Training one model consumed more energy than most households use in years. Then they trained the next version.
Literal meaning: The single, computationally intensive process during which an AI model learns from vast datasets. It happens before the model is ever used — but accounts for a substantial share of its total environmental cost.
Origin: The term comes from machine learning engineering. It gained broader analytical significance when researchers began publishing energy figures for specific training runs — most notably GPT-3 (2020) and subsequent large language models — making the scale of pre-deployment energy use publicly legible for the first time.
The one-time but enormous energy expenditure of building an AI model — which then needs to be repeated for every new version.
The Appeal: For developers, the training run is where capability is created. Spend energy once, deploy indefinitely. Amortised across millions of users and billions of queries, the per-interaction cost can be made to look small. This logic is not dishonest — it reflects how the economics genuinely work.
The Friction: The amortisation logic conceals the absolute scale. A single training run for a frontier model can emit hundreds of tonnes of CO₂ — a figure that does not shrink because many people benefit. And training is not a one-time event: models are retrained, fine-tuned, and replaced in rapid cycles. Externalized Costs — environmental burdens placed on communities rather than absorbed by companies — mean that the communities near the data centres bear the concentrated impact while the benefits are distributed globally.
Why This Matters: The training run makes visible a structural asymmetry: cost is concentrated in place and time, benefit is diffuse and global. “A new model was released today” becomes a different kind of sentence once you know what it cost.
Related terms: AI Energy Consumption · Carbon Footprint of AI · Inference Cost · Externalized Costs · Sacrifice Zones · Digital Colonialism · Greenwashing
Read more: -Energy and Policy Considerations for Deep Learning in NLP — Strubell, Ganesh & McCallum (2019), ACL -Carbon Emissions and Large Neural Network Training — Patterson, D. et al. (2021). arXiv