Computers will keep doubling in speed every two years, forever

What actually happened

Two things got conflated in the classroom. Moore's 1965 observation was about transistor density and its economics, not speed. Dennard scaling, from 1974, was the reason smaller transistors also meant faster and more efficient ones. Density kept improving; Dennard scaling broke down in the mid-2000s because leakage current and heat stopped cooperating. That is why desktop clock speeds have hovered around 3 to 5 GHz for two decades instead of reaching the 10 GHz that straight-line extrapolation promised.

The industry response was to go sideways: multiple cores, then specialized silicon. The most consequential performance gains of the last decade came from GPUs and machine-learning accelerators, which are fast because they are parallel and narrow, not because their transistors switch faster.

The International Technology Roadmap for Semiconductors, which coordinated the industry for decades, issued its last report in 2016 and handed off to a successor organized around applications rather than shrinking. Chips are still improving, through 3D stacking, chiplets, and new transistor geometries, and "Moore's Law is dead" gets declared too confidently. But the simple promised doubling that schools taught as a law of nature was an economic trend, and it slowed.