AI Energy Use

The two scales tell different stories. Per task, the numbers are small: Google estimates that a median text prompt to its Gemini app uses about 0.24 watt-hours, roughly what a modern TV draws in a few seconds. Company-reported figures depend on what they include and are hard to compare across providers. In total, the numbers are large and rising: the International Energy Agency estimates data centers used about 1.5% of the world’s electricity in 2024 and projects their use will more than double by 2030, with AI the main driver.

Training a large model is a big one-time cost; inference, answering everyone’s requests, is spent every day and grows with use. Reasoning models and agents that generate many tokens per task use more energy per answer than a short chat reply. Efficiency gains from better chips, quantization and smaller models push the other way.

For the AI industry, power has become a practical constraint as much as an environmental one. New facilities need grid connections, which can take years to arrange, so access to electricity now shapes where and how fast AI capacity gets built. For an individual user, the environmental case is mostly about those system-level choices rather than any single prompt; What AI actually costs the environment covers the water and carbon debates in more detail.

Guests come at it from both scales. In episode 45 at 35:11, writer Andy Masley argues that “computing is just so wildly energy efficient” that most individual computer use carries a small climate cost; that is his position in a contested debate. Episode 45 was sponsored by Galileo (sponsor thanks at 0:00). In episode 44 at 32:42, Cisco’s Jeetu Patel observes that data centers “are getting built where the power availability is the highest.” During that episode, Conor notes that the show’s sponsor at the time, Galileo, is an NVIDIA AI factory partner (31:57).