In a recent Substack post called “The Marginal Revolution is Dead, Long Live the Overhead Revolution”, Arnold Kling makes the point that, for many businesses today, the marginal cost to create an incremental additional good is close to zero. A commercial flight costs essentially the same amount no matter how many people are on the plane, so filling the empty seat is pure profit. An amusement park can admit an additional guest without significant incremental additional costs. A school can serve an incremental additional student without hiring more teachers or building new classrooms, and so on.
And this is even more true for digital goods, whether it’s software or a piece of music or a YouTube Video. A piece of software is all overhead, i.e., sunk costs, and the cost of distribution is trivial relative to the incremental additional revenue from a new sale.
As I was reading through the post, I couldn’t help but think about the semiconductor industry. The successful firms operate on a similar set of principles. But I’ve never seen anyone quite articulate these principles when discussing the semiconductor industry. Most of the commentary instead focuses on the industry’s traditionally cyclical nature, on the intense amount of competition, on the manufacturing process, on the secular trends (e.g., AI, mobile, automotive, etc.) that affect the industry. I think Kling’s “overhead revolution” may provide an additional useful framework to understand what’s going on in the industry.
I’m an amateur economist (at best!), but I figured I’d take a stab at explaining how the semiconductor industry works, using a bit of Kling’s framework.
Some background.
I often think of the modern semiconductor manufacturing process as something like a book-printing operation. There’s a design phase, a front-end manufacturing phase, and a back-end manufacturing phase, which are roughly equivalent to the writing phase, the printing phase, and the binding phase of a book. Very roughly. Bear with me…
The design phase starts out as something similar to writing a piece of software. Engineers write code that describes the various logical parts of a specific semiconductor design, that perform a particular set of functions. That initial set of, essentially, software code is transformed into a three-dimensional design file, describing a layout of transistors and other features on multiple layers of metal. There is an entire economy in the digital space, consisting of third-party “IP providers” like ARM who provide some of the building blocks for semiconductor designs, and “EDA Tool Vendors” like Cadence and Synopsis who provide the software that enables a semiconductor company to design, simulate, and validate its semiconductor designs before production. But for the purposes of this discussion, those details aren’t particularly important.
What is important is that the output of the semiconductor design process is a fully realized semiconductor design as a digital file, and nothing more. To make a semiconductor that file needs to go through the manufacturing process. In my book-making analogy, this is equivalent to the writing, editing, and layout phases – you end up with some sort of digital file, maybe a pdf, that you send to the print shop.
The front-end of the manufacturing process consists of wafer fabrication (at a “fab” or manufacturing facility), which is really the high-value, high-technology part of the process. TSMC is the once-obscure, now-famous company that currently leads the world in this sort of manufacturing. In my analogy, they’re essentially the big printing press that the “publishing houses” – i.e., “fab-less” semiconductor companies – use to produce their “books”. There are different processes, called “process nodes”, that describe the different generations of production, with each generation representing smaller and smaller features. So you might hear 28 nanometer, 64 nanometer, 3 nanometer, and so on – these all describe different generations of process technology. The wafer itself is a circle of glass coated in layers of metals, with multiple copies of a particular semiconductor design printed in a grid pattern on its surface.
The back-end of the manufacturing process consists of packaging and testing. This is a lower-tech and more competitive part of the supply chain. There are a lot of companies in this area that most people never would have heard of. These companies cut up the silicon wafer into single “dies”, or single copies of a semiconductor chip, and package those on what’s called a “substrate” – a plastic circuit board that takes the tiny features on the semiconductor die and makes them larger for connecting to a traditional circuit board – in a protective plastic capsule, to create a finished product that looks basically like what you’d think of when you think of a “chip”. In my book-making analogy, think of this part as where the large sheets of paper that come off of the printing press get cut up and bound into a book with a hard cover.
Here’s where we get to the “overhead revolution”
Here’s the thing: it costs basically the same to make every book. If you’re making an obscure treatise on Foucault, it costs the same as the latest book from Dan Brown. And it’s the same way in the semiconductor industry: a fabless semiconductor company – and these are really who we’re talking about when we refer to a “semiconductor company” – pays more-or-less the same to make a wafer’s-worth of semiconductor products on a given process node.
All else being equal, if your design costs are sunk, and your manufacturing costs are fixed, you’d prefer to “print” copies of semiconductors that are going to generate the highest value. And just like the book industry, some semiconductors are remarkably profitable, and others are kind of meh. For example, a GPU is going to fetch top dollar in today’s AI-fueled boom, whereas a CPU for a mid-level laptop will not.
Now, you need scale to justify the cost managing a complex semiconductor supply chain and to get the time and attention of important suppliers like TSMC, so you need to manufacture a sufficient volume of products. You can’t just run the high-end stuff all of the time. And you need enough diversity in the products you produce to maintain that scale, and revenue, across the business cycles of the downstream industries you serve. So there are some tempering factors.
But as a general matter, you are going to try to produce as many of the products as you can that will fetch the highest dollar you can get, because the cost of an incremental additional semiconductor – i.e., the manufacturing costs of the packaged wafer – is the same no matter what you manufacture, and that cost should be small compared to what you’re able to charge for the finished good.
Again, back to the book analogy, a bible and a new book by J.K. Rowling probably cost the same to print, but one is going to cost marginally more than the printing cost whereas the other is going to cost an amount that makes the printing cost look trivial.
Kling notes: “Managers put a lot of effort into coming up with ways to implement price discrimination. They don’t put effort into making marginalist calculations.”
And so it is!
Design is Overhead
Kling notes, “Garett Jones once observed in a tweet that most people don’t produce widgets. Instead, they work on producing organizational capital. Organizational capital includes new products as well as projects to improve the efficiency of production and distribution.”
In the context of a semiconductor company, this is certainly true, particularly with respect to the core design work that is transformed into the semiconductor itself. Getting from “no semiconductor design” to “version 1 of a commercialized semiconductor design” is a tremendous undertaking. And I want to emphasize the “commercialized” part of the prior formula: it is not enough to get to a design, you must get to a design that is integrated successfully into other products and provides the intended real-world functionality. This is a non-trivial and under-appreciated undertaking. Going from no idea to an idea is much easier than turning that idea into something that you can produce in the real world, and much-much easier than turning that “something” you can do once into something that you can mass produce reliably.
Once you have that version 1, and let’s assume that version 1 is successful in the marketplace, going to the incremental versions 2, 3, 4, and so on is a much-less significant effort. And, as with everything in the technology space, if you’re just a bit better than the next person in line, you tend to win the bulk of the business. Certainly, in many areas, the top 2-3 seem to generally win out in a way that is nearly insurmountable for others to overcome, except through a fundamental technology shift that puts an incremental improver at a disadvantage to the innovator.
I think this is where we stretch our book-analogy a little thin: it’s certainly true that the big “investment” with J.K. Rowling was the first Harry Potter book, that once that was a hit, and once she’d built the world and characters, going to the incremental second through seventh books was relatively less effort.
In contrast, in the semiconductor industry, it’s almost as if Rowling could re-write the first book, adding a chapter, or maybe weaving in a new character here-and-there, and sell a book all over again. I mean, they try to do this with special leather-bound editions. And they sort of do this through the movies, although there are tremendous production and marketing costs there. But there’s still a lot of incremental effort that goes into getting the next egg out of the goose compared to semiconductors.
Semiconductors aren’t alone in this regard: the iPhone has settled into a routine where each new version is somehow the same thing with some upgrades. For a few generations, the music industry made incredible amounts of money selling the same people the same albums in LP, then Philips cassette tape, then CDs, then MP3s. (See, e.g., the Dark Side of the Moon is still on the top 200 album sales chart after 972 weeks.) Maybe Porsche is able to do this somehow with the 911, and Chevrolet with the Corvette: i.e., iterate on something iconic to generate an entirely new sale. Amusement parks, going back to the example earlier in the post, and that Kling mentions, are selling you the same experience over and over again, maybe adding a Star Wars ride to get you to purchase again what they’ve already sold to you.
The classic example of this, of course, is software. How many copies of Photoshop and Illustrator has Adobe sold over-and-over to the same group of people? Certainly the first version and today’s version are almost laughably indistinguishable from each other. But if you were to look at each incremental version, the improvements would be very small and subtle. And the software industry has mostly abandoned the pretext of “selling” you anything at this point, instead opting to simply charge you a monthly subscription fee to access their current version.
So what does all of this mean?
As I mentioned in the top of this post, people often miss the real core of what drives the semiconductor industry, and Kling’s notion of the “overhead revolution” captures what I think is missed – at least in much of the commentary and reporting that I read. One might be tempted to draw the conclusion from my comments that the original creation is the important thing, I suppose. But, as I see more of the industry and as I get more experience in life, I think it’s somehow the opposite. And this is the insight I suppose that struck me in Kling’s article: it’s the incremental improvement on a differentiated foundation that is really the important thing to creating value, in an environment where the incremental cost of production is trivial.