GPUs.

As Marc Andreesen points out in his recent new essay Technology Saves the World:

We are coming out of COVID years early, with many livelihoods and businesses preserved, compared to what we had any right to expect. And overwhelming credit goes to our spectacular technology industry.

There is more to technology than electronics, of course, most notably vaccine development technology. But there is no doubt that almost anything that you can think of—DoorDash or Zoom, say—required a boost in the capacity of their infrastructure to handle the step-function in their business. And that typically meant more electronic systems, and especially more semiconductors.

The Semiconductor Industry

To me, the most interesting question is to what extent this boost in demand will turn out to be limited to the period we were all locked down, and to what extent it is a secular increase in the size of the market for technology in general and electronics in particular.

It is actually harder to tell than you would think. When I was at VLSI Technology 30 years ago, the semiconductor industry was much more cyclical than it seems to be today. Once business went "on allocation" with customers not getting their full orders in a timely manner, they would start playing games, ordering twice what they needed to hope they got their true need, or ordering ahead so that even when their shipments were late they were when they were really required. The truth would only become apparent when the mismatch between supply and demand eased and all the unnecessary orders would be canceled. You could have the odd phenomenon of the market improving and the book-to-bill ratio going negative due to cancellations (actually, "negative" means less than one, but everyone said negative).

The biggest leading indicator to what the industry truly believes is whether they are investing in building new fabs. Even at the best of times, it is hard to make the decision to build a new fab. In the days of a more cyclical industry, you would have to make that decision during the downturn, because by the time the upturn came along and was obvious, it would be too late. I can't find the quote, but the CEO of one semiconductor company said that these decisions are like a slow version of Russian roulette: you put the gun to your head and pull the trigger, and two years later you find out if went bankrupt.

Capital Investment in the 1980s

I happened to come across this Dataquest table on Twitter recently. This is the capital investment in semiconductor fabs back in the 1980s. The numbers are almost comically small. In 1990, the total capital investment for the entire semiconductor industry was forecast to be $14B. Also, note that Asia (outside of Japan) barely existed (I'm guessing that most of the ROW number was actually Asia-Pac). But no Korea, no Taiwan, no China, no Singapore. Japan still bigger than the US.

By comparison, from Statista, here's how it looks today, with Korea, Taiwan, and China being about 75% of capital investment (technically this is equipment spending, but the difference is a rounding error):

Some capital expenditure (capex) goes into outfitting existing fabs. The standard way to build a fab is to build the concrete for a much larger fab and equip it more slowly. There are a few reasons for this. One is simply that when a single EUV machine costs $150M then you can buy a lot of concrete for that. The second is the lead time issue. Pouring concrete takes longer than moving equipment into an existing shell. Thirdly, there is vibration. You really do not want to start pile-driving the foundations for a new fab right next to an existing running fab.

The World Fab Forecast

But the real leading indicator is new fabs. And the go-to guy for anything to do with new fabs is Christian Dieseldorff who maintains SEMI's fab database and writes its reports. I Googled him to try and find a photo for this post, and it turns out that he is a professional salsa instructor. The page I landed on is all about different types of rumba and salsa. But apparently, he has a day job, too: "He works as an industry analyst for the computer chip industry at an international trade organization in Silicon Valley."

If you really want to dig into the details, you have to pay for access to the fab database and associated reports. But every so often when a new version of the World Fab Forecast is produced, SEMI also puts out the headline numbers. They did this a couple of weeks ago:

Semiconductor manufacturers worldwide will have started construction on 19 new high-volume fabs by the end of this year and break ground on another 10 in 2022 to meet accelerating demand for chips across a wide range of markets including communications, computing, healthcare, online services, and automotive.
...

Equipment spending for these 29 fabs is expected to surpass $140 billion over the next few years.

I happened to see a YouTube video about semiconductor in Europe. The biggest fab in Europe is GF's Dresden Fab 1. They are expanding it. Nearby, Infineon has a fab, they are expanding it. Also nearby, automotive tier-1 Bosch, who has never built a fab before as far as I know, is building a $1B fab. Another conclusion from these numbers is that Dresden is the silicon manufacturing hub of Europe.

And Europe is not even that significant on the global scale. SEMI again:

China and Taiwan will lead the way in the new fab construction starts with eight each, followed by the Americas with six, Europe/Mideast with three, Japan and Korea with two each. Fabs that produce 300mm wafers will account for most of the new facilities – 15 – in 2021 and again in 2022, when seven fabs will begin construction. The remaining seven fabs planned over the two-year period will be 100mm, 150mm, and 200mm facilities. The 29 fabs could produce as many as 2.6 million wafers per month (in 200mm equivalents).

I prefer to think in 300mm equivalents since that is what the leading-edge companies always report, so that is 1.3M 300mm wafers per month, or about 15M wafers per year.

To put all this in perspective, here is a quote I reported in my post Dr C.C. Wei's Keynote at TSMC Symposium. He said:

In 2020, total managed capacity reached 12M 12" (300mm) equivalent wafers. In 2021, TSMC's capex increased "ambitiously" to $30B.

Another interesting question is what these fabs are going to be used for. Are they all memory? Christian has the answer to that:

Of the 29 fabs starting construction in 2021 and 2022, 15 are foundry facilities with capacities ranging from 15,000 to 110,000 wafers per month (300mm equivalents). The memory sector will begin construction on four fabs over the two-year span. Those facilities will boast higher capacities ranging from 50,000 to 200,000 wafers per month (300mm equivalents).

I formatted that as a quote, but actually the original was in 200mm equivalents. Like someone's going to build a 200mm memory fab! So I converted it all to 300mm.

It is worth emphasizing (sorry, automotive industry!) that this capacity will not come online immediately. It takes roughly two years from starting construction to starting production. But Christian's database has a much better handle on this than me:

Of the semiconductor makers beginning construction of new fabs this year, many won’t start installing equipment until 2023 since it takes up to two years after ground is broken to reach that phase, though some could begin equipping as soon as the first half of next year.

But wait, there's more, as the old ads used to say. There is probably even more in the pipeline:

While the World Fab Forecast shows 10 high-volume fabs starting construction next year, that number is likely to climb as chipmakers announce new facilities. 
...
In addition to the 29 fabs expected to begin construction in 2021 and 2022, the World Fab Forecast report tracks eight low-probability projects that could start construction over the same period.

For example, Pat Gelsinger, Intel's newish CEO has already hinted that they want to build another fab in Europe...and I assume that's not in Christian's database...yet.

Summary

I said above that the semiconductor industry used to be more cyclical than today. But memory was and is still cyclical. We will have to see about logic and other technologies. The driver for the cyclicality is that supply and demand are never that well matched in semiconductor, and that is not something that can simply be fixed by price like in Economics 101. So if there is undercapacity, prices are high and life is good. If there is overcapacity, then cash-flow is everything. It is better to lose money on a wafer with some revenue, than lose money on what is effectively the same wafer with no revenue by not manufacturing it at all. You can't put unused wafer starts in inventory for when times improve.

There is a truly massive amount of capacity being built (or forecast). We will have to wait and see whether the industry has hit the sweet spot.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.