Discover more from The Asianometry Newsletter
Where The Real Chip Shortage Is
Author’s note: If you want to watch the video, here it is:
I don’t have much to add, except that despite the weak macroeconomic situation, chip facility buildouts in Taiwan continue to move forward. Notably, UMC - which specializes in the 65 to 28 nanometer range - reported monthly revenue growth of 35% in August 2022.
Things might be getting difficult on the higher end. But supply seems to remain short for these older chips.
Let me introduce to you the ATSAMD21 microcontroller or MCU, made and sold by Arizona-based Microchip Technology.
It is a small 48-megahertz ARM chip with 256K of flash and 32K of SRAM. It sells for about $3-4 per unit and can be deployed in automation, metering, and industrial manufacturing.
As of this writing, this MCU is out of stock for the next year. Same issue with many related systems: Out of stock until 2023 or beyond.
I have spent a lot of time on this channel talking about sexy, leading edge engineering things like Extreme Ultraviolet Lithography and so on.
But the reality is that the chip shortage's greatest hurt hits far from the leading edge. In this video, we are going to talk about the massive shortage in trailing-edge semiconductors, and why it's so hard to fix.
The Trailing Edge
Most chips aren’t sexy 3-nanometer monsters. They are unheralded workhorses doin' work and just living their lives as happy little chips. Those guys are built with what are called “trailing-edge” process nodes.
The definition of what makes a “trailing node” differs depending on your perspective. TSMC classifies that as anything at 16 nanometers or older. These nodes accounted for roughly 50% of 2021 revenue.
You can go a bit deeper from there. You can further subdivide them to mature and legacy nodes. Keeping in mind that definitions differ depending on the perspective, but mature nodes would generally be considered in the 40 to 450 nanometer range.
The legacy nodes would be anything older than that. Most microcontrollers for their part are made with a 110 or 180 nanometer process.
There are roughly 200 fabs right now operating around the world. Most of them are old - having been built before 2000.
At the start, these factories were brand new, full of expensive chip-making equipment for making leading edge or near-leading edge chips.
Such chips sell with high margins and that helps pay back the investment for the expensive equipment at the start.
But this situation doesn't last for long. The industry quickly moves on. Within a year or two, there is a new set of "leading edge" chips.
Pretty soon, the fab's chips prices collapse as its products commoditize. But once a fab is built, it doesn't really require that much more effort or investment. So long the market price is a little bit higher than the individual factory’s variable costs, it can keep on going.
Fab capacity at trailing-edge nodes had been already relatively tight for several years before the start of the pandemic. This is due to a few secular industry factors we will speak on later.
Then the first pandemic lockdowns in early 2020 sent economic activity to record lows. People feared the worst.
Automotive companies and other manufacturers halted their production lines and canceled orders at suppliers.
This filtered all the way down the supply chain and those guys powered down too. STMicroelectronics for instance cut production 50% at their most advanced fabs in France.
But that very same summer saw a massive, unexpected bounce back in economic activity. And that means everyone is now trying to get product out the door - resulting in shortages.
Companies first went through the inventories they had on hand. They then started buying more than what they needed so that they can replenish their depleted stocks and guarantee future output of their product.
All the meanwhile, the greater semiconductor industry is experiencing unprecedented demand for chips thanks to the cloud, the wireless 5G transition, and of course, armed conflict.
Thus since 2019, fab utilization rates have risen far beyond their long term targets of 80% - hitting levels in the mid-90s. It is an indication that foundries are hitting the limit in terms of what they can squeeze out of their existing fabs.
Microchip executive chairman Steve Sanghi said in a recent earnings call that there's $150 billion of additional investment needed over the next five years to meet trailing-edge demand.
Growth in building out trailing edge capacity is very slow - about 2% annually as compared to 25+% at the leading edge.
The semiconductor industry has committed hundreds of billions in flashy announcements of building new fabs in Asia, the United States, and Europe.
Yet only 6% of that expenditure is dedicated to building out mature or legacy node capacity.
Not So Simple
Okay then. So what’s the problem? Let’s pour a bunch of money in, build a bunch of fabs, and take over the market! But it’s not that simple. For a long time, trailing edge chipmakers have been hesitant to invest in such a thing.
Customers are used to paying prices set by foundries with fully depreciated fabs. Wafers for these trailing nodes are very cheap. Despite contributing significant microprocessor volume, they account for only 10% of total sales revenues.
Even in this situation of heightened demand, it is expected that spending hundreds of millions or even billions on a new fab would mean having to endure a period of substantial underutilization and financial underperformance. They would rather stick with their backlogs.
Furthermore, many of these foundries use older, niche equipment. For instance, over 200 fabs make 200 millimeter wafers which are smaller and older than the current industry standard of 300 millimeters.
New 200 millimeter equipment is not readily available. Companies prefer to buy refurbished stuff from resellers. With all the 200 millimeter fabs running at high capacity, supply here is short right now.
Historically, big flagship companies like AMD and Apple always wanted to stay at the fastest, sexiest leading edge nodes. When the industry moved on, they moved on. Then you got other companies coming in afterwards, scaling up their designs accordingly.
This upscaling trend has slowed down over the years. It is now getting harder to move onto the next node. Customers have to balance the benefits against the costs.
Some types of chips simply don't scale up all that well. Analog, mixed-signal, and radio frequency chips are one example.
MEMS are a similar case. These are a specialty silicon found inside items like accelerometers. Demand for all these type of chips are through the roof especially in light of the Ukrainian war.
And then there is the 28/22 nanometer generation gap. Why 28? That is because 28 nanometers is probably the most advanced node still using what is called a planar gate.
Anything newer than that means having to deal with the more complex FinFET gate, rampant multi-patterning concerns, and possibly EUV. These contribute to additional complexity, less reliable product, lower yields, and higher cost.
Thus, 28 nanometers represents the bottom of the cost curve on a per-100 million gate basis. Things get more expensive on a per gate basis as the effects of these new things start to kick in. Customers don't need to get that much faster especially if it means dealing with these aforementioned tradeoffs.
Automotive companies in particular want above all reliability and effectiveness across a vast variety of challenging conditions. They are some of the biggest consumers of trailing-edge chips - taking up 40% of market demand.
Breaking the Jam: Supply
At some point, the industry is going to finally notice that backlogs in their older nodes are stretching 18 months and longer. And they will have to do something about it. What are some of those things?
Firstly, you can bite the bullet and finally build more fabs. And pass the additional cost down to the customers. Everyone’s interested in this TSMC Arizona plant. And to have an N5 plant in the United States is pretty cool.
But to me, what’s just as interesting is that TSMC is also building two new trailing edge fabs in Japan and Kaohsiung. Those fabs operate on the 28/22 nanometer node range. TSMC recently announced price increases on older nodes and I think that’s to pay for these new fabs.
Samsung also recently announced that they will invest more in trailing edge semiconductor capacity. The company has generally focused on delivering only the latest greatest thing, so this is a big step for them.
China is doing its part here too with up to a dozen new fabs under construction across the mainland. SMIC is the leading Chinese foundry and they are spending over $9 billion to add trailing edge wafer capacity in Shanghai and Shenzhen.
India is making moves too. The conglomerate Vedanta recently signed an MOU with Taiwan's Foxconn to build a trailing edge fab in the 65-28 nanometer range with substantial government support.
And the US CHIPS Act - an American bill that authorizes incentives for construction and modernization in the American semiconductor industry. They appropriated $2 billion specifically for "mature technology nodes" - with that definition to be set by the Secretary of Commerce.
Even with all the incentives I feel like companies out in the West still aren't really leaping to build brand new 90 nanometer fabs. But they don't have to. There are other things that the industry can do to improve the operating performance of their old trailing edge fabs.
For instance, moving from 200 millimeter wafers to 300 millimeter wafers. Thanks to the laws of numbers, a 300 millimeter wafer can yield over two times as many dies as a 200 millimeter wafer.
A wafer upgrade is a substantial one. The whole floor layout needs to get bigger, automation systems have to be installed and there are wafer supply concerns. But it doesn't require a new fab to be built and you get double the output.
Work can also be done from the design side. I spoke in a previous video that one goal of open source EDA software and PDKs is to make hardware development cheaper. New tools can be introduced to make it easier for designers to design for or do a die shrink onto another node.
For instance, an open source 130 nanometer PDK released by Google and SkyWater Technologies. If it gets certain players to finally do long-delayed die shrinks then consumers get to enjoy faster and better chips. That can’t suck, right?
Anyway this will all take some time - at least one or two years - so supply constraints at the trailing edge will continue on until at least 2023.
A few things are moving in the right direction. Suppliers are saying in earnings calls that growth in revenues for trailing edge nodes is slowing. The big players as well as China and India are racing to build their big new fabs.
And overhanging it all is an apparent slowing in the macroeconomy. Any such cyclical downturn will dampen overall demand - perhaps just in time for all that supply capacity to hit the market. If we are lucky, I guess.
I want to thank Doug of Fabricated Knowledge for his work in this topic. I highly suggest you check out his newsletter for amazing research on the semiconductor industry.