An Observation That Became a Deadline: Moore's Law

In 1965 Gordon Moore, who would soon co-found Intel, wrote a short magazine article making a modest-looking prediction. He had noticed that the number of components the industry could pack onto a single integrated circuit had been roughly doubling every year, and he guessed the trend would continue for at least another decade. He later revised the doubling time to about every two years. That casual extrapolation, drawn from just a handful of data points, became known as Moore’s law, and it went on to describe the most sustained exponential growth in the history of technology, holding for half a century.

The numbers are worth seeing laid out, because exponential growth is genuinely hard to feel. Here are representative chips across the decades, with their transistor counts plotted on a logarithmic scale, where each step up is a factor of ten and a steady doubling appears as a straight line.

 transistors (log scale)
 100 B |                                                  * M2 Ultra (2023)
  10 B |                                          * M1 (2020)
   1 B |
 100 M |                              * Core i7 (2008)
  10 M |                       * Pentium 4 (2000)
   1 M |                  * Pentium (1993)
 100 k |            * 386 (1985)
  10 k |       * 8086 (1978)
   1 k |  * 4004 (1971)
       +----+----+----+----+----+----+----+----+----+----
        1971  1978  1985  1993  2000  2008  2020  2023

Chip (year)	Transistors
4004 (1971)	2,300
8086 (1978)	29,000
386 (1985)	275,000
Pentium (1993)	3,100,000
Pentium 4 (2000)	42,000,000
Core i7 (2008)	731,000,000
Apple M1 (2020)	16,000,000,000
Apple M2 Ultra (2023)	134,000,000,000

The points fall close to a straight line on this log scale, which is the visual signature of exponential growth. From the 2,300 transistors of the first microprocessor to well over a hundred billion on a modern chip is a factor of roughly fifty million, achieved in about fifty years, by halving the size of the basic switching element again and again until transistors are now measured in handfuls of atoms.

What I want to argue is that Moore’s law was never really a law in the way physics has laws. It described no force of nature. It was, at first, an observation, and then it became something stranger and more interesting: a self-fulfilling prophecy. Once the whole industry believed the doubling would continue, everyone planned around it. Chip designers, equipment makers, and software companies all set their roadmaps to the expected cadence, each assuming the others would deliver their part on schedule. The semiconductor industry literally published shared roadmaps committing to the pace. The prediction became a deadline, and meeting it required staggering, coordinated investment, because each new generation of miniaturization demanded factories costing billions and physics problems solved on a timetable. Moore’s law held not because nature guaranteed it but because an entire global industry organized itself to make it hold.

That distinction matters now, because the cadence is faltering, and the reasons are instructive. One pillar fell around 2005. For decades, shrinking transistors had also let chips run at higher clock speeds while using the same power, a bonus relationship sometimes called Dennard scaling. That broke down: as transistors got very small, they began leaking current and generating heat that could not be removed fast enough, and clock speeds stopped climbing. The industry’s response was to stop making single processors faster and start putting many processors, or cores, on one chip, which is why the device you are reading this on has multiple cores rather than one impossibly fast one. The free lunch of ever-rising clock speeds was over.

The other pillar, the shrinking itself, is now running into the wall of the atom. A transistor feature only a few atoms across cannot be made meaningfully smaller, because below a certain size the quantum behavior of electrons, particularly their tendency to tunnel straight through barriers that should stop them, makes the switch unreliable. We are not at the absolute limit yet, but the easy, steady halving of two-dimensional features has slowed, and each further step costs more and buys less than it used to. The industry has turned to other tricks to keep gains coming: stacking circuitry vertically into three dimensions, building chips out of smaller specialized pieces called chiplets, and designing processors tailored to specific tasks rather than relying on general-purpose speedups. These keep useful performance rising, but they are not the same clean geometric shrink that Moore described.

So the honest assessment is mixed. The specific thing Moore predicted, the relentless doubling of transistors on a flat piece of silicon, is winding down, and proclamations of its death, premature for decades, are finally becoming reasonable. But the broader pattern, of computing capability growing fast through whatever means the industry can muster, persists by other routes. What is ending is less a law of physics than a particular fifty-year strategy that ran until it hit the edges of what atoms allow. That such a strategy worked for so long, sustained by belief and money as much as by engineering, is itself one of the more remarkable facts about the modern world. An offhand prediction from 1965 turned into a promise the whole industry spent half a century keeping.