For roughly half a century, computing rode two coupled trends. Moore’s Law, the 1965 observation by Gordon Moore, said the number of transistors on a chip would keep doubling every couple of years. Dennard scaling, named for a 1974 IBM paper, explained why those extra transistors could be switched faster without raising the chip’s power density. Together they delivered something extraordinary: software got faster every year for free, just by waiting for the next chip. The end of Moore’s Law refers to the gradual breakdown of this bargain.
The first crack came around 2005, when Dennard scaling failed. At very small transistor sizes, leakage and heat stopped cooperating, supply voltages could no longer be cut in step, and clock speeds stopped climbing. The industry’s response was to stop making single processors faster and start putting many cores on a chip. Transistor counts kept rising for a while longer, but the economic and physical returns from shrinking them thinned, and Moore’s Law itself slowed and grew far more expensive per generation.
This is the deep reason modern AI looks the way it does. When you can no longer count on a faster general-purpose CPU each year, the way to get more performance is parallelism and specialization - throwing many simple compute units at a problem and tailoring the silicon to one kind of workload. That is exactly what GPUs, Google’s TPU, Microsoft’s FPGA fabric, and the wave of AI accelerators do. The same pressure also revives interest in radically different substrates - neuromorphic, analog, optical, even quantum - as routes to keep improving when conventional scaling no longer delivers.
Why business readers should care: the end of Moore’s Law is the quiet force behind the AI hardware boom and its economics. Performance now comes from buying more specialized chips and burning more power, not from waiting for free speedups - which is why compute and energy have become strategic, budget-defining concerns.