AI Power Demand: What the boom means for the climate

Many people worry about the electricity behind AI because training and running large models requires powerful computers in data centres. At the same time, everyday uses of AI — voice assistants, image search, recommendation lists — are mostly inference tasks that consume far less energy per query but add up because billions of requests occur every day. The contrast matters: a single large training run can draw hundreds to thousands of megawatt‑hours, while a single search or recommendation often uses only a few watt‑seconds.

Assessing the climate effect requires clarity about three things: how much electricity data centres and networks use overall; how much of that growth is due to AI specifically; and how much of the power is supplied from low‑carbon sources. This article uses recent public studies and industry figures to offer numbers you can hold in your head, practical examples you can relate to, and realistic steps that reduce the climate footprint of AI workloads.

Measuring AI power demand starts with two layers: the compute used for model development (training) and the compute used to deliver models to users (inference). Training often uses many GPUs or specialised accelerators for days or weeks; inference spreads smaller work over many requests and runs continuously. Analysts express compute using FLOPs (floating‑point operations), GPU‑hours, or directly in kilowatt‑hours (kWh). FLOPs count the number of basic math operations a model performs; it is a useful comparative metric but not an energy number by itself.

Two infrastructure metrics are especially important and often cause confusion. Power Usage Effectiveness (PUE) is the ratio of total facility power to IT equipment power; a PUE of 1.1 means about 10 % overhead for cooling and power distribution. Second, the carbon intensity of the electricity grid (kg CO2e per kWh) determines how much greenhouse gas results from a given kWh.

Clear, standard reporting of kWh, PUE and the grid carbon intensity is the simplest step to make AI energy claims verifiable.

Here are a few rounded reference numbers that help put scale into context. Note that the global estimates come from an IEA report published in 2023 and are therefore more than two years old; they remain useful because they provide a transparent, methodical baseline.

The training example above is based on public research that estimated training energy for large language models; that work dates from 2021 and is therefore more than two years old. Methodologies have improved since then, but the magnitude — hundreds to thousands of megawatt‑hours for the biggest training runs — remains a helpful rule of thumb.

Most people encounter AI through services: search, translation, photo enhancement, streaming personalization and chatbots. For these tasks the per‑request energy is small. A web search with ranked results or a short text completion may consume only fractions of a joule to a few joules on the server side. The environmental impact becomes meaningful because these small amounts are multiplied by millions or billions of daily requests.

Training sits at the other end of the spectrum. Big models are trained infrequently compared with inference, but each training cycle can be intensive: many accelerators running at high power for days or weeks. Open discussions around compute growth have shown rapid increases in the largest training runs over the past decade; at the same time, algorithmic improvements have reduced the compute needed for the same capability. Both trends operate in parallel: researchers can afford larger experiments because hardware and cloud resources scale, while smarter algorithms reduce the compute needed to reach a given level of performance.

Providers manage the balance by moving common tasks to specialised, efficient inference hardware, batching requests, using quantized models (which use less energy per operation) and employing caching so repeated results do not recompute from scratch. For users, some choices matter: using on‑device features when available (phone‑based inference for voice commands) shifts energy from the data centre to a device already being powered; choosing services with transparent energy or carbon options helps move demand to cleaner hours.

There are clear ways AI growth can coexist with climate goals, but there are also real risks if nothing changes. Opportunities include algorithmic efficiency, better software to use compute more sparingly, and shifting heavy workloads to low‑carbon regions or to times when clean electricity is abundant. Industry initiatives that buy renewable power or match usage to hours of high wind and solar can sharply reduce the carbon intensity of identical kWh.

Operational improvements matter: modern hyperscale data centres achieve low PUE values (near 1.1 in efficient sites), and using wastewater or air heat recovery can reuse part of the waste heat for district heating. At the system level, sharing large, well‑utilized clusters for research reduces duplicate training runs and lowers total energy per published result.

Risks are concentrated rather than diffuse. Point events — training a very large foundation model — can create short, sharp spikes in electricity demand and associated emissions if they take place in grids with high fossil intensity. Cryptocurrency mining is a separate example of concentrated load that can materially change estimates if included. A general rebound effect is possible: cheaper, faster AI enables more features and services, which in turn raise total demand. Transparent reporting and procurement rules can reduce these risks.

Scenario planning points to three broad possibilities. One: continued rapid growth in peak compute demand, if model sizes and training cycles keep expanding faster than efficiency gains. Two: a slower growth or plateau, if algorithmic and hardware efficiency compensate for larger models. Three: a managed pathway in which transparency, procurement and cleaner grids reduce net emissions even as compute grows. Which path unfolds depends largely on policy, procurement practices and corporate transparency.

For organisations and engineers the most effective moves are measurable and immediate: publish training‑energy numbers (kWh), PUE and whether developmental tuning runs are included; use carbon‑aware scheduling to shift flexible workloads to cleaner hours or regions; and invest in model compression and distillation so the same services run on fewer resources. For regulators and buyers, requiring minimum disclosure of energy metrics and rewarding 24/7 carbon‑free electricity purchases will change incentives.

For individual readers who use cloud services or apps: check whether a provider offers carbon‑aware options, prefer on‑device features for simple tasks, and pause or batch heavy personal workloads where possible. Public pressure and consumer demand for transparent energy reporting will make cleaner options more attractive and cheaper over time.

AI power demand is real, measurable and concentrated in a few places: data centres, networks and occasional, energy‑intensive training runs. Broad estimates place global data‑infrastructure electricity use in the low hundreds of terawatt‑hours per year, while single large trainings use hundreds to a few thousand megawatt‑hours. Those numbers show that AI is not a dominant share of global electricity today, but it is large enough that choices about location, grid mix, procurement and algorithmic efficiency matter.

Improved transparency — publishing kWh, PUE and the carbon intensity of power used — combined with continued work on more efficient algorithms and a shift towards carbon‑aware operations, will determine whether the growth of AI leads to rising emissions or mostly cleaner computing. The key point is practical: governance, engineering and consumer choices can bend the curve without stopping innovation.