the environmental cost of training
/ the en-vy-run-MEN-tul kost of TRAY-ning /
Training a large AI model means running thousands of specialized chips flat-out for weeks or months, and that takes a lot of electricity — which, depending on where it comes from, can mean a lot of carbon emissions. It also takes water, much of it used to cool the data centers, and rare materials to build the hardware. "The environmental cost of training" is the accounting of that physical footprint: the energy, the carbon, the water, and the waste behind a capability that feels weightless on your screen.
It helps to separate two phases. Training is the one-time, intensive build of a model — eye-catching numbers (a single large model's training can emit as much carbon as several cars over their entire lifetimes). But inference — actually using the model, every query, every image generated, billions of times a day — adds up over time and, for popular models, can dwarf the training cost in total. The footprint depends heavily on details people often omit: the energy mix of the local grid (coal vs. hydro), the efficiency of the chips and data center, and whether spent heat or water is reclaimed.
Why it matters: as models get bigger and use spreads, AI's slice of global electricity demand is growing fast enough that grid operators now plan around it. The honest picture is mixed, not apocalyptic: the same AI can also help cut emissions (optimizing grids, materials, logistics), efficiency per computation keeps improving, and AI is still a small fraction of total energy use compared with, say, transport or heating. The fair critique isn't "AI will boil the planet" — it's that the costs are real, often hidden from users, unevenly distributed (the water and power are consumed somewhere specific), and rarely disclosed honestly.
A data center in a hot, dry region trains a model around the clock. The chips run on the local grid — partly coal-fired — and are cooled with evaporated water drawn from a stressed local supply. The same model, trained in a cold region on hydroelectric power with recycled cooling, could have a small fraction of the footprint. Same model, very different cost.
Where and how a model is trained matters as much as how big it is — the footprint isn't a single fixed number.
Beware single scary numbers quoted without context. A carbon figure means little without the grid's energy mix, and training cost alone ignores the often-larger lifetime cost of inference. The fair point isn't that AI is uniquely destructive — it's that the costs are real, location-dependent, and far too rarely measured or disclosed.