In 1965 Gordon Moore, then research director at Fairchild Semiconductor and later a co-founder of Intel, predicted that the number of components on a chip would keep doubling every year for the next decade. In 1975 he revised the pace to roughly every two years. The Computer History Museum notes that the prediction came true for decades, partly because the industry used it as a target: chipmakers planned their roadmaps to keep up.
The everyday result is why a phone today outperforms a room-sized computer from the 1970s at a fraction of the cost. For a business, it meant you could often wait a couple of years and buy the same computing power for much less.
For AI, Moore’s law is part of the backstory of deep learning. Neural networks had been studied for decades, but they only became practical once cheap, dense chips, especially GPUs, made it possible to train large models on large datasets. Rich Sutton’s essay on the bitter lesson argues that methods which ride this steady fall in the cost of computation beat clever hand-built ones over time.
The trend has slowed. Shrinking transistors is harder and more expensive than it was, and gains now come increasingly from specialized chips, larger packages and many chips working together. Editorially, it helps to keep Moore’s law separate from scaling laws. Moore’s law is about how much hardware you can buy; scaling laws describe how model quality improves when you spend more of that hardware on training. Today’s demand for AI compute is growing faster than Moore’s law alone would supply, which is one reason the industry is building ever larger data centers.
Sources
- Computer History Museum: Moore's Law — Explains Gordon Moore's 1965 prediction that components on integrated circuits would double every year for ten years, and that transistor counts on the largest microprocessors doubled about every two years from 1971 to 2010; checked October 24, 2026.