0
Every time a new foundation model is released, the accompanying blog post mentions scale, but rarely discusses energy consumption. I’ve seen estimates ranging from a few hundred megawatt-hours to terawatt-hours depending on the architecture. Are these figures reliable, and what factors actually drive the variance?
Beyond raw compute, I’m wondering about the hardware lifecycle. Do the GPUs and TPUs used for training get recycled, or do they end up as e-waste once a newer generation arrives? If anyone has looked at lifecycle assessments or carbon accounting reports from major labs, I’d like to understand which stages—data center cooling, manufacturing, or decommissioning—carry the heaviest burden.
