What a Data Center Is (andymasley.substack.com)

This explainer argues that data centers are essentially “building-sized computers” — highly optimized, hyper-efficient concentrations of many tiny compute tasks — and that targeting them as an environmental villain is a mistake. Using thought experiments, the author shows that running a small local AI (100 prompts/day) consumes roughly the energy of a 4-minute microwave run or 10 minutes of gaming, adding about 1/1,000th to an individual's daily emissions. When many people centralize similar workloads, economies of scale (better cooling, power delivery, and utilization) reduce per-task energy, so a large data center often uses far less energy per user than dispersed home devices. Even model training looks small when amortized: roughly ~50 Wh of energy per eventual user, comparable to running a laptop for an hour. The significance for AI/ML is twofold: first, scale usually improves energy efficiency, so cloud-hosted inference and shared training infrastructure can be greener per user than local alternatives; second, data centers create concentrated local grid demand even while representing only a small slice of global emissions (the piece argues data centers are a tiny part of total CO2 despite large absolute consumption). The policy implication is to prioritize per-task energy efficiency and grid planning over demonizing data center size — they’re an environmental achievement in delivering vast compute to millions with comparatively low energy per interaction.