Insights · Data centers
AI and energy: how much electricity an AI data center needs
Every query to an artificial intelligence model uses electricity and produces heat in a server somewhere. A single query uses little; multiplied by billions, and added to the training of ever larger models, it becomes one of the sources of electricity demand growth. This article separates fact from myth and explains where that power can come from.
Originally published on · Updated on
How much energy does a query use?
For years it was repeated that a ChatGPT query used about 3 Wh. A more recent estimate by Epoch AI (February 2025) puts a typical GPT-4o text query at around 0.3 Wh, an order of magnitude less.
The issue is not the query but the scale: billions of daily queries, model training and reasoning tasks that use far more than a simple question. And the hardware has changed: a system such as the NVIDIA GB200 NVL72 connects 72 Blackwell GPUs and 36 Grace CPUs in a single liquid-cooled rack, at power densities that force a rethink of both power supply and cooling.
The demand curve to 2030
According to the International Energy Agency (IEA), in its Energy and AI report (2025):
- Data centers used around 415 TWh in 2024, about 1.5% of the world’s electricity.
- Their consumption will more than double to around 945 TWh by 2030, slightly more than Japan’s total electricity consumption today.
- In the base case, accelerated-server electricity use grows 30% a year and provides almost half of the net increase.
In Mexico, the Ministry of Energy’s planning scenario expects gross electricity consumption to grow 2.5% a year between 2025 and 2039 (PLADESE 2025–2040). AI data centers add to that curve, concentrated in a few locations.
The heat problem
All the electricity that goes into GPUs comes out as heat, and removing it takes energy. The IEA estimates that cooling ranges from about 7% of consumption in an efficient hyperscale data center to over 30% in less efficient enterprise data centers. With high-density racks, cooling is one of the most reducible consumption items.
That is where trigeneration comes in: where there is gas-fired generation, engine heat drives absorption chillers that produce chilled water without using electricity for a compressor.
How that power is obtained
- Intense, continuous demand: on-site generation with solar and storage (BESS) as primary power, and grid and gas as backup. This is the Green Data Center by GIS Power model.
- Heat density: absorption chilled water using heat from gas-fired generation.
- Availability: N+1 or 2N redundancy, with the grid as an additional source.
- Time: if interconnection is slow, generation starts as bridge power.
- Growth: in phases as halls fill, with project finance per phase.
See also cogeneration for data centers: a technical guide.
The group’s experience
In 2013–2014 the group designed and built the first data center with cogeneration in Mexico for Alestra in Querétaro: 4.2 MW that supplied 100% of its power and chilled water. The group has built critical-power projects since 1983 and has developed 56+ MW of cogeneration in Mexico. See the Alestra case.
Sources
- International Energy Agency, Energy and AI (2025): data center consumption, accelerated servers and cooling.
- Epoch AI, “How much energy does ChatGPT use?”, February 7, 2025.
- NVIDIA: GB200 NVL72 specifications.
- SENER, PLADESE 2025–2040, DOF October 17, 2025.
- Group project history: Alestra Green Data Center.
Does your AI data center need firm power?
IT load, location and schedule: we design the scheme that powers and cools it.
By sending you accept our privacy notice.