The semiconductor industry is highly volatile.
The booms follow the busts with frightening speed. These are not normal times. But even in
normal times, conditions are unpredictable. Semiconductor shipments can soar
70% one year and crash 40% the next. In this video, I want to think about the
semiconductor business cycle. Right now, we are working through a big chip shortage. I wonder how
soon before that turns into a glut. With everyone racing to build or expand their fabs, maybe sooner
than we might think. But I do want to note, don’t expect me make any this year or that predictions
in this video. Just a bunch of scattered thoughts. Let us start by looking at the
dynamics of demand and supply. When things are planned out well, then capacity
can be forecasted, shortages avoided, and everything will be fine. The chairman of Foxconn
said this recently about the recent shortages: > Since most of the customers we serve
are large customers, they all have proper precautionary planning ... Therefore, the impact
on these large customers is there, but limited Of course, the hard part is the actual
forecasting. End users do not directly buy microchips in large quantities. Instead,
the chips are purchased by vendors who put them into consumer goods like PCs and
smartphones that the users then buy. These chips don’t last long on the proverbial
shelves. There is always a new product coming out this or next season that threatens to make
your current iteration obsolete. Missing an annual launch cycle like what happened with AMD
in the mid 2000s can have dire consequences for the whole company. For more information about
that, you can watch my video about AMD and TSMC. So foundry customers are always wanting
to get them made and shipped yesterday. Their concern about securing stock in uncertain
times can lead them to double book orders, which means exactly what it sounds like. TSMC Chairman Mark Liu cites double booking as
one of the major reasons why the shortage is not as bad as it looks. He blames this double
booking on a number of very legitimate reasons: the pandemic and the recent China-America
trade tensions. The latter of which is the bigger concern. No end in sight for that one so
we should expect some double booking to continue. Demand is also heavily impacted by the greater
macro-economy. It makes sense. Electronics are consumer items and consumers buy fewer iPhones
and laptops when times are tough. For instance, three downturns in the semiconductor business
cycles can be attributed to oil shocks in 1975, 1982, and 1991. Another from the
Asian Financial Crisis of 1998. Everyone might be howling about
semiconductor shortages right now, but we forget that in March of 2020
the stock market was down 25% and experts were saying that the
world economy was falling to bits. Back then, companies were doing all they
can to cut their expenses and hunker down. But a flurry of stimulus packages and money
printing meant that the economy bounced back faster than anyone ever anticipated in Q3 2020.
I don't think anyone predicted that would happen. So that is macro. Everyone talks about
macro but here is something interesting. A study in 2005 found that what is more
significant in the triggering of semiconductor down cycles is not a decrease in demand. Rather,
it has more to do with surges in semiconductor inventory levels and fab capacity. In other
words, over-supply rather than under-demand. This makes sense to me. It is not like
foundries and semiconductor companies have never heard of the business cycle.
But rather, the nature and economics of the industry force them to make like
rabbits until the whole thing collapses. First, let us start with the fab. A
fab costs a lot of money to build. We all know about Moore's Law. Not an
actual natural law but rather an industry observation that the number of transistors on
an integrated circuit doubles every two years. It is famous and everyone likes to talk about it. But not as well known is Moore's Second Law.
Also known as Rock's Law, named after one of Intel's early investors. It states that the cost
of semiconductor tools doubles every four years. Such tools represent 75% of the total
cost of the fab, driving them upwards. In 2000, a leading edge fab cost a billion
dollars, twice as much as it was in 1995. In 2005, $3 billion. Now in 2020, TSMC's leading
edge N5 fab in Tainan cost roughly $17 billion. Its next generation N3 process fab (also in
Tainan) is estimated to cost $19.5 billion. Like with Moore's Law itself, Rock's Law
is an observation and not a natural law. As much research goes into slowing Rock's
Law as it does into accelerating Moore's. And there are a few techniques and
financial methods that companies have explored in reducing those costs.
And they have their own trade offs. Regardless, foundries still
have to put up billions of dollars before they can even start to compete. Most foundries would like their new billion
dollar fabs to make back their billion dollar investments. But with the way the chip market is,
those fabs start losing value almost right away. It is like what they say about
driving a new car off of the lot. Much of these fab construction costs
are fixed. But the marginal cost of making another chip is basically
nil. So you want to make and sell as many chips as possible to
amortize those fixed costs. What happens is that supply hits the market,
exceeds demand, and prices collapse. Everyone starts making less profit. Market supply
should pull back. But the foundries are incentivized to sell every chip that they
can possibly make. If they can generate even a bit of gross profit that can contribute
to their future survival costs, they sell. Furthermore, the industry's titans - TSMC and
Samsung - value market share over profitability. They refuse to mothball capacity lest that
business go to any potential competitors. And because they have massive economies
of scale, they can cause the market price to plunge to essentially their survival price -
far lower than everyone else's in the industry. The result is a fierce downcycle
that burns out the lower performers. Companies go bankrupt and get bought
out. Supply re-balances with demand and the cycle resets itself once more.
Waiting for the next chip shortage. As they say in the Romance of the Three
Kingdoms: "The empire, long divided, must unite; long united, must divide." I want to talk a bit about subsidies. There
has been a lot of chatter in the news about governments putting down cash to build fabs.
China is doing it, with their $20 billion semiconductor plan. President Joe Biden
included $50 billion in his stimulus plan. And the EU is putting together 145 billion
euro for a 2-nm cutting-edge fab of their own. I can't address all that I want to say here,
but I want to get to some essential points. In the US, there are two significant
pieces of US legislation in the works: The CHIPS act and the American Foundries
Act. They would establish a R&D fund and a 20-40% tax credit for the acquisition
of semiconductor production equipment. I think the tax credit's effects
would be more short-term, impacting three to five years or one
fab generation. On the other hand, I like the R&D fund, which I think will have
longer term lasting effects. Especially when it comes to spinning off public-private
partnerships like the one that created EUV. So I have mixed feelings about the whole venture.
I understand the value of national sovereignty, government involvement and investment in certain
high-priority things. I recognize that I have written a lot about TSMC and Taiwan, and thus
might harbor biases. By now, you probably have already decided for yourself whether or
not I am some sort of shill or otherwise. But I am also a big believer in commercial
things having to pay for themselves. Like with every other government action, subsidies
create distortions. That is the whole point of having them. The subsidy’s intention is to
bring more manufacturing to the USA. It looks like they will get their wish. 40% subsidies can
do that. But unless the subsidies go on forever (and they might! The government's
spent money on dumber stuff before), the subsidized fabs should have to stand
up on their own two feet afterwards. We will need to go back to Congress
to get those credits renewed, right? I bet then it gets a lot more scrutiny, especially
once the politically powerful automakers get their chips back, the F-150s are back on the
lots again, and we are on to the next crisis. I am doing this video because. Today, we are
in the midst of a large build-out of fabs across the world. Fabs are so hot right now. TSMC is working on its Arizona fab. That has
been going for a year now. And there have been whispers that this small 20,000 wafer fab is
being upscaled in size from the original plan. Samsung is in negotiations on a major
investment buildout of their Austin fab. They are going to spend tens of
billions of dollars on that as well. Even the smaller, second-tier foundries are
building out or upgrading their fab capacity. There has been a flurry of press
releases from all these companies, none of whom with a process node under
10nm, about new sites and projects. Global Foundries is spending $1.4 billion
in expansion and moving up their IPO. Maybe they need the capital. Maybe
their Middle East investors think they should strike while the iron
is hot and cash out their shares. And SMIC is getting money from the Shenzhen
government to build a $2.35 billion project. That wafer fab is expected to do 40,000
wafers a month once it is done in 2022. Just a small sample of the recent news
going on in the semiconductor world. And my immediate thoughts is that
it is a whole lot of capacity. For instance, once TSMC completes its Arizona
fab, the new facility will have about 25,000 to 100,000 wafer starts a month, depending on the
final size they decide. Along with everyone else, that is hundreds of thousands of wafers
coming to the market over the next 1-2 years. It would be irrational to say that that
won't have any effect on the market once that capacity is all at peak output. Did we
really create tens of millions of chips worth of demand over the past year? Now, people are
starting to go back to school and the office. Are PC Masterrace types or crypto miners buying
that many new graphics cards? I doubt it. The American government is not going to buy
all of those chips. The Department of Defense buys 1.9 billion semiconductors a year. If the
TSMC Arizona fab does 50,000 wafers a month, that alone is a theoretical 3.5
billion dies a month. One month. So I am watching this build-out with some
interest and curiosity. The chip shortage has been in the Wall Street Journal and the Verge
for a really long time. Semis are super hot right now. I wonder how long this boom will
last, and when the bust will inevitably happen.
