Segment 1 — The Cold Open

Segment 2 — The Old Way (Reasoning By Analogy)

Before 1947 the standard way to control current was the vacuum tube. Engineers at RCA, Western Electric, and General Electric built each tube as a small glass envelope containing a heated filament, a cathode, one or more grids, and an anode. The parts had to be aligned inside a vacuum, sealed, and tested individually. A typical receiving tube in a radio or early computer cost between two and five dollars in 1940s money, roughly the price of a good meal. The convention had been set by the radio industry in the 1920s and simply scaled up for radar and computing during the war. Because every tube was a discrete mechanical assembly, the cost floor was set by the number of hand operations and the yield losses from cracked glass or contaminated seals. No one inside the tube industry questioned whether a solid-state alternative could exist; the assumption was that amplification and switching required a vacuum and heated electrodes. That assumption kept the price of logic and amplification high enough that a machine with thousands of switches remained the domain of governments and large corporations. One objection often raised is that tubes were reliable enough for their era and that incremental improvements in sealing or filament materials could have narrowed the gap. Yet those tweaks still required separate glassworking, evacuation, and alignment steps for each device, so the per-unit labor and scrap stayed locked in place. Another point is that the radio market had already amortized tube tooling across millions of units, which made the design feel inevitable rather than chosen. The result was that any system needing more than a few hundred active elements stayed expensive by construction, because each added element multiplied the same assembly sequence rather than sharing a common surface.

Segment 3 — The First-Principles Move

Bell Labs researchers in 1947 started from the physics of semiconductors rather than from the tube catalog. They knew that a junction between p-type and n-type silicon could rectify current, and that a third electrode could modulate it. The first point-contact transistor still required hand assembly of two metal whiskers on a germanium crystal, so its cost remained high. The decisive step came with the grown-junction and later diffused-junction transistors of the early 1950s. These devices were formed by adding tiny amounts of dopant atoms to a silicon crystal during growth or by diffusing them through a mask. The raw material cost per device was already low: metallurgical-grade silicon cost pennies per gram, and a single gram could supply the atoms for thousands of transistors once sliced into wafers. The Idiot Index of the old tube was therefore enormous; a finished tube costing several dollars contained perhaps ten cents of metal and glass. The new transistors began to close that gap once photolithographic masking replaced individual wire attachments. Each mask step defined millions of junctions at once across a wafer. The key engineering trade-off was achieving uniform doping and clean oxide layers so that yield did not collapse as feature sizes shrank. Once the planar process developed at Fairchild in 1959 allowed the entire transistor, including its contacts, to be formed on a flat surface without breaking vacuum, batch fabrication became reliable. A single 1-inch wafer could now carry hundreds of transistors instead of one. Later, the move to integrated circuits placed multiple transistors and their wiring on the same die, eliminating separate packaging steps for each device. Each of these changes attacked a different term in the cost equation: the number of manual alignments, the number of separate packages, and the fraction of the wafer lost to defects. Because the dominant cost was no longer the silicon itself but the sequence of photographic and chemical steps, further reductions came from larger wafers, better steppers, and higher yields rather than from cheaper raw material. The magic-wand floor remained the commodity price of purified silicon plus a few dopant atoms and aluminum interconnects, a figure already below one cent per die by the mid-1960s and far lower today. One natural question is whether the same outcome could have been reached by simply scaling tube production; the answer lies in the physical limit that each tube still needed its own glass envelope and vacuum seal, so labor and material waste scaled linearly with quantity instead of being amortized across an entire wafer surface. Another objection is that early transistors were fragile and temperature-sensitive, yet those problems were solved by moving to silicon and controlled oxide growth rather than by refining the tube envelope. The arithmetic becomes clearer when the wafer is viewed as the unit of production: if a polished silicon slice costs a few dollars and yields several hundred working transistors after all process steps, the material contribution per switch drops into the millicent range even before accounting for later improvements in wafer diameter and defect density.

Segment 4 — The Result & The Limits

By the mid-1970s the price per transistor in high-volume logic had fallen below 0.1 cent, and it has continued to drop by roughly an order of magnitude every decade since. The Idiot Index for a modern logic transistor is now close to one; the finished cost is dominated by the amortized cost of the fab and the energy of the process steps rather than by any scarcity of atoms. The remaining hard limits are the physics of atomic-scale leakage, the capital cost of extreme-ultraviolet lithography tools, and the yield losses that still occur when a single defect kills an entire large die. Those constraints are genuine and have slowed the cost curve in recent nodes, yet they are still orders of magnitude below the cost structure of vacuum-tube manufacturing. A further point worth noting is that even when feature sizes approach atomic dimensions, the batch nature of the process keeps the per-device overhead from rising back toward the old discrete-assembly baseline. The capital intensity of the latest tools is high, but that cost is spread across billions of transistors per wafer rather than across a few hundred hand-built units.

Segment 5 — The Lesson

A device whose function can be printed across an entire surface at once reveals that its price is set by the area of silicon processed rather than by the number of individual assemblies. The second principle is that once the atoms themselves are cheap, the remaining work is to make every subsequent patterning step cover more devices without adding proportional cost. The same logic that collapsed the transistor now points toward other fields where batch surface processing could replace piece-by-piece construction. What would be the first product whose price would shift if someone applied the same wafer-scale arithmetic to it?

💬 Reply to this email — Patrick reads every one.

Share: X · LinkedIn · WhatsApp

Forwarded this email? Subscribe here — it's free.