The cost of using machine learning intelligence is decreasing by several orders of magnitude each year, with no signs of slowing. LLMs are expected to become ubiquitous infrastructure, running locally on commodity hardware within 3‑6 years. Evidence includes proprietary models like GPT‑6 Astra and open‑weight models such as GLM‑5.3‑flash, as well as local models like Muse Glimmer and Qwen3 Coder.

GPU power efficiency follows a logarithmic curve, doubling roughly every two years—comparable to historic Moore's Law gains. Model pricing is per token, but task‑level cost is falling sharply; the Pareto frontier in 2026 shows models spanning from cheap, less intelligent options to expensive, high‑performance ones. Inference engines are rapidly improving, delivering 10‑50% yearly gains.

As token costs become negligible, quality and access are emerging as the new limiting factors for AI adoption.