TSMC Deep Dive Podcast
TSMC Deep Dive Podcast
On the Compounding Curiosity Podcast
I want to share a recent podcast I was on, and enjoyed being on - the Compounding Curiosity podcast.
I had a blast. I felt so excited that I must have been talking a hundred miles an hour! Kalani makes podcasting fun and productive with a thoughtful preparation process and excellent questions.
I wasn’t quite able to get to everything that I had written up for the podcast - just did not fit in the flow. So for this newsletter I want to share some of the outlines and notes I prepared for this podcast.
TSMC Origins and setting the scene
A fun way, at least for me as a listener, to start my deep dives is with your history and background for what we’re about to dive into. So what’s you’re history with TSMC, what got you interested, and why are you passionate about this topic?
Taiwan is obsessed with TSMC. And that has been the case for a very long time.
It’s hard to put it into perspective and context for outsiders
Morris Chang is like Elon Musk here
The Chinese nickname for the company is literally translated as the “Silicon Sacred Mountain” or the “Silicon Shield”
Recently I was in Kaohsiung and TSMC recently announced that they are going to put a new fab there - not sure what generation but not a leading one.
And it was all everyone can talk about in the cafes and the conferences. I’m at Starbucks and there’s little old ladies talking about it.
At a startup festival I was attending, the keynote speaker by a member of the municipality - I was later told he was the Mayor of Kaohsiung, after Han Guoyu got voted out - he came in and I didn’t understand what he said except for that him saying TSMC TSMC TSMC repeatedly. Like a chant.
TSMC's new foundry consumed a third of Taiwan's newly generated power;
There is a generation of Taiwanese who are now rich because they bought TSMC stock.
its 40,000 plus employees also contributed more than 1% of total newborns in Taiwan in 2019.
Go to Hsinchu, you will find a plethora of children and families
As for history of me and TSMC, I don’t speak to TSMC people and I almost never hear from them. On the other hand I’ve gotten many emails from ASML employees.
And Dutch people are so nice.
For listeners to get a bit of background and understanding, could you explain the chip development process from start to finish? So from maybe the original idea all the way through to ending up in our iPhones, could you detail that?
You have four stages: Design, fabrication, packaging, and testing
Semiconductor design starts with a specification.
That specification is turned into a high level design representation of the chip - referred to as an RTL - using a language like Verilog.
Modern SOCs are architected by integrating hundreds of pre-designed and pre-verified hardware blocks or "cores". Without it, I don't think modern semiconductor design is possible at a commercial level.
These cores are provided by an ecosystem of industry players - Original Equipment Manufacturers, 3rd party Semiconductor design houses, or TSMC.
TSMC has a massive repo of IP property that it can license to you
Then that high level representation is converted into a netlist, a bunch of gates. These are done with third party EDA tools like Cadence Genus Synthesis Solution.
The gates in the netlist are then placed and physically routed using EDA software.
The design is then transmitted to a foundry for fabrication, usually in a file format called GDSII.
After fabrication, the wafer is cut, packaged, and tested before being shipped into the rest of the supply chain
These steps are called "assembly" (I have also seen it referred to as “packaging”) and "test". Together make up what we call the "back end".
Someone has to cut the wafer into chips, encase them within ceramic or plastic so to make it easier to handle, and test them to ensure quality.
The work is mundane and repetitive but at the same time requires dexterity and close attention.
Furthermore, productivity is relatively low. 50-70% of the time, the guy is waiting for another machine to finish what it is doing.
So TSMC is the first foundry to provide 7-nanometre and 5-nanometre production capabilities, and to put that into context, a human blood cell is about 2,500 nanometres, and in a single drop of blood, there’s around ~5 million red blood cells. So it’s hard to even fathom how they’re doing this kind of work. Could you maybe set the scene for how technologically difficult and impressive the work TSMC is doing? Especially because sometimes its easy to get a little lost in the numbers and nanometres and whatnot.
Wafer manufacturing is hard. They do it in mask layers, a mask is used to reproduce a pattern onto a substrate, so you would put a whole chip pattern or part of a whole chip pattern into one mask layer
Each mask layer requires steps with 15-25 per each mask layer. So there can be up to 600 steps, for a simple 180 nanometer process chip. That’s a chip from like 20 years ago.
The today ones - N14, N16, and N10 go up to 60, 70, 80 mask layers.
This is super super hard. On a process with 400 steps, even if each step has a yield of 98%, the total yield of your wafer is just 0.03%, which means 3 good dies out of 10,000
TSMC is historically known for good yields. I was looking at Canaan financial filings. Company that makes BTC miners
Their yields at 16 nanometers were 96% in 2017, 96% in 2018 and 95% in the six months ended June 30, 2019
TSMC 16nm has some 60 wafer layers, which means about 800-1200 steps. To achieve this, you need better than 99.99% yield on each of the 800 steps
Again for a little context, but could you give a little colour as to what the fabs look like and how they’re built? Because this is where the magic happens, so knowing the setup and infrastructure that allows this to happen is extremely interesting and important.
A fab is a factory just like any other. Fabs are built in phases. You got phase 1, phase 2, phase 3, usually when someone quotes big cost numbers, they are summing up all the phases together. You generally want to get these done within a year.
A single fab cleanroom is a massive undertaking. Cleanroom size is directly correlated to expense. A fab with a cleanroom sized 8,000 square meters needs:
Enough concrete for 25 km of road, 120 kilometers of ultra clean pipes, 250 kilometers of electric wire, 4,500 kilometers of steel bar
It uses 24 swimming pools of ultra pure water each hour (though much of that gets recycled)
It uses enough compressed air to fill 3,000 balloons each minute
It runs 24/7 all the time - night and day.
And guess what? A TSMC Giga-Fab have cleanrooms 13 times larger across all their various phases
Fab 18 in Tainan where N5 and N3 are being done is 20 times bigger, 160,000 square meters. (Note: Thanks to Ryan, Asianometry Discord member, for his correction that this is over multiple buildings - not 1 building).
Why is it so hard to have an American semi conductor industry? Is the the talent hurdle too high?
It is hard to catch up to the leading edge once you stop going for it -
And because the new leader is moving as fast as they are moving. If you want to catch up with the leader, you have to run faster than them because it is very easy and more financially profitable to just say “this is too hard and we don’t need a faster chip”
When it comes to Semiconductor manufacturing you are not only racing on a technical basis, you are doing it on an economic basis too
Showing stuff on in a lab is nothing compared to actually doing it for customers at a profit
Advanced manufacturing is very hard and it needs scale. The more scale that TSMC gets, the better it gets at manufacturing - which means their customers get faster chips at a cheaper price
Hyper competitive manufacturing is a bad business. Why do you think all these tech companies are going to D2C subscriptions? It’s an easier life.
Taiwan lives and breathes semiconductors. The USA is going to be very good at a lot of things, and that is why America is so successful.
But Taiwan is the Jamaica of semiconductors, the USA of NBA talent. It is like what Malcolm Gladwell said about running talent in Jamaica - there is a system set up around it.
On average, Taiwan had one large foundry per 1,000 square kilometers, making it the leading region in foundry density globally.
And then for the economics behind building the fabs, so TSMC goes to great lengths to accommodate Apples product rollout, which is fair enough when they represent about a quarter of their revenue, so what’s the economics and numbers look like behind building a fab that’s likely to be dedicated to a sole customer?
Semiconductor fabs are all about depreciation. 70% of the cost of a wafer is depreciation. Soon as you finish the fab and it starts working, the capital equipment inside starts to depreciate.
Depreciation is an accounting metric that attempts to replicate the ongoing expense of maintenance
CapEx is expense that kills foundries because that’s cash out the door that won’t get paid back until later.
What the financial analyst likes in me is to see something what is happening with Global Foundries, when the company is showing very high cash generation because all the capEx had already been invested previously
Once the equipment is fully depreciated, then suddenly the company looks super profitable
The reality of course is that the depreciation exists for a reason
You got to keep investing in the machines so that they produce at a very high rate
TSMC poured $9 billion to start a new fab for Apple.
It is a huge financial risk for the foundry to start a new fab.
But I think it’s actually a financial risk that TSMC is more than willing to do. It’s actually a slam dunk.
How many Apples are there?
You have a partner who is willing to work with you at the leading edge. Like how AMD and GlobalFoundries had been
Now TSMC is starting to do the same for Intel and it’s interesting to see how that relationship is going to turn out.
Forgive me for asking a dumb question, but could you explain how TSMC adds value compared to ASML who provides the machines to fab? There’s the description that ASML sells the pots - TSMC sells the cooking, how accurate is that?
I don’t think it’s a dumb question. I get a lot of comments saying like “Well ASML is a European company” and they are the only ones making EUV. Why are we selling them to TSMC? Why aren’t we making a European fab?
The answer is that lithography is by far the most expensive part of the process, but it is not the only one.
The fallacy is presuming the most expensive equipment is the most proportional important one.
There’s other equipment involved in the process. And it needs to be sourced from all over. For instance, the wafers come from Japan, the masks from Japan, the packaging services are in Taiwan, and so on. Applied Materials is in the United States.
Competitors and the landscape
Could you talk a little about TSMC’s supply chain and their relationship with ASML? Are you worried at all about TSMC’s reliance and dependance on ASML?
ASML and TSMC work closely together. They have worked together ever since ASML was part of Philips.
You can do N7 without EUV, but you won’t do it profitably. It’s a science project. And you can’t do N5 and so on at all.
But what is the terror case here? That ASML decides that they will only sell to Intel and not sell to TSMC?
I reckon that’s not going to happen unless there’s a geopolitical incident.
I think there is a possibility that Canon or Nikon - mostly Nikon I think - break through to EUV and whatever is beyond that
TSMC recently announced a $44 billion CAPEX plan, competitors have announced similar spending plans, are we heading for a glut?
Recently, Intel has been doing a lot of committed spending - $20 billion for an Ohio site with possible $100 billion commitment
It’s hard to not say that we won’t be in a glut at some point. Especially as the economy starts to slow down and interest rates rise.
But that does not mean that you guys are going to see cheaper GPUs. I keep getting comments about that. Talk to Nvidia.
But I think TSMC sees the situation as a net benefit for them, because they tend to want market share over profitability
They’ll set the market on fire and they know they can take the heat because again - their capital base is solid
When the crash happens then they’ll buy nonprofitable capacity off the losers
Is it naive to think that Intel can catch up just by spending about the same amount as TSMC? How else can Intel catch up to TSMC?
Intel and TSMC had never really competed together before, mostly because they seemed to have played in different circles.
Intel and TSMC have been friendly. Andy Grove was one of Morris Chang’s first customers and Intel patiently worked with TSMC to make it happen
Samsung has always been TSMC’s biggest competitor. It is the company that actually is willing to steal talents - and has - and can splash out money like no other company
Samsung has ambitions too. They spread out and about
Taiwan used to have a substantial memory and display LCD panel business and Samsung took all that away from them. Samsung Electronics has been around since the 1990s and they will continue to compete
Intel has been working on its foundry, but it’s hard to make sense of what to do about it until we start seeing some revenue come into the books.
If they break it out into a real segment and we start seeing it take share from TSMC and GlobalFoundries and Samsung then we can look back and say, “wow they were not talking out of their butts”
And it will make me look exceptionally stupid.
So Never doubt a $75 billion company
But right now, I tick off a bunch of reasons why Intel’s Custom Foundry work won’t work:
The EDA tool integrations are not ready yet and Intel has generally done what it has done by itself.
Intel has squeezed a lot out of its design-manufacturing integration. Are they the best GPU designer? The best CPU designer? Are they the best manufacturer? They will need to be both.
Let’s say TSMC enables AWS to make Graviton3s and so on that are SO good that Intel finally starts losing substantial share in the server business.
Foundry is a service business and Intel has never done that sort of thing before. Taiwan has a Japan-like flavor in the sense that they really do customer service. Broad generalization
And of course, what to make about the rumors that Intel is going to use TSMC’s N3 process? How can Intel sell foundry services to their customers while using another foundry for the leading edge?
But I think it’s just as likely that Intel Foundry ends up something like GlobalFoundries,
Which right now runs at an 84% utilization rate, as compared to TSMC and UMC’s 95%, in 2020. And is losing a lot of money despite hundreds of millions of dollars of government subsidies
The majority of their wafer volume comes out of Singapore due to their Chartered acquisition in 2009
Each year, they are selling wafers at a cheaper price, because each year the competition is coming up behind them
Culture-wise, I think TSMC is a very aggressive company. You work long hours. You work day and night.
TSMC gets away with it because they pay pretty good and Taiwanese labor laws are pretty weak
But also because people work like hell for it because you are literally part of the Silicon Sacred Mountain. The Silicon Shield.
Intel doesn’t have that status in the US. Which is why when a guy on Glassdoor writes that “TSMC will have to change to an eight-hour work day five days a week or run three days on, four days off, three days off, four days on”. I don’t think that’s on the table.
Are you worried at all that TSMC has all their eggs in one basket? The China threat has been talked about into the ground, but even another event like the 1999 Jiji earthquake could totally halt production. I know you’ve mentioned how the clean rooms are elevated, but I don’t think it’s crazy to say that TSMC is probably THE most important company in the global supply chain, so how are they mitigating these risks?
The company has put a lot of effort into earthquake amelioration.
They work with the government to predict when an earthquake will happen so that the fabs get enough time to shut things down
And they’re diversified across Taiwan too. You got multiple fabs in Taichung, Tainan (where my family is from), Kaohsiung, Miaoli, Baoshan.
Most of the worst earthquakes have been in Nantou, Yilan, Hualien
The fact of the matter is that Taiwan won’t let TSMC move its best people, equipment, or facilities off-island.
I don’t think a fab in the US really ameliorates for the risk. Taiwan’s fabs are doing over ten million wafers a year. You think two or three more fabs in Texas, Ohio and Arizona doing even 50,000 wafers - which would be big - are going to fill that need? That they can ramp up as fast in time?
I read a comment on HackerNews that said that the United States should maintain as much capacity in the US as Taiwan, just in case.
That’s something like 1.1 million 12-inch wafers a month to make up. Intel alone does 880K 8-inch wafers a month.
12 inch wafers are roughly twice the size of 8 inch wafers. Maybe more.
So TSMC is over twice as big as Intel in terms of scale. I’ve never heard of a physical manufacturing company tripling its capacity at that size -
And according to that comment, the capacity is just going to sit there doing nothing waiting. It’s depreciating and many of this equipment doesn’t make financial sense unless it’s being used 100% of the time
Everyone is running on a financial tightrope here. You cannot subsidize all of it
So I think TW is fine for something up to a 8.0. Nothing can be done about something like a 10.0 earthquake.
Is Apple being their biggest customer a risk? Whilst building your own fab is a massive investment, and technically very difficult, Apple is probably one of the few companies with the resources to do so. Anything to be worried about?
I’d say it’s the biggest risk. Apple anchors TSMC’s customer base. Without Apple I do not think TSMC gets to be the giant that it is today.
So yes there is a failure mode where Apple one day announces they will second source M1 chips from Samsung or Intel and TSMC stock will take a massive hit
Losing any one of these core chip customers - AMD, Nvidia, Qualcomm, MediaTek - will hit the stock hard but Apple most of all
But at the same time, I think Apple and TSMC work very closely together - and it is hard to step off that treadmill when you are on it
Apple and TSMC get to work together and define the PDK - the process development kit
This is called Design Technology Co-Optimization and it’s really really tight. Like literally, TSMC engineers will tell Apple engineers, “You can’t arrange these logic gates this way because of this FinFet misalignment” and the Apple engineers will say, “Okay we will draw the gates elsewhere”.
Something bad has to happen and it has to be really bad
I said in a previous video that TSMC is always going to try to get a new process ready for Apple for a new iPhone. I think that’s less the case now because TSMC’s N3 is late, 2.5 years rather than 2 years, and it is going to miss the iPhone and even maybe the iPad.
But TSMC is still Apple’s exclusive fab. Because the two are working so closely together that it is hard to separate now
I don’t think Apple can build its own semiconductor fabs.
They are already moving towards more of that system integrator Boeing model, where they pick and choose suppliers and make money from services and users of their ecosystem.
I think TSMC has been searching for another client that is going to sit alongside Apple at the leading edge - that used to be Huawei. Now it is probably going to be Intel
I don’t think Apple will leave TSMC because of Intel, after all they were competing head on with Huawei, but maybe AMD will lower their business with TSMC because of it.
AMD and Samsung working together for an Exynos chip I think is pretty interesting. And it makes me wonder which came first.
TSMC is probably going to play AMD and Intel off each other for a little bit.
If Intel thinks they can just shift those N3 wafers back in-house, maybe TSMC can give the whole fab capacity to AMD? Who knows. But that’s good capacity and someone will take it.
Misc
Is Moore's Law really dead?
Moore’s Law was never a Law of course. It’s a pace-setter - like those electric rabbits in the dog race - for the industry to continue pushing technological progress forward. But it is getting harder.
It is interesting to see that starting in 2000 and 2001, the return on investment for capital build up in the overall semiconductor industry started to decline
Total assets stopped growing and it sorta peaked for a few decades. Installed capacity stopped growing and peaked
I see two reasons for that:
First the dot com bubble popped and the economy fell into a big recession
Second, that is when 193 nm ARF lasers started to hitting the market, and these DUV machines represented a generational change that shook out a lot of customers
Yet the semi space is so rich and so money heavy that you can afford to throw smart people and cash at the problem until it goes away.
I am about to say a bunch of very abstract things off the top of my head so forgive me for being wrong in the details. But for the first thirty or so years of the industry - the pathway to a faster chip was simple
You needed more transistors - more gates switching on and off.
So the industry was like, yeah just squeeze the gates together more and more. Conceptually simple.
But that stopped being the case at around the 32 nanometer node.
The scientists realized that they should make better gates rather than denser gates. So they introduced FinFet at N22 and N16
The laser lithography situation got hard when 193 topped out with 193i and EUV hit delays. So they overlaid 193i over and over again, 2, 3, 4, 5, 6 times until EUV came about.
The semiconductor industry drives forward not necessarily because some dude 45 years ago said so, but because people want faster, better looking games. Because software programmers write crappy software and they just assume the chips will get faster and can run it. Because everyone is working on the cloud. So on.
So long as the market is going to reward an electronics product for being better, the semiconductor industry will push forward
You’ve mentioned how TSMC is very open, except when it comes to their fabs. Compare this to ASML who are extremely open with their research findings, any theories as to why TSMC is so secretive when it comes to their fabs?
ASML is very open about their conceptual research, but they are never going to tell you how they actually put it together.
That’s where the trade secret is
Also, the optics stuff is just way too hard for any human to comprehend.
TSMC is a conceptually simple company. They make and sell wafers.
Their trade secret is their fab and it used to be that you can read white papers about how that stuff works.
How the automated systems works
What they focused on
How the recipes are kinda there
But that’s gone now. Samsung and SMIC have taught them a very nasty lesson about what it means to be too open
Basically you are not going to find anything about their operations or work styles anything past 2010.
The best thing I know of is this article in the CommonWealth, and it discusses how the fabs are more automated than ever before. Where they use large data sets to help optimize recipes and operations
So it kind of makes sense to me then that we don’t hear about how TSMC works and optimizes - because nobody really knows
Q4 earnings
So a while ago, I had been looking at TSMC and seeing it be so highly valued felt strange to me. TSMC right now has a $650 billion market cap. This is the 9th or 8th most valuable company in the world. Why?
The company is simple to value
The company is simple to predict revenues - let’s talk more on that later
People perceive it to have scale, which it does, and that it has a monopoly, which I would argue that it doesn’t.
A quirk of the market? Everyone in Taiwan buys TSMC stock. And they never sell. If you think bitcoin is full of HODL, you haven’t seen Taiwanese aunties.
This really smart analyst named Sebastian Hou - that’s the name of my sister’s cat - and he talked about this on the latest call
Sebastian, my sister’s cat There is something really interesting happening with TSMC’s capEx and its future revenues.
Take the capex for a year, multiply it by about 3X, and you can broadly predict revenues for the next year
This makes sense since it generally takes a fab a year to get on track
This trend has been going on for almost a decade.
With current exchange rates, pulling from cash flow statements
2020, TSMC spent $18.3 billion on capital expenses
One year later 2021, it generated $56.8 billion revenues, 3x
2021 TSMC spent $30 billion on capital expenditures
I’ve seen other more conservative estimates saying $72 billion
So in 2022, is TSMC making $90 billion in revenue?
At 38% net margin from 2021 that’s $34 billion in net profit
That’s Exxon Mobil in 2003 and Gazprom status
That’s how much Google/Alphabet made in 2019
2022 TSMC says it will spend $40-44 billion. What if they hit the high case for that CapEx?
So in 2023, is TSMC making $130 billion revenue?
At 38% net profit margin from 2021 that’s $53 billion in profit
That’s the same as Apple’s net profit in 2020 and more than MSFT in 2020
So this leading indicator is flashing utterly insane growth.
TSMC: We expect our long-term revenue to be between 15% and 20% CAGR over the next several years in U.S. dollar terms
TSMC 10-year revenue growth CAGR is 14%.
That’s the kind of growth that got them to the largest fab in the industry, yet their long-term revenue guide is now actually a call that their revenue will accelerate on a larger base.
This is not a software company. This is a hardware company, making actual things
The weird thing is that nobody is yet clear on where this industry-wide growth is coming from. More than just Intel onboarding.
“Yes it is from AI and data centers”, but what trend?
AI? AI has been around since 2013. Why now?
Work from home? That’s been around since 2020. Why now?
Shortages? Inflation? Everyone says that it is about the 22nm and lower chips, and that’s not a huge revenue driver.
Concluding Thoughts
What have we not talked about that's consequential about TSMC’s future, in your opinion?
The relationship between TSMC and the Taiwan government is sort of interesting. The government owns a significant share of TSMC and sometimes gets involved in its operations
I reckon that’s why TSMC is building a joint plant with Sony.
I reckon that’s why TSMC is building a plant in Kaohsiung
It’s confirmed that the Tsai administration got involved with helping TSMC decide to put a plant there
But I also know that the government bends over for TSMC too.
Their R&D centers source new technologies
Tax benefits for the employees
Tax benefits for the companies