Vals AI finds long-horizon jobs can leave a 10,000-times larger environmental footprint than simple queries

Longer, more complex jobs for AI models — such as building a software app — can carry an environmental footprint as much as 10,000 times larger than a simple query, and the energy used to build a web app with some models is equivalent to powering a typical U.S. home for two and a half hours, Bloomberg reported on a new audit of real-world tasks. The figures come from Vals AI, an independent benchmarking firm that estimated energy, carbon, and water use for open-weight models on long-horizon work rather than one-off chat prompts.