Training a large language model in a data center consumes orders of magnitude more power than a human brain.