An observation, not a law of nature
In 1965, Gordon Moore, who later co-founded a major chip company, published an article noting that the number of components that could be economically placed on a single integrated circuit had been roughly doubling every year, and he expected the trend to continue for some time. In 1975 he revised the expected pace to a doubling about every two years. The popular name “Moore’s law” came later.
It is worth being clear about what this is. It is an observation about industry trends and economics, not a physical law. It held for decades because companies treated it as a target and invested heavily to meet it. Popular versions often say that computers double in speed every two years or eighteen months. That is not what Moore described, and the statement about speed has been loosely attached to the original claim over time.
What doubling every two years means
Doubling compounds quickly. If a count doubles every two years, then after ten years it has doubled five times, which is 2 to the power of 5, or 32 times as many. After twenty years it has doubled ten times, which is 1,024 times as many. That kind of growth is why a phone today can outperform computers that once filled rooms. It also shows how hard the trend is to sustain: every step has to be bigger in absolute terms than the last.
Why the free ride ended
For a long stretch, shrinking transistors brought several benefits together. Smaller switches were cheaper per unit, faster, and used less power. Engineers sometimes describe the power part as Dennard scaling, the observation that as transistors shrank, their power density stayed roughly constant. Around the middle of the 2000s that stopped working so neatly. Shrinking further made electricity leak more, and chips began generating more heat per area than could be removed cheaply. Clock speeds, which had been rising steadily, largely leveled off.
The industry response was to place several processor cores on one chip rather than push one core ever faster. That helps only for software that can divide its work across cores, which is not always easy.
The rising cost of each step
Transistor counts have continued to rise, but several pressures make progress slower and more expensive than before:
- Physics. Features are now only a few dozen atoms across in some dimensions. At such sizes, quantum effects such as electrons passing through thin barriers make switches harder to control.
- Heat. Packing more transistors into the same area makes the power density and cooling problem harder.
- Manufacturing cost. The equipment needed for the finest patterns is extremely expensive, and the factories that use it cost very large sums. As a result, fewer companies can afford to stay at the leading edge. The basic steps involved are described in explanations of semiconductor manufacturing.
- Memory and data movement. Processors have often become faster than the memory that feeds them, so moving data can cost more time and energy than calculating with it.
Most observers agree that the doubling cadence has slowed compared with its heyday, though how much depends on what is measured, and companies disagree about how long it can be sustained. Process names in nanometers are no longer reliable measures of real feature size, so comparing generations by name alone can mislead.
What is carrying progress forward
If shrinking alone gives less each time, improvement has to come from other directions. These are possibilities that the industry pursues, with varying success.
Better transistor structures. Designs have moved from flat transistors to raised and wrapped-around shapes that give tighter electrical control at small sizes.
Specialized chips. Instead of one general-purpose processor, designers add circuits tuned for a particular job, such as video, signal processing or AI. The rise of graphics processors in AI work is one example of hardware that suits a particular kind of calculation. Specialization can bring large gains for the right task but gives little for others.
Three-dimensional stacking and chiplets. Stacking layers of circuits, and combining several smaller dies in one package, can shorten the distance data travels and improve manufacturing yield.
Software and algorithms. Efficiency gains in code, compilers and methods used in machine learning can deliver improvements without changing the hardware at all.
New materials and devices. Researchers study alternatives to silicon for some parts of a chip, and new kinds of memory.
Different computing models. Some research aims to rethink the basics rather than shrink them. Neuromorphic computing borrows ideas from the brain, photonic computing uses light, and quantum computing, built around qubits, targets certain problems rather than general tasks. These are active research areas; none is a drop-in replacement for today’s processors, and timelines are uncertain.
What it means for everyday users
For consumers, the slowdown shows up as less dramatic generation-to-generation jumps in raw processor speed than in the past, devices that stay useful longer, and a greater emphasis on features such as battery life, cameras and built-in AI functions. It also means that the price of cutting-edge manufacturing is shaping which companies can compete.
For software, it means that performance can no longer be assumed to arrive for free with new hardware. Programs have to be written more efficiently and to spread work across many processing units.
A cautious outlook
Predictions that the trend would end have been made for decades and have repeatedly been proven premature, which is a reason for humility. At the same time, the economic and physical pressures are real. A reasonable expectation is that computing will keep improving, but through a mix of many smaller advances, in design, packaging, software and specialized hardware, rather than the single steady doubling that the original observation described.