Luccioni, Strubell & Crawford (FAccT 2025) invoke the Jevons paradox to argue that per-model AI efficiency gains will not reduce, and may increase, total AI energy consumption

In a FAccT 2025 paper (arXiv:2501.16548), Sasha Luccioni, Emma Strubell and Kate Crawford apply the nineteenth-century Jevons paradox directly to artificial intelligence: "just because an AI model becomes more efficient, that does not imply that overall AI resource consumption will decrease, and in fact the inverse effect is highly plausible." The argument is the coal-and-steam rebound effect transposed to compute — cheaper, more efficient inference lowers the effective price of an AI query and so expands the volume of queries, model sizes, and deployments, potentially raising total energy demand even as each unit becomes more efficient. Where the rebound exceeds 100%, efficiency is net-consuming.

The authors ground this with the observation that per-unit hardware efficiency gains have coincided with soaring aggregate demand — illustrated by the industry shipping on the order of millions of GPUs in 2024 (the specific ≈3.7M figure is the paper's cited market number, flagged as [unverified-quant] and routed to question-verify-luccioni-2025-nvidia-gpu-rebound-figures). This is the "road home" leg of a cross-century bridge: it connects a Victorian economist's coal argument to the present-day compute cluster the vault tracks — the collapse of per-query inference cost, the fact that inference now dominates AI compute, and the resulting grid pressure captured in claim-ercot-large-load-queue-quadrupled-2025-phantom-load and claim-bit2watt-gpu-scheduling-destabilizes-power-grid. See observation-jevons-double-bridge-victorian-logic-machine-to-ai-energy and moc-inference-economics.