1981: The Open Box That Won

"IBM published the circuit diagrams. Everything the PC industry became follows from that one decision."

A Box Built From Other People's Parts

On August 12, 1981, IBM unveiled the IBM Personal Computer. By IBM's own account of its history, the team "embraced an open architecture and even published a technical reference of the circuit designs and source codes". The processor was Intel's, the operating system was licensed from Microsoft, the expansion slots followed a bus anyone could build a card for, and the reference manual told you exactly how. IBM did this because it wanted the machine out in a year and could not build every part itself. What it got, within a few years, was a market where anyone could build the whole machine — the clones — and a business that IBM eventually left altogether.

That is the founding event of the modular computer industry: a machine defined by its interfaces rather than by its maker. A slot has a pinout; anything with the right pins works. A socket has a specification; any chip that meets it drops in. A disk connector, a memory module, a display port, a keyboard — every one of them became a boundary with a published contract on each side, and every boundary became a market.

From Vertical Silos to Horizontal Layers

Andy Grove, who ran Intel through the transition, described the shape of what happened in Only the Paranoid Survive (1996): the computer industry stopped being a set of vertical companies, each making its own chips, machines, operating systems and software, and became a set of horizontal layers — chips from a few vendors, machines from many, one or two operating systems, applications from thousands — with each layer competing inside itself and meeting the layers above and below at standard interfaces. The Windows-and-Intel stack was the layered industry's most successful instance, and for two decades it was, for practical purposes, the industry.

The layered model has properties worth naming precisely, because the rest of this track weighs them:

Substitution. Any part can be replaced by a competitor's part that meets the interface. This is where price competition and choice come from.
Independent pace. Each layer improves on its own schedule. A faster GPU does not wait for a new CPU.
Repair and upgrade. A failed or outgrown part is swapped, not the machine.
Distributed innovation. Nobody has to ask permission to build a better card.

Every one of those is a real good, and the founder of this household has built on all of them for decades. When he says he is against monolithic design, this is the thing he is for.

Apple, the Counter-Example, Then and Now

Apple was the vertical company that did not disaggregate. It designed the machine, the operating system and, for long stretches, the applications, and it sold them as one object. Through the 1990s that nearly killed it — the layered industry was cheaper and faster on every axis that mattered to buyers then. The Intel Macs of 2006 to 2020 were a partial surrender to the horizontal model: standard processors, standard memory, standard graphics parts. Apple silicon is the un-surrender. The processor vendor is Apple; the memory is on the package; there is no graphics card. The Mac of 2026 is more vertically integrated than any Mac since the 1980s, and in the same period Apple added tiers the Intel era never sold in this shape — a desktop whose GPU can address 512 GB (the 2019 Intel Mac Pro took 1.5 TB of DDR4, but its graphics cards saw only their own VRAM), a laptop with 128 GB — rather than retreating to the commodity middle.

That pairing — the most integrated design in the market, and the one reaching furthest up — is not a paradox. It is exactly what Christensen's theory predicts under a specific condition, and lesson four will name it. First, two lessons on what an interface actually is and what it costs.

Code

layers.py — a vertical stack vs a horizontal industry, as data·python

#!/usr/bin/env python3
"""Two ways to build a computer industry. The interfaces are the difference:
in the horizontal model every boundary between layers is a market."""

VERTICAL = {           # one company owns the column
    "Apple (2026)": ["processor", "memory packaging", "GPU", "OS", "frameworks", "machine"],
}

HORIZONTAL = {         # each layer is its own market; rows meet at interfaces
    "processor":  ["Intel", "AMD", "..."],
    "memory":     ["any DIMM vendor meeting the JEDEC spec"],
    "GPU":        ["NVIDIA", "AMD", "Intel"],
    "OS":         ["Microsoft", "Linux distributions"],
    "machine":    ["hundreds of OEMs and every home builder"],
}

INTERFACES = ["x86 ISA", "DIMM socket (JEDEC)", "PCIe slot", "ATX board + PSU", "USB", "DisplayPort"]

print("vertical:")
for owner, stack in VERTICAL.items():
    print(f"  {owner}: {' -> '.join(stack)}  (one vendor, no published seams)")

print("\nhorizontal:")
for layer, vendors in HORIZONTAL.items():
    print(f"  {layer:10} {', '.join(vendors)}")
print(f"  seams: {', '.join(INTERFACES)}")
print(f"\n{len(INTERFACES)} published interfaces = {len(INTERFACES)} places a competitor can enter")

Exercise

List every published interface in the last PC you owned or built — the slots, sockets and connectors a competitor could design a part for. Then list the interfaces in the Mac on your desk that an owner can reach after purchase. Count both. Which of the four modular goods (substitution, independent pace, repair, permissionless innovation) does each list support, and which does the Mac list not support at all?

Hint

The Mac list is short and mostly about the edges of the machine: Thunderbolt, USB, the display, storage over Thunderbolt. Inside the package there is nothing to list. That asymmetry is the whole argument, and it is exactly as real as the bandwidth number it buys.