AI · SEMICONDUCTOR · FACT CHECK
SKD WIREPartly True
AI inference — the cost of running large language models (LLMs) once trained — is falling at a staggering pace. Epoch AI's research on "the plunging price of thought" puts the decline at roughly 47% per quarter for some models, a figure that supports the widely repeated claim. But the number applies to specific benchmark conditions rather than the market as a whole, making the blanket assertion only partly accurate.
Epoch AI's analysis tracks the falling price of LLM inference over time, with its figures illustrated in charts showing steep downward curves in the cost per unit of AI "thought." The trend indicates that for a given level of capability, the price of running AI models has dropped sharply quarter over quarter.
The decline is not uniform across all models, providers, or use cases. The rate of cost reduction depends on which models are compared, over what time period, and at what performance level. Costs for older-generation models fall faster than headline prices for frontier models, which often debut at a premium. As a result, a single per-quarter percentage figure — such as 47% — cannot be applied indiscriminately to all AI inference spending.
Epoch AI reviewed ten primary sources in compiling its analysis, and its published figures are accompanied by images of the underlying graphs, including a headline chart on the price decline and a follow-up figure breaking the trend down further.
The directional claim is solid: inference costs for LLMs are plummeting, and Epoch AI's data supports declines in the neighborhood of 47% per quarter under the conditions it measures. But the figure describes specific, benchmarked scenarios rather than a universal market rate, so the unqualified claim about AI inference costs falls short of full accuracy.
The claim that AI inference costs fall 47% every quarter is Partly True.
Verdict: Partly True
