Data Center Consumption: Problems and Solutions
Training a large language model consumes as much electricity as thousands of households use in a year. Cooling the servers that run it evaporates millions of liters of water. And the companies building them are signing contracts with nuclear plants to make sure the lights don’t go out.
The International Energy Agency (IEA) compiles electricity figures in its report Key Questions on Energy and AI, and they’ve been growing at a steady pace for several years now. In 2025, data centers’ electricity consumption grew 17%, while global electricity demand grew 3% [1]. The sector that underpins the internet and artificial intelligence is taking an ever-larger share of the world’s electricity.
Remember that on Pensar es Gratis we only cite reliable data and sources that pass our content filter. At the end of each post there’s a practical guide on the topic if it makes sense for one to exist.
The outsized spending of AI data centers is mainly due to either training or inference. Training a model is the process of showing it millions of examples so it learns patterns. Inference is what happens every time the model responds to a query: individually it consumes little, but multiplied by millions of users a day it ends up weighing as much as, or more than, training.
PUE (Power Usage Effectiveness) is a ratio that measures how much of a data center’s electricity actually goes toward computing versus how much is lost to cooling, lighting, and other auxiliary systems. A PUE of 1.2 means that for every watt spent on computing, 0.2 more is spent keeping the building running.
How much energy it takes to train an AI
The energy consumption of a training run is calculated by multiplying the number of chips by the training hours and by the power draw of each chip, adjusted for the data center’s PUE. OpenAI has never published GPT-4’s or GPT-5’s consumption, so the reliable figures come from earlier or open models.
GPT-3, from 2020, consumed about 1,287 MWh and emitted 552 tonnes of CO₂ equivalent, according to an analysis by Google and UC Berkeley [3]. Meta’s Llama 3.1 405B needed 30.84 million H100 GPU-hours at a reference power of 700 watts per chip [4]. Multiplying that out gives roughly 21.6 GWh — seventeen times more than GPT-3, and equivalent to the annual consumption of about 5,900 Spanish households, assuming 10 kWh per household per day.
The trend points upward. According to the Epoch AI institute, the power required to train frontier models has doubled every year over the past decade. Frontier training runs already exceed 100 MW, and the largest single training run of 2030 could require between 4 and 16 GW [6].
Not all training costs the same. Pretraining is the phase in which the model reads trillions of words and learns to predict the next one; it’s the phase that consumes almost all the energy. After that comes post-training, where the model is taught to follow instructions using examples and to improve its responses through rewards. In DeepSeek-V3, pretraining cost 2.664 million GPU-hours, and all the later phases combined cost 0.1 million [7]. Fine-tuning, which adapts an already-built model to a specific task, is cheaper still, since it starts from work already done.
As for inference, the cost per query varies enormously. According to Stanford’s AI Index 2026, DeepSeek V3 consumes about 23 Wh for a query of average length, versus about 5 Wh for Claude 4 Opus [8]. Google has measured 0.24 Wh per median text query on Gemini — less than nine seconds of watching television [9]. A 2026 study by Microsoft researchers published in the journal Joule gives a median of 0.31 Wh for models with more than 200 billion parameters on standard queries, rising to 3.91 Wh for long, reasoning-heavy responses [10]. The range depends mostly on response length, along with model size and hardware. An interesting detail: Google’s last official figure for a normal search dates from 2009 and was 0.3 Wh. By that old number, asking Gemini a question today uses less energy than a Google search did seventeen years ago.
Training and the first 18 months of use of Mistral Large 2 consumed about 281,000 cubic meters of water, which fits into 112 Olympic swimming pools. This is a full life-cycle analysis that also includes hardware manufacturing, and training plus inference account for 91% of that water [12].
The water that evaporates so AI doesn’t overheat
Evaporative cooling is one of the most widespread methods in data centers. It works well and is cheap, but the heat from the servers evaporates the water, and that water doesn’t come back. In 2025, according to the consultancy Rystad Energy, data centers worldwide consumed about 222 billion liters of water for cooling alone. Without efficiency improvements, that figure could approach 644 billion liters by 2030 [13].
The link between water and energy is direct: reducing water consumption usually means spending more electricity, because the heat that used to be dissipated by evaporating water now has to be extracted with pumps and fans. Rystad quantifies it: dry cooling saves about 2.15 liters of water per kWh of computing load, but adds between 0.30 and 0.74 kWh of electricity.
Nvidia announced at CES 2026 that its Vera Rubin platform can be cooled with water at 45°C, which allows it to do without chillers [14]. According to Jensen Huang, that could save around 6% of the world’s data center electricity [15]. The water temperature is the same as in the previous generation, Grace Blackwell, so the novelty isn’t the thermal jump but that Vera Rubin dissipates twice the power with the same flow rate and the same inlet temperature. Even so, the heat still has to leave the building: without chillers, a system to evacuate it outside is still needed.
Water-efficiency figures vary a lot by operator and location. AWS reported 0.12 liters per kWh of computing load in 2025, and Equinix 0.91. Within AWS, the range runs from 0.02 l/kWh in Stockholm to 2.85 l/kWh in Jakarta [13]. On top of that comes the water used by the power plants generating the electricity: in the United States that can exceed double the data centers’ direct consumption.
Tech giants turn to nuclear power
A data center runs 24 hours a day, 7 days a week, and solar and wind alone aren’t enough. The intermittency of renewables requires either fossil backup or massive storage, and neither option is cheap or quick to deploy. That’s why big tech companies are signing deals with the nuclear sector.
Constellation Energy will restart Unit 1 of Three Mile Island, in Pennsylvania, and Microsoft will buy its entire output, 835 MW, for 20 years, with first electricity expected in the second half of 2027. Google has signed with Kairos Power for 500 MW of small modular reactors (SMRs): the first is slated for 2030 and the full total for 2035. Amazon already receives electricity from the Susquehanna plant and invested $700 million in X-energy [16]. Meta announced deals in January 2026 with Oklo, Vistra, and TerraPower for up to 6.6 GW by 2035 [17]. By the end of 2025, the portfolio of deals between data centers and SMR manufacturers totaled 45 GW, according to the IEA [1].
SMRs are reactors of up to 300 MW, designed for series production. They’re built in modules and assembled on site, which in theory reduces costs and construction timelines. The first SMRs committed to tech companies won’t deliver electricity before 2030, while data center demand is growing at 15% a year [2].
Efficiency, reused heat, and leaner models
Efficiency per task is improving very fast. Google cut the energy of its median query by a factor of 33 in just twelve months [9]. The IEA estimates that energy per AI task has fallen by at least an order of magnitude per year in recent years [1].
Another path is designing models that use less. DeepSeek-V3 has 671 billion parameters but only activates 37 billion for each word it processes. It was also trained using 8-bit numbers instead of 16-bit, and its full training run cost 2.788 million GPU-hours [7] — roughly eleven times less than Llama 3.1 405B for comparable performance.
The third path is to stop throwing away the heat. A team from the University of Florence simulated integrating a data center’s heat into the gas-fed district heating network of Calenzano. Depending on the data center’s size, between 100 and 500 kW, the waste heat would cover between 12% and 59% of the network’s demand, and gas consumption would drop between 11% and 58%. A 100 kW center without heat integration releases about 926 MWh of heat into the environment per year [18]. In Brescia, a real-world case is already running: the French company Qarnot installed 30 liquid-cooled computing units at an A2A plant that recover heat up to 65°C and feed it directly into the urban network — about 800 MWh of heat a year [19].
What’s happening outside the United States
In 2024, the United States accounted for 45% of the world’s data center electricity consumption, China 25%, and Europe 15% [2]. According to China’s National Energy Administration, the country’s computing centers consumed 170 billion kWh in 2025, 1.6% of its electricity. The official forecast is to reach 800 billion kWh by 2030, around 6% [20]. The eight major hubs of the national computing network increased their consumption by an average of 39.5% annually over the past three years. In Inner Mongolia the pace was 66.5%, because the load is shifting toward energy-rich regions. The new centers in those hubs aim for more than 80% of their electricity to be renewable.
Ireland is Europe’s most extreme case. Data centers went from 5% of the country’s electricity in 2015 to 23% in 2025, at 7,663 GWh, 10% more than the previous year. All of the country’s homes combined consumed 28%. By 2023, data centers had already overtaken all of Ireland’s urban households combined [21]. For several years there was a de facto freeze on new grid connections. It has since been lifted, and new centers must now cover at least 80% of their annual demand with new renewable electricity [22].
The Nordic countries are taking a different route. In Finland, Fortum is building heat pumps alongside Microsoft’s data centers in Espoo and Kirkkonummi. Once running at full capacity, they will cover around 40% of the district heating for Espoo, Kauniainen, and Kirkkonummi, and will avoid about 400,000 tonnes of CO₂ a year [23].
My own view is that, if the current trend holds, it seems likely that within five to ten years data centers will stop being passive electricity consumers and become flexible energy assets: modulating their load according to the availability of renewables on the grid and selling grid-balancing services back to the system. There are already pilot projects moving in that direction, and the European regulatory framework is starting to recognize demand flexibility as a resource. I think a data center of the future could make money not just by processing data, but also by helping stabilize the grid.
On another note, it seems clear that waste-heat reuse could become as important a siting criterion as fiber optics or land availability. If a data center’s heat can warm homes at a competitive cost, cities will have an incentive to attract these facilities rather than reject them. Qarnot is already doing this in Brescia, and Fortum in Espoo. It wouldn’t surprise me to see city councils competing to host data centers for their ability to hand off heat to the urban grid.
References
[1] International Energy Agency (2026). Key Questions on Energy and AI. Paris: IEA.
[2] International Energy Agency (2025). Energy and AI. Paris: IEA.
[3] Patterson, D., Gonzalez, J., Le, Q., Liang, C., Munguia, L.-M., Rothchild, D., So, D., Texier, M. & Dean, J. (2021). Carbon Emissions and Large Neural Network Training. arXiv:2104.10350.
[4] Meta (2024). Llama 3.1 405B, Model Card, “Training Energy Use” section.
[5] Luccioni, A. S., Viguier, S. & Ligozat, A.-L. (2023). Estimating the Carbon Footprint of BLOOM, a 176B Parameter Language Model. Journal of Machine Learning Research, 24(253), 1-15.
[6] EPRI & Epoch AI (2025). Scaling Intelligence: The Exponential Growth of AI’s Power Needs.
[7] DeepSeek-AI (2024). DeepSeek-V3 Technical Report. arXiv:2412.19437.
[8] Stanford HAI (2026). AI Index Report 2026.
[9] Elsworth, C., Huang, K., Patterson, D. et al. (2025). Measuring the environmental impact of delivering AI at Google Scale. arXiv:2508.15734.
[10] Energy use of AI inference, efficiency pathways, and test-time scaling (2026). Joule.
[11] Li, P., Yang, J., Islam, M. A. & Ren, S. (2023, revised 2025). Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models. arXiv:2304.03271.
[12] Mistral AI (2025). Our contribution to a global environmental standard for AI.
[13] Rystad Energy (2026). Tech Thirst: Data center water consumption could triple by 2030 without efficiency gains.
[14] DatacenterDynamics (January 2026). Vera Rubin hot water cooling reveal triggers HVAC share drop.
[15] Fierce Network (January 2026). Supercomputers can stay chill with hot water says Nvidia.
[16] Axis Intelligence (2026). Nuclear Energy for Data Centers 2026.
[17] Utility Dive (January 9, 2026). Meta inks nuclear deals for up to 6.6 GW from Oklo, Vistra, TerraPower.
[18] Socci, L., Rocchetti, A., Verzino, A., Zini, A. & Talluri, L. (2024). Enhancing third-generation district heating networks with data centre waste heat recovery. Energy, 313, 134013.
[19] Il Sole 24 Ore. A2A inaugurates heat recovery data centre with Qarnot.
[20] Xinhua (May 27, 2026).
[21] Central Statistics Office, Ireland (2025). Data Centres Metered Electricity Consumption 2024.
[22] IrishCentral. Data centres consume nearly a quarter of Ireland’s electricity.
[23] Fortum (2023). Construction of Fortum’s heat pump plant has started on Microsoft’s data centre site in Kirkkonummi.
Notes on source reliability: References [3], [4], [7], [9], [12], [14] and [15] come from the companies whose own systems are being measured, without external validation, except for the consultancy review in [12]. [8], [13], [16], [19], [20], [22] and [23] are consultancy estimates, official data without published methodology, or journalistic roundups of announcements. [11] is an estimate based on average water-efficiency assumptions. [18] is a single-case simulation. [1], [6] and [10] are recent: they include current-year estimates or projections not yet corroborated.