Moore's Law
The pattern that ran for sixty years
Description
Few empirical observations have shaped a technology industry as thoroughly as the one Gordon Moore made in 1965. Writing in the trade magazine Electronics, the thirty-six-year-old director of research at Fairchild Semiconductor noted that the number of components engineers could fit on an integrated circuit had been doubling roughly every year, and he projected the trend would continue for another decade. The article was four pages long, the graph at its center had only five data points, and the title gave no hint that it would later be cited as one of the most consequential forecasts of the twentieth century. Moore was describing a curve. The industry that grew up around him decided to make it a destination.
What followed was sixty years of compounding. The doubling cadence held, with adjustments. Transistor counts went from around sixty in 1965 to billions by the late 2010s. The smartphone in a back pocket today contains more computing capacity than the supercomputers that simulated nuclear weapons in the 1980s, and costs less than a week of groceries. The world we inhabit search engines, video calls, machine learning, GPS in every car, the entire software industry sits on top of that curve. When people talk about the digital revolution, they are often describing the cumulative effect of Moore's Law running for two human generations.
The curve has now slowed in ways that matter. Transistors have reached dimensions measured in handfuls of atoms. The cost per transistor, which fell reliably for fifty years, has stopped falling at the leading nodes. New approaches are filling the gap, but the era when a single number doubled on schedule is ending. What Moore actually wrote, how a forecast became a roadmap, why physics pushed back, and what the post-Moore landscape looks like together tell the story of how computing got cheap.
The question we're asking: what was Moore's Law, why did it hold for so long, and what does its slowdown mean?
What we'll see: the original 1965 paper, the self-fulfilling roadmap, the physical limits, and the post-Moore era.
Table of contents
01The 1965 paper
Gordon Moore wrote the article that would carry his name because the editors of Electronics asked him to. The magazine was preparing a thirty-fifth anniversary issue and wanted predictions about where electronics was heading. Moore had access to production data nobody outside the industry could see. He plotted the number of components per chip against the year of introduction, drew a line through the points, and extrapolated. The data ran from 1959 to 1965, and the line suggested a doubling roughly every twelve months. He projected it forward to 1975 and predicted around sixty-five thousand components on a single chip by then.
The paper itself was not framed as a law, and Moore did not call it one. The term Moore's Law was coined around 1970 by Carver Mead, a Caltech professor who used the phrase in lectures. What Moore actually argued was more nuanced. He suggested that the economic optimum the chip complexity that minimized cost per component was itself moving, because manufacturing yields improved as engineers learned to work at smaller scales. The doubling was a statement about the economics of a young industry, and about the rate at which engineers were learning to make things smaller without making them more expensive.
02From forecast to roadmap
A forecast becomes a roadmap when enough actors decide to coordinate around it. By the 1980s, the semiconductor industry had become a complex web of suppliers, equipment makers, design firms, fabrication plants, and customers, each needing to plan multi-year investments. The doubling cadence gave them a shared timeline. Equipment manufacturers knew chip companies would need new lithography tools by a particular year. Chip companies knew their customers would expect a particular price-performance improvement on schedule. Customers wrote software assuming the hardware underneath would get faster on the curve. The expectation was self-reinforcing.
The Semiconductor Industry Association formalized the coordination in 1992 by publishing the first National Technology Roadmap for Semiconductors, later expanded into the International Technology Roadmap for Semiconductors. The document specified, year by year, the technical milestones the industry needed to hit feature sizes, wafer dimensions, lithography wavelengths, materials. It was produced by consortia of competing companies who agreed they had a shared interest in keeping the curve going. Suppliers built tools to meet the next milestone, foundries planned plants for the one after that, and the cadence continued.
03The physical limits
Physics began to push back in earnest around 2005. The clock speed of processors, which had been climbing alongside transistor counts, hit a wall near four gigahertz. The reason was heat. Transistors that switch faster dissipate more power, and chips were starting to consume hundreds of watts in areas the size of a thumbnail. Higher clock speeds meant chips that would melt without exotic cooling. The industry pivoted to multicore designs, putting several slower processors on the same chip. Transistor counts kept climbing, but user-visible gains slowed. The single-thread performance that defined personal computing for decades stopped doubling.
The shrinking of the transistors continued, but the names of the process nodes became misleading. A node labeled seven nanometers does not contain features that are seven nanometers wide; the number is a marketing convention that rolled over from when it described a physical dimension. The smallest features on a leading-edge chip today are around twenty nanometers in some directions and below ten in others. At these scales, the channel through which current flows is only a few dozen atoms across. Quantum effects electrons tunneling through barriers that classical physics says should block them become significant, and a single misplaced atom can change a transistor's behavior.
04The post-Moore era
Specialization is the first major response to the slowdown. For decades, general-purpose processors got faster every generation, so it rarely paid to design custom chips for narrow tasks. Once that improvement curve flattened, custom chips became attractive. Graphics processors, originally designed for video games, turned out to be ideal for the matrix arithmetic machine learning requires, and Nvidia rode that match into one of the most valuable companies in the world. Google designed its own Tensor Processing Units for AI workloads. Apple designs its own chips for iPhones and Macs, optimized for the workloads its devices run.
Three-dimensional stacking is the second major direction. If features cannot easily get smaller in two dimensions, engineers can build upward. Modern memory chips already stack dozens of layers of storage cells, and logic chips are starting to follow, with companies bonding silicon dies together vertically and connecting them with through-silicon vias. The most aggressive packaging schemes treat what was traditionally a single chip as a system of chiplets smaller dies, each manufactured on the process best suited to its function, packaged together with high-bandwidth interconnects. The unit of innovation is shifting from the transistor to the package.
05Conclusion
Moore's Law was never a law in the sense that gravity is a law. It was an observation about a young industry that became a coordination mechanism once enough actors decided to plan around it. Its sixty-year run is one of the most remarkable feats of sustained industrial progress in history, and the connected, computational world we take for granted is largely its accumulated output. The slowdown does not erase any of that. The transistors already shipped are still shipping, and the next decade of computing will be built on the same silicon foundations as the last.