AI Scaling Laws

Overview

AI scaling laws describe empirical relationships between compute investment and model capability. Researchers have identified three distinct scaling regimes: pre-training scaling (more data and compute yields better base models), post-training scaling (reinforcement learning and fine-tuning further improve capability), and test-time scaling (more inference compute via "long thinking" yields higher quality outputs). Together, these three laws drive 100× growth in compute requirements at data center scale.

AI development has passed through three waves since AlexNet in 2012: Perception AI (speech recognition, medical imaging, image classification), Generative AI (multimodal content creation enabled by transformer architectures and unlabeled data), and Reasoning/Agentic AI (planning, autonomous vehicles, and robotics). Each wave has required progressively more compute, with the agentic AI wave now demanding compute resources orders of magnitude larger than the generative AI era.