In my previous post, I talked about Apple’s strategy with the M1 chip, and the technology that allows it to compete with Intel’s chips. The M1 uses a RISC instruction set architecture (ISA), a simpler architecture than Intel’s which allows for lower power consumption, better cooling, and more efficient out-of-order execution. But this is not Apple’s first time using RISC; before switching to Intel Apple used the RISC-based PowerPC architecture. This architecture should have all the advantages the M1 has today, and yet it failed to gain adoption and died off after 10 years. Why?

To understand, first we need to take a look at the CPU industry, and take a look at what allowed Intel to remain the dominant CPU vendor for 40 years.

Intel’s Competitive Advantages

The CPU industry’s market share has been stable since the early 1980s, after the introduction of the IBM PC. IBM’s open PC design had two proprietary components: the Microsoft MS-DOS operating system and the Intel CPU. The decision for IBM to outsource these two components of the PC has led these two to become the largest earners from the PC boom.

Intel has kept their marketshare dominance for a majority of the past 40 years. In the early days Motorola was a major competitor but has since fallen behind, and AMD sometimes makes an inroad but usually Intel manages to catch up. Others such as NEC and Texas Instruments have attempted to gain entry into the market but have failed. While keeping competitors at bay, Intel has reaped massive profits, with average returns on capital at 30 percent after tax. Ratio of market value to estimated replacement costs of assets has continually exceeded 3 to 1, which means each dollar invested by Intel creates three dollars in shareholder value. (Greenwald & Kahn 61)

How is this possible? How can Intel simultaneously be extremely competitive and be extremely profitable? In the CPU industry there is free entry for anyone who wants to create an Intel competitor which produces a similar result at a lesser price. Other silicon products such as controllers, flash memory, and DRAM are extremely competitive industries where dozens of producers push for the lowest price possible for their product, leading to average returns on capital before tax of 6–8%.

The difference between the CPU industry and other silicon-based industries is that the CPU industry has a set of competitive advantages to the incumbent player. The CPU industry has (1) high economies of scale and (2) consumer captivity. These two things combined are a very powerful way to keep competitive forces at bay and profits unusually high. Intel has these advantages on two fronts: their CPU R&D and their manufacturing plants.

Let’s start with economies of scale. CPUs have a high fixed-to-variable cost ratio. Fixed costs include R&D costs and costs of buying/renting manufacturing equipment, while variable costs are the costs of the raw materials used per chip. Fixed costs for computer chips are in the tens of billions while the variable cost for each chip is very small, less than $100 per chip. If Intel were to sell a single chip, not a single model but literally make one CPU, that chip would cost tens of billions of dollars. This chip would cost so much because you need to recoup your R&D costs and costs of buying manufacturing equipment. But now that you’ve got the chip design and the factory equipment, why not produce another chip? So Intel produces two chips, and now each chip costs slightly over half of the single chip from before. The cost of each chip is now fixed cost ÷ 2 + variable cost. If Intel produces three chips, each chip will cost fixed cost ÷ 3 + variable cost. Continue this process and you see that the cost of chips goes down the more of them you make. To make a profit, Intel has to sell millions of chips and spread their fixed cost over all chips sold. This concept of efficiency growing with the size of the operation is called economies of scale. Competitive advantages from economies of scale come from relative sizes between firms; smaller firms won’t be able to match the costs of a larger one even though they have equal access to technology and resources because they can’t reach the same scale of operation. Since average cost¹ goes down with the quantity sold, a larger firm can be highly profitable at a level where smaller firms, with higher average costs, are losing money. (Greenwald & Kahn 38)

The automotive industry also has high fixed costs in manufacturing, and automotive parts need to be sold in the millions to recoup the costs of the factories. This is why there is so much consolidation among automotive groups; it’s more efficient to share parts over a number of brands instead of each brand having their own manufacturing plant. But the car industry is rife with competitors, unlike the CPU industry where Intel has remained dominant. This is because the car industry is missing something Intel has: captive customers.

If an entrant has equal access to customers that the incumbent has, they will be able to reach the incumbent’s scale. A market where all firms have equal access to customers and common cost structures, where entrants and incumbents offer similar products on similar terms, should divide up evenly among competitors. This theory holds true for the other silicon manufacturers I mentioned before. (Greenwald & Kahn 38)

For economies of scale to be a legitimate advantage, you have to combine it with some consumer captivity. This way you can prevent entrants from gaining the customer base needed to lower their average costs to a reasonable level. If the incumbent has consumer captivity, entrants will have a difficult time catching up and will remain on the wrong side of the economies of scale differential. (Greenwald & Kahn 38)

Computer manufacturers are accustomed to working with Intel and appreciate the level of quality, supply stability, and service support they receive from them. Competitors such as AMD may perform just as well in these areas, but with a smaller market share they haven’t been able to form the same relationships with manufacturers. If Intel and AMD produce next-generation CPUs with similar prices, features, and release dates, you can expect Intel to capture a dominant market share. All Intel has to do is match AMD’s product to retain its market share. Even if AMD manages to produce a superior product to Intel, computer manufacturers will almost certainly give Intel a grace period to catch up, rather than switch immediately to the new supplier. Out of the 40 years of Intel’s dominance, only in the past three has AMD managed to begin cutting away at Intel’s marketshare. (Greenwald & Kahn 40–41)

PowerPC

The transition to Apple silicon is Apple’s fourth CPU architecture transition. Apple initially used Motorola 68k processors (CISC), then moved to AIM PowerPC (RISC), then Intel x86 (CISC), and now finally Apple silicon (RISC).

After Steve Jobs was ousted from Apple in 1985, CEO John Scully became the sole leader of the company. From his vantage point he could see that, although Apple computers were doing well at the time, they would need a major change to remain competitive in the future. Out of the many problems Apple was facing, one was the fact that the Motorola chipsets were falling behind Intel’s. The Mac was becoming the machine with a higher price tag and lower performance.

The reason for Motorola’s failure to compete with Intel was due to their lack of economies of scale. Intel sold their chips for Windows machines which made up over 80% of the market, while Apple’s marketshare peaked at 12%. Motorola’s average costs were many times greater than Intel’s, choking their R&D. Eventually they were eliminated as a competitor.

Scully decided it was time for Apple to switch its CPU architecture. But instead of switching to Intel, he switched to PowerPC, a family of chips built by the AIM alliance: a collaboration between Apple, IBM, and Motorola.

Surely Scully knew why Motorola was failing against Intel. But he actively chose to create the AIM alliance and compete with Intel instead of adopting them. The reason is RISC: experimental chips at both Apple and IBM have shown that the technology had the chance to dominate Intel’s x86. And they did for the first couple generations; PowerPC chips benchmarked well against Intel chips for a number of years. Microsoft even released a version of Windows NT for PowerPC. Unfortunately customers of Windows showed very little demand for the new chips, as most applications never bothered to support the new architecture.

Even though Intel was using a technology worse at its core, the switching cost for developers and customers was too high. Developers and consumers are stuck in a negative cycle: consumers won’t buy computers with the new architecture if developers don’t bring their applications there, and developers won’t bring their applications to the new platform if consumers don’t buy it. This cycle caused PowerPC to fail immediately on every operating system except Mac OS, where Apple created a 68k emulator to help old applications to run on the new platform for the time being. Unfortunately the Mac was a minor player compared to Wintel machines, and PowerPC faced higher average costs from Intel for its entire run. Just like Motorola, PowerPC died from high fixed costs.

Although Intel had a source of consumer captivity against AMD, it is not remotely as strong as the captivity Intel has over other instruction set architectures. Intel and AMD both use the x86 instruction set, so programs that run on one will run on the other just fine. The consumer captivity Intel has against AMD is from their reliability as a partner for computer manufacturers, who don’t want to switch to a new vendor for a component as critical as the CPU. However other ISAs such as PowerPC can’t simply run x86 applications on their own, they need support from developers to port the apps over. Switching from one ISA to another takes significant development time, and can easily take over a year for some applications. If the application contains third party libraries or is built on third party tools, then switching will be put on hold until the third parties update their code. Consumers won’t ever make the switch if the future for their preferred applications is this uncertain, so they will prefer chips from Intel or AMD.

Why Apple Silicon is Different

Apple is trying a RISC architecture again, this time with Apple silicon. Although it seems like they are making the same mistake again, this time around there are differences in the market which should make the switch successful.

ARM

Apple silicon uses the ARM ISA². ARM is developed by ARM Holdings, a company based in the UK. The people at ARM develop the ISA and reference processors, but they do not compete in the processor market directly. Instead they license their architecture to companies like Apple, who develop a processor around it for their own custom use case. ARM’s license is fairly cheap and very generous in how much the client can change the ISA to suit their needs. This is in stark contrast to Intel/AMD’s x86 ISA, which can’t be licensed to anybody on any terms. ARM’s ISA is widely used among a variety of processor designers, which is what allowed them to slowly take marketshare from Intel. Instead of tackling Intel head-on like AIM tried to with PowerPC, ARM instead simply created the ISA and put it on sale, allowing many smaller suppliers to use it, and these suppliers combined provide the economies of scale to tackle Intel.

ARM’s RISC architecture got a boost when the smartphone market formed. Intel’s chips use excessive power and run hot, which is unideal for small battery-powered devices with no active cooling. The smartphone market started using ARM right from the beginning, and now ARM has become the dominant supplier in that market. Because smartphones were new, naturally there wasn’t any consumer captivity at the start because there were no consumers of smartphones before they were invented. A budding market provides opportunity for the superior technology to take foothold, which is exactly what ARM did. Every smartphone operating system, from iPhone to Android to KaiOS, runs on processors using the ARM ISA. In fact, smartphone sales have been higher than PC sales for the past couple of years, so ARM actually has the economies of scale in their favor compared to Intel/AMD.

Apple is extending ARM’s architecture from the iPhone to the Mac. Because this architecture is widely used, developers already have experience with it and are much more likely to pick it up. In addition, Apple’s Rosetta 2 (Intel x86 translation layer for Apple silicon) can help smooth over the transition in the meantime, making sure customers don’t immediately reject the new computers like they did with Windows on PowerPC.

TSMC

Intel not only has economies of scale with their R&D, they also have it with their fabrication factories. The tooling required for making a new line of chips costs billions, so you have to churn out a lot of chips to offset the costs. Apple currently has 9% global marketshare of PCs, so if they were to manufacture their chips in-house they would face much higher average costs than Intel. But Apple is not manufacturing their chips in-house, instead they are outsourcing the manufacturing to TSMC (Taiwan Semiconductor Manufacturing Company). Just like ARM, TSMC takes many clients, as opposed to Intel who only fabricates their own chips. Therefore TSMC has better economies of scale than Intel. And it shows, because Intel has been having issues manufacturing anything smaller than 7nm while TSMC has already announced 4nm. Apple’s M1 is using the 5nm process, so right off the gate they have an advantage over Intel, an advantage Intel has had a difficult time catching up on.

A Larger Market

Growth of the market is the enemy of economies of scale advantages. A larger market means that your fixed-to-variable cost ratio will go down. Say it takes $1 billion in R&D to make the latest generation CPU, a cost that stays constant no matter the size of the market. Say Intel has $25B in revenue and ARM has $5B. Then Intel’s R&D makes up 5% of the firm’s costs while it takes up 20% of ARM’s. But if the market doubles, where Intel is making $50B and ARM is making $10B, then R&D makes up 2.5% and 10% of each firm’s costs. The differential has shrunk, and the fixed costs which were once choking ARM are less of a problem now. (Greenwald & Kahn 41)

With the advent of smartphones, the market for CPUs have indeed doubled in size. But ARM Holding’s has taken the smartphone market all for itself, giving them a larger marketshare than Intel. But my point stands, a larger market means that economies of scale barriers dissipate and more specialized and niche firms now have the chance to tackle the incumbent. Where once Intel did all the ISA R&D, processor design, and fabrication, now specialized firms can do all these tasks individually. ARM Holdings develops the ISA, Apple develops the processor specifications, and TSMC manufactures the processor. A larger market has created room for these new firms to rise and create new economies of scale advantages of their own. Niches create riches, and Intel’s field has shifted from a niche market to a mass one.

---

In short, Intel’s economies of scale advantages in R&D and in manufacturing have been eaten away by the competition. The ARM ISA took a foothold in the smartphone market and grew until they had larger marketshare than Intel. ARM Holdings decided to license their ISA instead of make chips in-house, which allows them to spread the cost of R&D over multiple processor developers. Just as ARM specializes in R&D over their ISA, TSMC specializes in chip fabrication, allowing many smaller processor developers to use their factory, spreading the cost of chip manufacturing over these clients. Apple is utilizing ARM’s ISA and TSMC’s chip fabrication to avoid the economies of scale problems they had with PowerPC. Finally, Apple is using Rosetta 2 to help early adopter customers use x86 applications on the new Macs, preventing an early death like PowerPC on Windows. All of these things combined means Apple silicon has a much better chance of succeeding than PowerPC did.

Sources:

Competition Demystified: A Radically Simplified Approach to Business Strategy by Bruce Greenwald and Judd Kahn. Some sections of this post are near verbatim from this book. Those sections have been marked with an inline citation.

PowerPC

AMD marketshare expanding

Intel cost of R&D

Example of Intel’s cost of a factory expansion

Operating system market share. These statistics line up with the more commonly cited “mobile vs desktop marketshare” statistics.

Evidence of Apple using custom instructions

PS: Here you can see why Apple decided to transition to ARM instead of the newer RISC-V. Although RISC-V has many benefits over ARM, it still hasn’t gained mass adoption yet and developers are not used to coding for the standard. The ISA doesn’t have the consumer base needed to tackle ARM.
In the future Apple can convert their coprocessors to RISC-V while keeping the central CPU on ARM. Because the coprocessors’ instruction set extensions are locked behind libraries anyway, there would be no need for a transition.

PSS: I mentioned that AMD has recently started to cut into Intel’s marketshare; they’ve done this by getting rid of their factories and outsourcing to TSMC. Since TSMC has higher economies of scale than Intel, they saved AMD enough resources to make a comeback.

Updated 2021-03-24, added inline citations