Moore's Law

Overview

Moore's Law is the empirical observation made by Intel co-founder Gordon Moore in 1965 that the number of transistors on an integrated circuit doubles approximately every two years, and by extension that computing performance roughly doubles at the same rate. In the AI era, the law remains partially true — physical design scaling delivers approximately 2x improvement per year in computing TOPs/watt, interconnect bandwidth (GB/s), and HBM memory bandwidth — but the rate of improvement is slowing as process nodes approach atomic limits.

The structural challenge for AI is that demand for training compute has grown at approximately 10x per year for the past six years (as evidenced by NVIDIA GPU scaling from GPT-1 through GPT-5), creating a 5x annual gap between supply (Moore's Law) and demand (AI). This gap cannot be closed by process scaling alone, driving investment in advanced packaging (3.5D), chiplet architectures, custom ASICs, and Co-Packaged Optics as the primary mechanisms for continuing to scale AI compute capability.