Scaling laws

Research at OpenAI in 2020 and DeepMind's 2022 Chinchilla work established that loss falls smoothly with scale and that compute is best spent balancing model size against training data volume — findings that let labs forecast the payoff of nine-figure training runs before committing to them, and that fueled the frontier compute race.

The original laws govern pre-training, but the same logic now extends along other axes: post-training quality, and most prominently inference-time scaling, where reasoning models convert extra thinking tokens at answer time into higher accuracy.

For buyers, scaling laws are the engine behind the market's defining trend: the price of any fixed capability level keeps collapsing as newer, better-scaled models push older performance tiers down the price ladder. Cost models that assume today's price-performance ratio will hold for years are systematically wrong.