Training a large language model can require thousands of specialized chips running for weeks. Its carbon footprint is the greenhouse-gas emissions attributable to that work, commonly expressed as carbon-dioxide equivalent. Electricity used during training is one part of the picture, but the environmental cost does not end when training stops. section 1.14 section 11.3
The boundary of an estimate matters. A narrow calculation might count electricity consumed during training. A broader one can include power used while equipment is idle, estimated emissions associated with servers and chips, and power consumed when people use the model. That use is called inference: generating responses or predictions with a trained model. section 1.14 section 11.3
Putting an estimate in perspective
One study estimated that electricity consumed while training GPT-3 caused 500 metric tons of carbon dioxide emissions. For comparison, the cited annual averages were about 18 metric tons per American and 7.4 metric tons per person globally. These figures provide scale; they do not establish the footprint of every model or of GPT-3’s later use. section 1.14
Footprints are difficult to compare precisely. Chip efficiency, data centers, electricity sources, model size and architecture all affect an estimate, while companies may withhold relevant details. Accounting methods are not standardized. Inference also deserves attention: repeated model responses and tool calls can increase an agent’s energy use, though agents are not inherently more energy-intensive. A useful comparison therefore asks what activities were counted, not just which model has the smaller reported number. section 1.14 section 11.3
