Every question has a bill. Small per query. Enormous in aggregate.
Literal meaning: The energy consumed each time a trained AI model generates a response. Inference — applying a model to new input — is distinguished from training — building the model. The cost per query is small. Multiplied across hundreds of millions of daily queries, it is not.
Origin: The term is native to machine learning engineering. Its emergence as a public accountability concept followed the rapid scaling of AI usage from 2022 onward. Researcher Alex de Vries published the first systematic estimates of cumulative inference energy demand in Joule in 2023, bringing the figure into public and policy debate.
The energy consumed each time an AI model responds — invisible to the user, real at the infrastructure level, and growing with every new deployment.
The Appeal: Per query, inference costs are genuinely small — fractions of watt-hours for a typical exchange. AI can answer questions faster and at lower marginal cost than many human-staffed alternatives. Researchers and companies building on AI infrastructure have legitimate reasons to find this cost acceptable.
The Friction: The problem is aggregation and invisibility. At hundreds of millions of daily queries, small numbers compound into large ones — and Carbon Footprint of AI calculations almost never include them. Companies regularly publish training emissions; inference emissions are rarely disclosed. The interface is designed to feel weightless: instant, frictionless, free. Externalized Costs make the weightlessness possible. The Attention Economy sets the volume: a business that earns more as it raises the number of interactions per user is also raising the number of inferences per user. The design that makes an interface hard to put down is the design that runs up the bill. Someone pays. It just is not the person typing.
Why This Matters: Inference cost introduces a new environmental reading: how often matters as much as how large. A daily habit becomes infrastructure. The daily conversation with an AI assistant is a recurring small claim on shared energy resources, invisible by design.
Related terms: Training Run · AI Energy Consumption · Carbon Footprint of AI · Externalized Costs · Greenwashing · Sacrifice Zones · Attention Economy
Read more:
- The growing energy footprint of artificial intelligence — de Vries, A. (2023). Joule
- Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models — Li, P. et al. (2023). arXiv