Moore's law is the observation, made by Intel co-founder Gordon Moore in 1965, that the number of transistors on an integrated circuit doubles roughly every two years, a rule of thumb that held with remarkable consistency for over four decades and became a general shorthand for the exponential growth of computing power. In artificial intelligence, "Moore's law" is often invoked loosely to describe a related but distinct trend: the rapid growth of computational power applied to training and running AI models, which has at times moved even faster than classic transistor scaling.
Traditional transistor-density Moore's law began slowing significantly in the 2010s as chipmakers approached physical limits of miniaturization, prompting the semiconductor industry to rely more on specialized architectures, chip stacking, and larger dies rather than pure density gains.
Compute scaling in AI
A widely cited 2018 analysis from OpenAI, "AI and Compute," found that the amount of compute used in the largest AI training runs had doubled roughly every 3.4 months between 2012 and 2018, far outpacing the classic two-year doubling of Moore's law, driven by researchers' growing willingness to spend on larger training runs rather than by chip improvements alone. This compute growth underlies empirical Scaling laws describing how model performance improves predictably as parameters, data, and compute increase together.
Hardware behind the curve
General-purpose GPUs, originally built for graphics rendering, became the dominant AI training hardware because their parallel architecture suits the matrix multiplications at the core of neural networks. Nvidia's CUDA software platform, launched in 2006, gave it a durable advantage once deep learning took off after the 2012 AlexNet result. Google's custom TPU accelerators, first deployed in 2016, offered an alternative path optimized specifically for neural network workloads and have been used to train models such as Gemini. Nvidia CEO Jensen Huang has argued that combined hardware and software gains in AI compute now outpace historical Moore's law, a claim central to Nvidia's rise to one of the world's most valuable companies by the mid-2020s.
Limits and pushback
Rising power consumption, chip fabrication costs, and data center construction have become binding constraints distinct from transistor density itself. Events such as the January 2025 DeepSeek shock, in which a competitive model was reportedly trained far more cheaply than assumed, renewed debate over whether raw compute scaling is the only path to better AI, with techniques like mixture of experts and Quantization offered as ways to extract more capability per unit of compute rather than simply spending more compute.