Data · AI · Law · Ethics

The magazine for decision-makers

·

How much energy does AI consume? Numbers, costs and what companies can do

Data centers could more than double their electricity consumption by 2030, driven by AI. What the numbers say, how much a single prompt weighs and which levers companies have to cut costs and energy.

by

·

·

4–6 minutes
Traliccio dell'alta tensione in silhouette contro un cielo rosso al tramonto

Key points

  • The big picture: according to the IEA, data centers reached around 415 TWh in 2024 and could reach around 945 TWh in 2030, mainly because of AI.
  • A single prompt consumes little (fractions of a watt-hour), but the volumes are enormous and growing.
  • For companies, energy and cost are one and the same: choosing the right model and measuring the cost per task is already a sustainability policy.

Every time we ask an AI assistant something, an accelerator switches on somewhere. Taken on its own, the consumption is minimal. Multiplied by billions of requests a day, it becomes a matter of energy policy, corporate costs and ethics. AI energy is no longer just an issue for environmentalists: it concerns CFOs, CIOs and those who plan electricity grids.

The IEA numbers

In its April 2025 Energy and AI report, the International Energy Agency estimated that data centers consumed around 415 TWh in 2024, roughly 1.5% of global electricity. In the base case scenario, this rises to around 945 TWh in 2030, growth driven above all by the accelerated computing used for AI. In April 2026, the IEA reported that consumption continued to grow strongly in 2025.

GLOBAL DATA CENTER ELECTRICITY CONSUMPTION 4008000 415 2024 ~945 2030 (estimate) × 2.3 TWh Source: IEA, Energy and AI (April 2025), base case scenario. Chart by QrmServizi.
In six years, data center consumption could more than double.

How much a single prompt consumes

For a long time there was no official data. Since 2025, some providers have begun publishing estimates:

  • Google (August 2025): a median text prompt on Gemini consumes around 0.24 Wh, emits 0.03 g of CO₂ equivalent and uses 0.26 ml of water. According to Google, energy per prompt fell 33-fold in twelve months.
  • OpenAI: CEO Sam Altman wrote in June 2025 that an average query uses around 0.34 Wh. This is a statement, not an independently verified measurement.
  • Mistral (July 2025) published a life-cycle analysis of Mistral Large 2, carried out with Carbone 4 and the French agency ADEME: around 1.14 g of CO₂ equivalent and 45 ml of water for a 400-token response, also accounting for the share of training and infrastructure.
ENERGY PER SINGLE REQUEST (WATT-HOURS) 10 W LED bulb for 1 minute 0.17 Gemini, median promptGoogle, August 2025 0.24 ChatGPT, average queryOpenAI statement, June 2025 0.34 For comparison: a full smartphone charge requires around 15-20 Wh, the equivalent of 60-80 prompts at 0.24 Wh. Provider methodologies are not uniform. Chart by QrmServizi.
A single prompt weighs little: the problem is scale.

Two caveats. First, the methodologies are not comparable with one another, and models that “reason” at length or generate images and video consume far more than a text prompt. Second, efficiency per prompt is improving, but usage is growing faster. It is a classic rebound effect.

What Europe is doing

The EU Energy Efficiency Directive already requires data centers to report their energy performance indicators, under Delegated Regulation (EU) 2024/1364. On September 21, 2026, the Commission proposed a common rating scheme for data centers and opened a consultation on minimum performance standards, which remains open until December 14, 2026. The legislative proposal is expected in the second quarter of 2027.

The paradox is clear: the EU wants to multiply its computing capacity so as not to depend on others and, at the same time, to contain consumption and emissions. In Italy, where electricity costs more than in many European countries, grid connections for new data centers have become an industrial issue.

Five levers for companies

The good news is that, for those who use AI, cutting energy and cutting the bill are almost always the same thing.

1 2 3 4 5 Right modelsmall where it suffices Cache and reusedon’t redo the work Batchgroup requests together Where and whenregions with clean energy Measurecost per task QrmServizi diagram. In most cases, less computing means both less energy and lower cost.
Energy efficiency and economic efficiency point in the same direction.
  1. Use the smallest model that does the job. Classifying an email does not require the most powerful model on the market. A routing system can send only the difficult requests to large models.
  2. Reuse. Caching frequent prompts and responses avoids repeated computation.
  3. Batch. Non-urgent processing in batches costs less and makes better use of the hardware.
  4. Choose where and when. The cloud region and the time of day affect the carbon intensity of the energy used.
  5. Measure the cost per task. Not the cost of the model, but how much it costs to handle a case, resolve a ticket or analyze a document. This is the metric that brings FinOps and sustainability together.

The ethical dimension

Using AI is not wasteful in itself: if it replaces more energy-intensive processes or improves important decisions, the balance can be positive. The problem is indiscriminate use: the huge model used for everything, the image generated “just to try”. Provider transparency, measurement and frugal design are the three pillars of AI that is responsible from an energy standpoint too.


The compass

AI energy is not governed by proclamations, but by design choices. Those who measure the cost of every task and use the right model for each job save money and reduce their impact at the same time.

Sources

Leave a Reply

Your email address will not be published. Required fields are marked *