In 2019, all the fabs combined made $X chips per year, and projected a growth of $delta_X for the next year.
By the end of 2020, $X + $delta_X is significantly less than the actual demand.
There are only three possible expectations:
1. The fabs made less than $X + $delta_X chips
2. There has been an explosive growth in sales of existing electronic devices exceeding $delta_X that either surprised the OEMs or that OEMs failed to communicate to the chip fabs by booking extra capacity in advance.
3. All fabs are not equally equipped to handle fab demand. If an OEM changes to a different fab vendor:
3.a. Some other OEM was denied capacity at the new vendor
3.b. The capacity that was opened up at the old vendor isn't usable by the OEM that got dropped by the new vendor
Has anyone done a deep dive and done the math to see how much capacity shortage was due to #1 and #2?
Apple famously moved a huge amount of their chips from Intel to TSMC when they made the m1 Mac. Intel has traditionally only used their fabs to make their own chips - so the OEMs that got dropped at TSMC couldn't just take their business to Intel to rebalance the available fab capacity.
Does the amount of m1 macs made/sold help explain the difference that #1 and #2 can not explain?
Is there another equally interesting story in the chip fab business that would explain this shortage?
M1 is insignificant, or at most just a very small piece of the puzzle. There are currently huge shortages of parts that are made on 10 year old silicon processes. Those share almost nothing in terms of supply chain or manufacturing capabilities with state-of-the-art CPUs like M1 or what Intel makes. Except maybe at the very lowest levels, like wafer and reagent suppliers.
My reading of this whole thing is that we just made a brittle system that fell apart at the first slightly larger global disturbance (the pandemic).
It's lean manufacturing and just-in-time inventory. When you cancel orders or people can't work for days/weeks/months due to a pandemic, it creates a massive backlog. We've created incredibly fragile supply chain.
Everybody drained down their inventory heading into Covid assuming demand was going to crash. Simply replenishing that was going to take time even under the best of circumstances.
And, to be fair, there likely has been a bit of uptick in semiconductor demand--lots of people needed webcams, new monitors, new mice, new graphics cards, etc.
Also, fabs are normally running pretty close to capacity in general. An idle fab is a waste of money. This leads to the famous "feast or famine" in the semiconductor industry--either your fab is running full out and people are scrambling to build more, or you have oversupply and your fabs are completely idle. There is basically no middle ground.
I suspect there is also some overbuying going on that is going to unwind on December 31 when a whole bunch of CFOs realize that there is a big tax bill for a bunch of overbought inventory.
It's lots of things all coming together, but, at the end of the day, it's simply the fact that nobody holds inventory, and nobody will get fired for not holding it.
> Is there another equally interesting story in the chip fab business that would explain this shortage?
Automakers cancelled their orders at the last minute, so the fabs sold the slots to other companies.
When demands for cars actually increased in 2020 automakers tried to get their slots back but were told to wait for 2021 for new ones to open. Not sure what the fabs made instead of auto-parts.
M1 is 5nm, those are all factored in 5nm capacity planning. i.e It is part of $delta_X.
>Is there another equally interesting story in the chip fab business that would explain this shortage?
1. I have always thought the toilet roll incident in pandemic would give every one a lesson in commodity and supply chain. The Chip shortage is no different. One can also look Lumber, Steel or any other commodity. But basically instead of having a negative impact on demand, COVID actually make huge increase in demand. PC industry saw ~15% increase than previous projection. One only need to look how a 5% misprojection of Smartphone shipment in ~2018 lead to an avalanche of NAND and DRAM pricing.
2. Old mature node ( sub 28nm node ) has always been in short / tight supply for nearly a whole decade. You just dont hear about it in mainstream media because they are not headline worthy and doesn't drive clicks. Generally speaking the older you are and the more you know about an industry, the more you realise how news are always very late to the cycle.
3. None of these factored in trading. If any market is a simple game of supply and demand one could always balance things out. But you have people betting / trading these commodities. Which are not actually "demand" as in usage but investment. Considering the premium / profit one could gain from chip trading could easily be multiple if not in some cases order of magnitude. This also has a negative feedback loop, as wholesale or channels see difficulty in getting hold of something they will want to stock pile more of it. Something similar to what is described as Bullwhip effect.
4. One often overlooked sector is Crypto. It basically absorb any sort of capacity as long as the Crypto dollar is profitable. Right now this doesn't seems to add much to the constrain. But at one point Crypto players could buy up whole patch of wafer capacity at a much higher premium.
5. Supply will always be much slower to react, the higher capital investment, the more risk there is in over supply. There used to be an old Intel joke about The most expensive Fab isn't leading edge Fabs, but will always be an empty Fab. This is just as true today. It is the same with DRAM and NAND, that is why you see vendors reluctant to build additional capacity until absolutely necessary. ( Whether it is a cartel or not depends on your perspective, so I will leave that part out. But it is absolutely common for any commodity industry to meet and discuss these issues. )
Well, all THREE of those scenarios (and sub-scenarios) are routinely happening. And add in some other reasons as well.
2019 was predicted to be a "low cycle point" going back to market projections in 2016. So fabs were already dialing back production. This then was amplified by COVID. So production volumes were dialed down and the boom was not expected because it didn't follow decades established cycles of supply and demand.
Fabs are absolutely NOT equal even when they target the same baseline features/performance.
Usually fabs from the same company can be equivalent but there are plenty of situations when the particular fab for a 2nd or 3rd tier foundry will not be up to the same standards so you'll often want one particular site to do your parts or to be avoided.
TSMC is head-and-shoulders above Samsung, who is head-and-shoulders above UMC or GF. Fab services are absolutely NOT interchangeable beyond a fairly low common denominator.
Fab services are NOT fungible in the general case between fab companies!!
If you are forced accept a 2nd tier fab product, you WILL have to change (often radically) what your entire marketing and business plan will be. You will NOT be able to promise the same things that getting a TSMC product would allow. You will NOT be able to charge the same product prices that getting a TSMC product would allow. And that assumes you have a top-flight IC design team.
This is why TSMC is the gold standard. They are entirely focused and have the best of the best of people.
Being one of TSMC's vendors for anything is a very harsh but profitable duty; but that's because they are the best and demand the best (Intel sort still was like this but nothing like TSMC anymore). I've worked with other fab companies, and the difference in focus and commitment among TSMC employees is very noticeable in the details and attitudes. They are scary good.
For at least a decade, TSMC and other top-flight fabs have indeed been routinely turning away or delaying fab runs to fabless clients who's order volumes or dollar cost was below a planning threshold. This is NOT new. So in many ways there's been shortages for TSMC level foundry services for more than a decade. It's only gotten much worse recently.
TSMC has been conservative with expanding production capacity because they are Taiwanese and that's a cultural norm.
Only "knowing someone" and/or becoming personally involved e.g. flying to Taiwan or Korea and talking directly with management, as a C-level executive of your company, will break that lock but only so much - you can and will always be outbid and then bumped. TSMC literally turns away most fabless companies that request their services from the get-go. You fit their schedule because that's where the power is due to demand.
Apple is easily TSMC's largest customer with Nvidia being 2nd or 3rd. But Apple has been using TSMC for iOS/iPad processors for a long time now so they are already established and the M1 et al. merely upped an already well-established and reliable working relationship to high dollar value. They have a relationship and system of working together.
Apple has special priority for everything because of their already high volume. When you enter lobbies at TSMC and all other Taiwan semiconductor fabs, there's always a sign welcoming some Apple employee for a visit that week. Apple is super-engaged with its vendors like this. So nothing is left to chance for them. Lots of other fabless do the relationship "at remote arm's length" which puts them at a disadvantage!
Many IDM semiconductor companies like Texas Instruments, and others, are also using TSMC to manufacture many of their "standard digital product" ICs as well because the cost to use TSMC is cheaper than running 20-year-old+ fabs of their own.
These are multi-year, high value contracts for much of TSMC's baseline volume (i.e. there is portfolio value to fulfilling them). So these are also impacting supply when demand bumps up - long term contracts trump short-term or unscheduled work.
To spin up a new fab costs between US$ 30B-100B and has a 3-5 year timeline on a good day. So adding capacity is both a major capital commitment and never instantaneous. TSMC does have the money (hence the new Phoenix AZ fab already under construction) but it will not have any impact for a few years yet.
This cost/delay is then multiplied by the tacit technical knowledge required which can NOT be easily spun up or duplicated. TSMC is adding new factories like crazy but even they are limited by reality.
This latter tacit knowledge is why there's so much Epic Fail with Chinese semiconductor firms now (MOST are outright fraudulent scams, e.g. Evergrande and EVERY semiconductor company associated with them) and why no one can just walk in and compete with TSMC.
Even Intel has not been able to jump into being a foundry. They UTTERLY FAILED when they tried become a foundry about 10 years ago - they've been an IDM too long to really have the corporate culture to make being a foundry work. As it is now, Intel is actually outsourcing a lot of their "foundry wins" to the likes of TSMC and Samsung. It's NOT THEM. Same problem to this day.
I work in this market. TSMC is one of my customers. We also use TSMC fab services for custom chips we put into our products (a recent run we did required we stuff multiple chips from several groups into the same multi-chip run to get a spot - we then saw them up ourselves and package them).
I've worked in semiconductor since the 1980s. I was involved in the creation of the Fabless-Foundry model back in the early 1980s (MOSIS, USC-ISI, etc.) I've been inside pretty much every fab of every company around the world over the last decades as vendor, employee or customer.
SMIC is entirely "trialing edge" process technology-wise. If you want 7 nm or lower for competitive advantage, you don't go to SMIC. Also you can NOT trust your IP to remain secure if you use them.
My mental model is this:
In 2019, all the fabs combined made $X chips per year, and projected a growth of $delta_X for the next year.
By the end of 2020, $X + $delta_X is significantly less than the actual demand.
There are only three possible expectations:
1. The fabs made less than $X + $delta_X chips
2. There has been an explosive growth in sales of existing electronic devices exceeding $delta_X that either surprised the OEMs or that OEMs failed to communicate to the chip fabs by booking extra capacity in advance.
3. All fabs are not equally equipped to handle fab demand. If an OEM changes to a different fab vendor:
3.a. Some other OEM was denied capacity at the new vendor
3.b. The capacity that was opened up at the old vendor isn't usable by the OEM that got dropped by the new vendor
Has anyone done a deep dive and done the math to see how much capacity shortage was due to #1 and #2?
Apple famously moved a huge amount of their chips from Intel to TSMC when they made the m1 Mac. Intel has traditionally only used their fabs to make their own chips - so the OEMs that got dropped at TSMC couldn't just take their business to Intel to rebalance the available fab capacity.
Does the amount of m1 macs made/sold help explain the difference that #1 and #2 can not explain?
Is there another equally interesting story in the chip fab business that would explain this shortage?