In their 2019 ACL paper “Energy and Policy Considerations for Deep Learning in NLP,” Emma Strubell, Ananya Ganesh, and Andrew McCallum estimated that running a full neural architecture search to tune a transformer model - including all the trial models trained along the way - emitted roughly 626,155 pounds of CO2 equivalent. For comparison, the same paper cited a car including fuel over its entire lifetime at about 126,000 lbs, so the search was on the order of five car-lifetimes. The figure became the most-cited single number in the early debate over the carbon cost of AI.