Discover more from jonstokes.com
Why a Chinese invasion of Taiwan would be a catastrophe for China and the world
If they do it, it won't be for the microchips
For more content like this, please consider subscribing and following @doxanotes on Twitter.
Worries are increasing that China may invade Taiwan, amid a combination of Chinese military saber-rattling and US military warnings that the PRC’s timeline for “reunification” has moved up and may happen any time within the next few years.
Most coverage of a possible Chinese invasion of Taiwan focuses on the area as a potential flashpoint for a US vs. China war, but I haven’t found any detailed coverage of what such an invasion would mean in practical terms beyond generic talk of nuclear escalation risks and widespread damage to the global economy.
In an effort to think through the impact of such an event on the technology sector, along with potential ripple effects elsewhere in the global supply chain, I talked to David Kanter, Executive Director of ML Commons. David’s an old friend of mine who’s spent many years covering semiconductors as an analyst and journalist, and he helped me update my slightly out-of-date semi knowledge and better understand how the current market is structured.
In short, the market for advanced microchips fits together like this:
The equipment and tooling for making leading-edge chip fabrication plants comes from the US and Europe
The two leading-edge fab makers, who have the proprietary know-how to turn that equipment and tooling into fabs that turn out chips more advanced than anyone else’s, are in Taiwan’s TSMC and South Korea’s Samsung. (There actually sort of two-and-a-half leading-edge fab makers, with Intel being the half. They’ve stumbled but will recover and join the other two, soon.)
The customers for the outputs of those leading-edge fabs are global, including in the US and Europe.
Porting a chip design from one company’s fab process to another is costly and painful. It’s not like you send an email with some files to a different address and a few days later identical chips come off a new production line in a different factory. No, customers are tied to specific companies’ fabs, and suddenly losing access to that fab means suddenly no more chips until you complete a lengthy, costly re-targeting process.
So fab equipment goes from the west into Taiwan, where TSMC puts all the parts together into fabs that can turn raw materials into advanced chips, and the chips then go back out to the west and rest of the world.
Thus the semiconductor market in 2021 is a fully baked cake. You can’t just swap some ingredients out in response to a one-off military invasion, and then keep trucking along. No, we’ll have to bake a whole new cake if TSMC goes bye-bye. And that will be painful for everyone, everywhere.
To unpack all this in a bit more detail:
The world’s largest, most valuable tech companies are dependent either directly or indirectly on the steady output of TSMC’s fabs. If those fabs went offline or became unavailable in the west because they were controlled by PRC, it would immediately devastate the global economy. An unknowable number of large companies just wouldn’t be able to refill their inventories for an indeterminately long time.
TSMC’s status as a leading-edge semiconductor powerhouse depends critically on inputs — tooling, equipment, software — from the US and Europe. So if China did seize TSMC and could keep it running (a big “if,” and probably unlikely), it would be stuck and unable to advance the state-of-the-art. It will not realize its tech leadership dreams through military adventures.
More likely is that China would not even be able to take over TSMC’s fabs and talent via military means and keep it all running, status quo. They’d take over disabled fabs that had been damaged either by the invasion or by the Taiwanese (in an effort to protect customer IP), and wouldn’t be able to locate and/or keep critical talent.
The world economy would likely go into crisis if TSMC suddenly stopped working, but in the long-term, the US is already in the process of decoupling from dependence on Asia and is on-shoring more of the semiconductor value chain. So the US would eventually recover in a few years, and companies outside China would eventually re-target their designs for other fabs.
China is a lot further behind the US in its national semiconductor sovereignty efforts and would suffer more in the medium- and long term.
Given that an invasion of Taiwan would not magically give China leadership in the global semiconductor market, and would in fact set back its timeline for being an advanced semiconductor superpower by probably decades, then if such an invasion does happen then it’s not because they want TSMC.
TSMC’s lone $12B fab in Arizona is a nice gesture, but it is not relevant to any part of this wargaming. It will only make a tiny slice of TSMC’s global output when it comes online in a few years and will be trailing-edge at that point. It would generally be pretty useless without the rest of TSMC in existence.
The decapitation strike
TSMC operates a foundry business that takes chip designs from its clients — like NVIDA’s GPU designs, or the designs for the ARM chips that power Apple devices — and manufactures them on its advanced processes. All told, TSMC has over half the global foundry market (55 percent), and even more of the foundry market for the most advanced microchips.
So if the PRC invades Taiwan and seizes TSMC, it will be something like a decapitation strike on the global semiconductor market — a big chunk of the world’s most advanced, powerful chips made on the world’s most advanced manufacturing process will instantly stop being available during whatever invasion and transition period. This alone will upend markets and global supply chains while it plays out.
Assuming China can get TSMC is up and running again after an invasion — this is a really big assumption — all of TSMC’s clients in the west will have a painful choice to make:
Drop TSMC immediately, so that there are just no new products — no new iPhones, no new NVIDIA GPUs, no new who-knows-what other kinds of advanced products that rely on TSMC chips either in the product itself or in the production process — for two or three years until everyone can find replacement foundry capacity at a comparable level of sophistication somewhere else and then complete the costly work of porting their chips over to that foundry’s process.
Try to keep existing product lines at TSMC while making plans to migrate new product lines to other foundries as those catch up with TSMC in capability and capacity.
There are three options that are actually not on the table, and later in this piece, I’ll explain why for people who don’t know the industry and who may mistakenly think either of these is viable:
Keep on doing business-as-usual with TSMC, but under de facto, Huawei-style PRC ownership and control.
Shift existing TSMC order volume to a rival like Samsung, which shares process leadership with TSMC, or to trailing-edge fabs owned by Global Foundries, Intel, or the like.
Just move advanced chip manufacturing to the US and/or Europe. After all, TSMC is working on a new fab in Arizona, and Intel is building new fabs in the US. So imagine a kind of Project Warp Speed, but for advanced chip fabrication.
There may be some other options I haven’t thought of, but these are the possibilities that come to mind.
It’ll go through these in the rest of this article, including the last three that are not likely to happen. My main goals here are to do a little bit of scenario planning, and a lot of educating and debunking.
Dropping TSMC like a hot potato
My first option above — everyone outside China drops TSMC immediately — would immediately halt the flow of iPhones, GPUs, parts for fighter jets and supercomputers, and many other complex electronics that make up the global economy.
It’s really difficult to say how widespread the carnage would be because it’s hard to say where TSMC’s chips are sprinkled throughout the global supply chain. But my guess is it would be very grim.
As more products, from cars to home appliances, have begun to rely on more advanced logic for features like voice recognition, image processing, complex user interfaces, and the like, our markets have become more reliant on high-end semiconductors. I really don’t have a good sense of the number of consumer products that are using high-end chips from TSMC, much less the prevalence of these chips in military, manufacturing, shipping and logistics, and other critical points in the supply chain. But I wouldn’t be surprised at literally anything going out-of-stock for a few years if this happens.
As for why this worst-case scenario might happen, it’s not at all clear that China could just step in after a military invasion and send everyone at TSMC right back to work in their old jobs. Some of the fabs may well be damaged from conventional weapons or a cyberattack, employees could be dead or missing, and many key employees would have expatriated.
Furthermore, there’s a lot of critical IP at TSMC headquarters that either TSMC itself or western powers would probably destroy to keep China from getting their hands on it. The masks for things like iPhone CPUs or the proprietary TPUs that power Google’s AI efforts may not be usable on the day the PRC tries to turn the lights back on at TSMC.
(Update: I got some smart feedback to the effect that Applied Materials and other US-based suppliers to TSMC could close down the company quickly by just cutting them off from the ongoing flow of critical materials and other inputs. So the US effectively has a kill switch on TSMC, and even if China were to take it intact with all the critical personnel willing to keep working, the US could shut it down by ordering its domestic suppliers to drop it.)
All of this means it’s very possible, maybe even likely, that the day after an invasion China would have a non-functional mess on its hands, and there would just be no more TSMC production for a very long time, if ever.
Phased migration from TSMC
Assuming TSMC is still intact after a Chinese invasion and that China can get it all running again, then some consumer product makers like Apple might be willing and able to keep the iPhones and new Macbooks flowing from PRC-controlled fabs while looking for an alternate supplier.
Big tech mega-platforms, like Google and Amazon, that are using some TSMC silicon in their clouds or other efforts, might likewise be suck up the likely IP theft in order to continue to source existing products from “Chinese Taipei” while making plans to migrate the next versions of all their chips to other foundries.
However, companies that sell chips to the US government will run into problems. These companies will be barred by law and other forms of pressure from selling chips made in PRC-controlled fabs to the government. So a lot of NVIDIA GPUs that go into our government supercomputers and the Xilinx chips that go into our F-35’s would suddenly have to find other buyers the moment the PRC takes over Taiwan.
Some firms would survive this sudden loss of revenue while they solve their foundry problem, while others would not.
I should note that already something like the beginnings of a phased migration away from TSMC is probably already in the works, in light of all the military activity around Taiwan at the moment. Every entity with a dependence on TSMC is right now in meetings trying to figure how to decrease that dependence so that if anything happens to TSMC they’re able to survive it. I’m sure Samsung’s phone is ringing off the hook.
Chip fab basics
Now let’s take a look at the options for dealing with a Chinese-occupied Taiwan that are not viable. To understand why certain things are off the table, you have to know a bit about how the global semiconductor market works.
How advanced, high-performance, and power-efficient a microprocessor is depends mainly on the manufacturing process used to make it — i.e. how small a feature size that process is capable of producing.
Smaller feature sizes are better and more advanced, and whoever has the smallest feature size at any given moment is winning the technology race. (Right now, that’s TSMC at 5nm.)
Every transition to a smaller feature size is built on highly proprietary, secret know-how that the chipmaker learned from all the previous transitions.
There’s no way to magically jump the line by throwing money at the problem. The Gulf nations tried this by throwing $20 billion at Global Foundries, largely to no avail. You can’t just buy the know-how to make a leading-edge fab. You have to build that know-how before you can build the fab.
The secret know-how for fab building resides meaningfully in the parent company, not in the fab itself. So if you somehow had a super-advanced fab in one location — let’s just say Arizona, for instance — you couldn’t take all the employees and equipment at that fab and just clone it all a couple of times by throwing money at it if the parent company didn’t cooperate. It’s like in Breaking Bad, how the cartel couldn’t just buy a state-of-the-art meth lab and expect to compete with Walter White — they actually needed Walter White there to run it.
There are only three companies that have the know-how to make fabs at the leading edge: Samsung (South Korea), and TSMC (Taiwan), and Intel (US). (Update: Originally I said “only two companies…” but this isn’t fair enough to Intel. Intel has the know-how, they’re just still getting their execution sorted.)
TSMC can’t keep making new, advanced fabs without critical inputs from the US and Europe. So the US can basically stop TSMC’s ability to advance the state-of-the-art dead in its tracks, and will undoubtedly do so if PRC controls the company.
The first thing to know about microprocessors is that performance improvements are largely driven by innovation in our ability to make the fabrication plants (“fabs” for short) that make the chips. Yes, clever chip design matters and innovation does happen there, but innovation in the manufacturing process trumps everything else for most of the attributes we care about, like application performance and power efficiency.
The chart above is a recap of Intel’s improvements to its manufacturing process over a 15-year span. Each of the columns — 90 nm, 65 nm, etc. — is a feature size (i.e. a measurement of a critical part of a transistor), and each feature size shrink involves a massive amount of engineering effort that builds on all the knowledge gained from the previous feature size shrink.
So Intel used its proprietary learnings from the 90 nm to 65 nm transition to design the 65 nm to 45 nm transition, it used its learnings from that transition to design the next one, and so on for each transition down to the latest 7 nm transition. (Intel actually stalled out at 7 nm and is just now starting to move ahead again. More on that in the next section.)
This constant shrinking of feature sizes is the engine behind Moore’s Law, and staying on track to keep delivering regular feature shrinks is how you stay on the leading edge.
The other important thing to know is that when a new semiconductor manufacturing process is first brought online, there are a lot of imperfections and other issues in the finished product. The result is that relatively few of the chips that get produced actually work to spec — you have to throw all the bad ones out.
As you run the process for a while, you start to work the kinks out. Eventually, your yield rises, and enough of the chips you’re making are sellable that you can actually run the fab profitably. Until you really get the process dialed in, though, you’re taking a loss on every production run.
My point is that it takes at least two years and tens of billions to build a modern fab, and then it takes another year or so to work the kinks out and get the yields up. Lost fab capacity will not be replaced quickly.
Given the above realities of how the semi market works, you can see how the US and the world are extremely dependent on TSMC for advanced processors. But this dependency is a two-way street — TSMC is dependent on the US and Europe for critical inputs to its fab building efforts. Without the west, TSMC couldn’t advance the state-of-the-art and would become frozen in time.
The technical secret to getting features smaller than 7 nm is extreme ultraviolet (EUV) lithography, an advanced process that only Samsung and TSMC have mastered.
As for US semiconductor giant Intel, Kanter tells me that their “EUV is 15 years late and 10X over budget. It’s so horrifically complicated there’s no way to get it aside from ASML, and trying to recreate it would take you a decade.”
ASML is a Dutch fab equipment maker, and the US has successfully pressured it not to sell EUV equipment to China, effectively locking PRC’s entire semiconductor sector out of the feature size race. (They can sell older equipment to China, just not the good stuff.)
If PRC controlled TSMC, then suddenly ASML would find itself in a bind — either anger the Americans by continuing to supply TSMC with EUV tech or cut off one of its two main customers for bleeding-edge fab equipment.
Furthermore, there are many other fab equipment suppliers and semiconductor design tooling makers that are based in the US, and who are directly subject to US export controls. None of these would have the option of keeping TSMC as a customer.
Simply put, a PRC-controlled TSMC would be a dead man walking, in terms of its ability to participate in the march of Moore’s Law. It would be frozen at whatever its smallest feature size is at the time of a Chinese takeover, and it wouldn’t progress further until China could replace all the US-based fab equipment and toolchain inputs with domestic equivalents. That would take a very long time if it’s even possible.
China is actually trying and failing to break this dependency on the west and to gather all the inputs to the chip fabrication industry on its soil. It’s planning to throw over $100 billion at the problem over the next decade, and it’s way behind schedule. Again, this is just not the kind of problem you can solve by throwing money at it or with a national mobilization; if it were, China would already be winning at it.
To make its silicon sovereignty dreams a reality, China needs TSMC’s cooperation and know-how. It seems unlikely to secure that by military means, which is a good reason to think that if China does invade Taiwan, it’s not because they’re trying to take TSMC by force.
Why certain options are off the table
Now that we know more about the chip market, let’s revisit the options I said were off the table, and explain why they won’t work. This is repeats of stuff I said above, but now I’m just putting the pieces together to summarize.
Keep on doing business-as-usual with TSMC:
TSMC won’t be able to innovate without access to tooling and equipment from the west. So business-as-usual won’t be possible, and any company that stays with TSMC is planning to leave the leading edge and live in the past.
Many companies will be barred by the US from getting chips from a PRC-controlled fab.
TSMC probably won’t meaningfully exist after a military invasion of Taiwan, anyway.
Shift existing TSMC order volume to a rival like Samsung:
TSMC is 55% of the foundry market, so Samsung can’t instantly absorb all that, nor can the rest of the market.
You can’t immediately make the same chips on a different foundry’s process. Re-targeting a design for a new process is expensive and time-consuming.
Just move advanced chip manufacturing to the US and/or Europe:
It will take another few years for Intel to ramp production at 5nm and to catch up on future processes.
You can’t just throw money at the problem of having a leading-edge semi fab on your own soil, because if you could then the Gulf states and China would’ve already done it. Talent and institutional experience are critical and non-fungible assets in this race.
TSMC’s Arizona fab is a few years from being online, will be a small percentage of global 5nm production when it is online, and is not something anyone will use if the parent company is out of the picture. So it doesn’t matter for this discussion.
An alternate scenario
Speculating about the why’s and wherefore of a Chinese invasion of Taiwan is a bit above my pay grade, but I want to toss something out there that occurred to me as I read this great Steve Blank piece on the geopolitics of semiconductors. Specifically, this part stood out to me:
Alternatively, Beijing may seek to negotiate with or coerce Taipei (or both) in order to allow China sole access to TSMC and block chip exports to the United States, thereby securing China’s own supply while crippling American industry.
If China were looking to strangle TSMC’s western business, so that the chipmaker would have to turn to the mainland for customers, then I’d say it’s probably doing a really good job of exactly that with all the military activity around Taiwan.
The prospect of an invasion is spooking every one of TSMC’s non-Chinese customers, and in the coming weeks and months we’re going to start seeing more announcements like Tesla’s recent decision to backtrack on plans to use TSMC, and instead go with Samsung for some critical chips.
If China can scare Apple and the rest away from TSMC with a bunch of bomber flyovers and maritime shenanigans, then TSMC might go back to making chips for Huawei and other Chinese firms. I don’t know how plausible this is as a rationale for what we’re seeing in and around Taiwanese airspace, but it’s a less stressful thought than an all-out invasion and a crash of the global economy.