The answer, according to a pre-print study, is about one. Researchers at AI company Hugging Face and Carnegie Mellon University found that general-purpose AI models like GPT-4 are “orders of magnitude” more power-hungry than purpose-made models powering products like Google Translate.

The study, though not yet peer-reviewed, puts into context the environmental cost of generative AI, particularly of inefficient models (one image from the least efficient image-creating model can use as much CO2 as an average gas car driving about 4 miles, for instance).