AI Energy Consumption
Why it matters
Deep Dive
The first thing to understand is the split between training and inference. Training a frontier model is a single, enormous burst: tens of thousands of accelerators running in lockstep for weeks or months, with energy estimates for the largest runs landing in the tens of gigawatt-hours — roughly the annual electricity use of a small town. Inference is the opposite profile: each request is cheap, but billions of requests a day accumulate, and as deployed usage has exploded, inference has overtaken training as the larger share of AI's total electricity demand. Both phases ultimately land in the same place — the data center — where GPUs and TPUs convert power into heat that then has to be removed, which is why modern AI campuses are engineered as much around substations and cooling as around silicon.
Training vs Inference
A frontier training run concentrates its energy into a single window. The cluster behind it — tens of thousands of accelerators connected by high-bandwidth networking — can draw 100 MW or more of continuous IT load for the duration of the run, which is why training sites are chosen around available substations and cheap, firm power. Smaller jobs matter too: fine-tuning an existing model is orders of magnitude cheaper than training from scratch, but thousands of organizations doing it daily adds up.
Inference spreads the energy out instead. Each generated token costs a tiny amount of compute, but a popular model serves millions of users around the clock, so the meter never stops. Google's own 2025 measurement put a median Gemini text prompt at about 0.24 watt-hours, and independent estimates for comparable chatbots land in the same few-tenths-of-a-watt-hour range — small per request, enormous in aggregate. Reasoning models have pushed per-request energy up sharply, because they generate long hidden chains of thought before producing a visible answer. The rule of thumb: training energy scales with model and dataset size as described by scaling laws, while inference energy scales with users, context length, and output length.
Gigawatt Campuses and Nuclear Deals
The industry's response has been a buildout measured in gigawatts, a core part of modern AI infrastructure. Announced AI campuses now plan for 1 GW or more of capacity at a single site — the output of a large nuclear reactor — and the IEA estimates global data-center consumption at roughly 415 TWh in 2024, about 1.5% of all electricity, with that figure more than doubling by 2030 and AI as the main driver. The strain is already visible locally: US data centers drew about 4.4% of national electricity in 2023, with federal projections of 7–12% by 2028, and in Ireland they account for roughly a fifth of national demand, prompting restrictions on new connections around Dublin.
Because grids cannot expand that fast, the biggest buyers have gone straight to the source. Microsoft signed a 20-year agreement to restart the undamaged reactor at Three Mile Island, securing roughly 835 MW of round-the-clock carbon-free power; Google struck a deal with Kairos Power for small modular reactors totaling about 500 MW by 2035; Amazon followed with nuclear-backed campus deals of its own. These agreements are less about green branding than about physics: training and serving clusters need firm, 24/7 power, and only nuclear and a handful of other sources deliver that at gigawatt scale.
One Query Won't Crash the Grid
A common claim is that every chatbot query is an environmental disaster, and the per-query numbers do not support that. At a few tenths of a watt-hour, a text prompt is comparable to a couple of minutes of an LED bulb and far below earlier viral estimates that assumed every query triggered a massive fresh computation. The per-query figure also keeps falling as models and serving stacks improve.
But the opposite claim — that AI energy is a non-issue — is equally wrong. The impact lives in the aggregate: billions of daily requests, heavier modalities like image and video generation that cost far more than text, and demand that compounds as AI gets embedded into search, office software, and customer support. Both the doom headlines and the dismissal miss the real picture, which is a grid-level planning problem rather than a per-query guilt problem.
Efficiency Gains and the Jevons Rebound
The efficiency curve is genuinely steep. Quantization, distillation, speculative decoding, and mixture-of-experts architectures each cut the compute per token by large factors, and every accelerator generation from NVIDIA and its competitors improves performance per watt substantially. Serving stacks have gotten smarter too, with KV-cache reuse and continuous batching keeping utilization high, and some inference is shifting to edge devices where small models run locally.
Yet total consumption keeps climbing, which is the Jevons paradox in action: when the cost per token falls, demand does not stay flat — applications that were previously uneconomical switch on, and usage expands to fill the budget. Token prices have fallen roughly an order of magnitude per year while total inference volume has grown even faster. The realistic outlook is not that efficiency saves the grid, but that efficiency determines how much AI a given grid can deliver.