The Semiconductor Bust. Still Coming?
If you want to watch the video, it is below:
I just released this video a week ago but I already have something that I want to add to it. Previous early signs of bad semiconductor business cycles have come through rising inventories in downstream industries like computer makers and automobiles. We will need to reconcile this, however, against a new general move away from Just in Time after the pandemic.
I pulled some of this data and it is hard to see any particular patterns. The 2019 semiconductor downturn saw a tiny little move up.
The automobile industry has a much more pronounced chart.
Again, I think it is hard to reconcile this against the general trend of companies moving away from just-in-time and embracing having more inventory.
A year ago, I posted a video about business cycles in the semiconductor industry. I titled it the “Coming Semiconductor Bust” and people seemed to have enjoyed it.
And by "enjoyed it", I mean that they screamed at me in the comments that I had no idea what I was talking about. Which is par for the course here.
Now that a year has progressed, I thought some of the topics might be worth revisiting. I have been thinking about a few things recently.
So I want to run some numbers, talk some recent news items, and make a few predictions about what might happen in the coming years.
Now, Asianometry is not a semiconductor industry analysis channel. There's no better way to look stupid than to try to predict the future. So I am admittedly out of my element here. Just wanted to do this follow-up.
Memory Versus Logic
In his interview with Ben Thompson, Intel CEO Pat Gelsinger spoke his thoughts with regards to a coming semiconductor shortage:
> When was the last time we had a logic surplus, not a memory surplus? ... The last memory surplus was about three and a half years ago. The last logic surplus was over a decade ago.
Pat made an interesting point. In my first video, I didn't delve into the differing natures of these two markets. Ben probably should have dug into that a little more than he did.
Broadly speaking, the industry has these two very large categories of product: Memory and Logic.
Memory chips store information. The two major types of memory chips are DRAM and NAND. These are more of a commodity. Each generation of memory has a very short life cycle and suppliers compete ferociously with one another - often on price.
Logic on the other hand involves things like microcontrollers, AI accelerators, CPUs - mobile, desktop, otherwise - and GPUs. They process information to handle a task.
A logic chip requires more customization and handles more complexity. This specialization implies that the product experiences less of a commodity market.
Pat's point is that memory might frequently suffer massive drops in sales, and they do. But since logic chips are so fundamentally different and are not commodities, their business cycles are far more resilient.
So let’s check in on this. Pat mentions the last memory glut was in 2019. That year had been described as a difficult year for the semiconductor industry.
Just FYI. 2019 was three years ago? Time moves so fast, man.
Anyway, it was reported then that Global semiconductor sales declined by 12% from 2018. Growth went negative, the first time it had happened since 2015.
But this headline number was largely driven by a 30% revenue decline in memory - representing a quarter of the market. All of the major memory makers - Samsung, SK Hynix, and Micron - saw revenue declines in the 30% range. Definitely a memory bust.
The rest of the industry however saw much more moderate declines. Sales of all non-memory products declined by just 1.7% in 2019.
So Gelsinger might be onto something here. I mean, a -1.7% growth rate is still negative growth but far from a "glut" or a "bust". So one might think of logic as being more chained to secular macro-trends than individual industry dynamics. Thus making that corner of the industry less cyclical than we might assume on the surface.
The worldwide semiconductor fab buildout has continued on. Based on the announcements we have heard, by 2025 we can expect to see new online logic fab capacity from Intel, TSMC, Samsung, UMC, Texas Instruments, Tower, STMicroelectronics, Toshiba, Infineon, and maybe SMIC - depending on what's going on in Shanghai.
In total, I eyeball about 33 logic projects - fab expansions or construction starts - announced to be coming online within the next 2-3 years. This excludes packaging and R&D center announcements.
By the way, I want to thank friend of the Channel Redfire for helping to collate and present this information. He keeps a tracker that's as good as anything I have ever seen. Go check it out.
On the memory side, we can expect to see new DRAM and NAND fab capacity coming from Micron, SK Hynix, Winbond, Nanya, and Kioxia. Maybe Samsung too depending on how they split between the two.
Again within the same 2-3 year time frame, I eyeball about 10 memory projects - expansions and new fab starts.
Announcements rarely translate into actual shipments, wafers have different sizes and I have to estimate a few capacity numbers based on the projected capital expenditures rather than actual reported values. So this is a super dirty, disgusting back of the envelope calculation.
But if we are to take all of this at face value then by 2025 we can expect to see 1.2 million new logic wafers and at least 513,000 new memory wafers hitting the market. Three quarters of that logic number comes from Intel, TSMC, and SMIC alone.
In 2021, the semiconductor industry reported shipments of about 14.8 billion square inches of chips. Assuming that half of that number are logic chips, then very roughly translated, this means we are adding about 20% additional total logic capacity into 2024-2025. And that's if all of these projects turn out as expected, definitely not a given.
20% additional capacity over the next 3 years does not feel wildly out of control. Especially if industry estimates are reporting that global semiconductor shipments - memory and logic - will rise by 10% in 2022 alone.
At the same time, the industry as a whole hasn't experienced a 20+% 3-year shipment growth period since 2018. And that led into 2019, which we talked about earlier. The question for me then is whether or not there’s enough leading edge logic demand for that much leading edge logic capacity coming up soon.
The Cloud Train
In the first semiconductor bust video, I asked the rhetorical question: Where is all this demand coming from? What’s going on?
A year later, I feel like I have a decent grasp on what the answer might be, and it is rather dull: The cloud.
It is a reversal from what drove logic chip growth over the past ten years, which was mobile. After the launch of the iPhone 3G, the App Store, and Android, smartphone revenues drove the logic industry out of the Global Financial Crisis.
Smartphones rapidly climbed the performance ladder faster than almost any previous consumer electronics product, demanding better cameras, faster apps, and higher resolution screens.
As recently as Q1 2020, TSMC was essentially a smartphone chip-making company. Smartphone chip revenues made up nearly 50% of their revenue mix.
By that time however, that segment was experiencing a bit of a stall - with revenue growth -9% quarter over quarter.
According to Gartner, worldwide smartphone sales volumes hit a steady state in 2016. Nearly 80% of people now own a phone, and people don't have to upgrade so often.
At this time TSMC and the rest of the semiconductor industry saw 5G as driving the next big wave of smartphone semiconductor revenues. They have been talking this investment thesis even as COVID started to take hold in late 2019 and early 2020. Here is a quote from the call in January 2020:
> We continue to see stronger deployment of 5G networks and smartphones in several major markets around the world. We reiterate mid-teens penetration rate for 5G smartphones ... We also forecast a faster penetration of 5G smartphone as compared to 4G over the next several years while silicon content of 5G smartphone will be substantially higher than that of a 4G smartphone.
The pandemic caught them and the rest of the industry unawares because the demand it generated came from the data center rather than the smartphone.
The exploding video game industry.
The Direct-to-Consumer streaming revolution.
The work from home miracle.
And the unprecedented surge in AI machine learning adoption and its use cases.
These all require computation in the cloud. The biggest players in this cloud computing ecosystem are the hyperscalers - Amazon, Google, and Microsoft. These cloud computing companies suddenly find themselves growing 30% year over year despite massive revenue bases.
In the fourth quarter of 2021, AWS revenues grew to nearly $18 billion a year - 40% year-over-year growth. The company has a $71 billion revenue run rate.
In the first quarter of 2022, Microsoft reported that their Azure cloud services grew 50% year over year. Revenue sizes weren't reported.
The hyperscalers are some of the only companies projected to spend more on capital expenditure than the chip foundries are: An estimated $150 billion in 2022. And that titanic spending filters through the entire supply chain.
In TSMC's recent Q1 2022 earnings, 41% of their revenues came from High Performance Computing or HPC - where they put their cloud revenues. Just two years ago it used to be 30%.
They also noted that HPC grew 26% quarter over quarter - far faster than almost every other category other than automotive, which is a far smaller part of the overall business.
The big question of course, is how long this cloud gravy train is going to last as the world emerges from the pandemic and the macro-economy experiences some speed bumps.
Over the past few months we have been seeing a number of "pandemic plays" like Zoom, Netflix, Peloton, and others experience reversals.
Recently, Netflix lost 35% of their market cap when they announced their first subscriber loss in 10 years and projected further declines in 2022.
Zoom mentioned in their recent Q1 2022 earnings call that their sales productivity - conversion rates - is returning to levels more like the pre-pandemic era.
And commerce is swinging away from e-commerce back towards brick-and-mortar.
The Wall Street Journal recently noted that E-commerce as a percentage of retail sales leapt to titanic highs in 2020, but have since returned to the secular trend. A snapback that's taken a toll on many e-commerce companies that scaled up during the pandemic.
I do think some things have changed. Many companies will stick to some form of work-from-home. Not all of the 55 million subscribers Netflix added in 2020 and 2021 are going away. E-commerce will still grow faster than the rest of the retail industry. Companies will continue migrating old on-premise installations to the cloud. And AI and electric vehicles are still a thing.
So the secular trend is still good. But I don't know if the cloud gravy train in the coming years will be as good as it was in 2020 and 2021.
Even if the cloud doesn't grow another 30, 40, 50% year over year however, things will be alright for the big foundries of the world and their new fabs. They can re-adjust their capacity pricing and find some other fabless customers with a cool idea to fill those capacity slots.
And like I mentioned in the first video, TSMC and Samsung prioritize market share over profitability. If things cool down, they will get more aggressive. But I do wonder what it means for everyone else, however. The second-tier fabs, the equipment suppliers, and the rest of the ecosystem.
In conclusion, here are some "things that I think".
One, cloud computing revenue growth will decelerate. Either because of the law of big numbers or a return to a form of pre-pandemic normal. This will cause some issues throughout the ecosystem.
Two, there's still a semiconductor bust coming. But it is more likely coming in memory than in logic.
Hyperscaler data centers have big demand for memory chips and memory makers will feel their slowdown more - because their products are still commodities.
Three, the massive build out in logic semiconductor capacity projected over the next 3 years probably won't cause a market-screeching crash like in 2009 - but it may result in slightly lower prices and some broken promises.
Four, just because we see a slowdown in the leading-edge does not mean we will see the current shortages in trailing-edge/legacy nodes - meaning 65 nanometers or older - going away. Those dynamics are way different. Same with specialty technologies like MEMS, analog, etc.
I did a video about the trailing edge shortage earlier that may or may not be out right now. Sign up for the Patreon if you don’t see it on this channel right now.
And last, I have no idea what this will do to any of the big chip companies or their stocks. Don’t ask me about that.
Alright then, a bunch of new things to make me look like an idiot.