The Environmental Cost of AI: What's Actually Being Measured
AI's environmental footprint is a genuine and contested topic. Here's what's actually being measured and debated.
Training and running large AI models requires substantial computing power, which translates into real electricity consumption and associated carbon emissions, though the specific scale of that impact is measured and estimated differently across studies.
What's relatively well established
Training a single very large AI model can consume electricity comparable to what a small number of households use in a year, and this energy use scales with model size, a genuinely documented and measurable cost of developing more capable models.
Where the numbers get more contested
Estimating the ongoing energy cost of actually running and using a trained model at massive scale across millions of users, as opposed to training it once, involves more assumptions and varies significantly depending on the specific method and boundaries a given study chooses, which is why total estimates of AI's environmental footprint vary widely across different reports.