In this article
I write "cloud" and my head hands me back vapor. It's the trap of the word, and anyone falls for it. Behind every query that seems to weigh nothing there are hectares of refrigerated corridors, diesel backup generators and towers exhaling evaporated water over a real landscape. The International Energy Agency put the world's electricity consumption from data centers at around 415 TWh in 2024, close to 1.5% of the planet's total, and projects some 945 TWh for 2030. The footprint of asking a chatbot something exists, it's measured and it's going to multiply. What bothers me isn't the figure. It's that the figure doesn't appear on any bill the user recognizes as their own.
The physical figures
It's worth fixing the magnitudes before arguing about anything, because the energy debate over AI is almost always waged with no numbers on the table, and that favors whoever has something to hide.
The global electricity consumption of data centers was around 415 TWh in 2024, roughly 1.5% of world electricity consumption, according to the Energy and AI report from the International Energy Agency (April 2025). That same report puts demand at around 945 TWh for 2030. The IEA doesn't split the total between "AI" and "the rest" with a closed percentage —which is why I distrust anyone who does— but it does distinguish rates: the accelerated servers that sustain AI grow at around 30% a year, against 9% for conventional servers. The slice of the pie AI eats widens year after year. Any exact percentage at the end of the decade, written by whoever writes it, is a projection and not a number carved in stone.
The capacity installed specifically for AI reached some 29.6 GW at the end of 2025, according to Stanford HAI's AI Index Report 2026, equivalent to the electricity consumption of a state like New York at its demand peak. And according to the same report that computing capacity has more than tripled every year since 2022. Multiplying by more than three every twelve months, sustained over years, is a regime for which the everyday language of bills has no word: not "rose," not "grew," not "shot up" give the measure.
The cost of training a leading model is counted in tens or hundreds of millions of dollars of compute, according to Epoch AI's estimates for 2024-2025. The energy footprint of training GPT-4 circulates as a figure on the order of 50 GWh, but I haven't found any source that signs it with rigor, so I leave it where I found it: as a number repeated by ear, not as data I could defend. What is documented is the raw scale of a real training run, because the company that paid for it published it. Meta trained Llama 3.1 405B on a cluster of around 16,000 H100 GPUs.
Individual inference, by contrast, seems ridiculous, and that's where the misunderstanding begins. A query to ChatGPT consumes on the order of 0.34 Wh of electricity, a figure Sam Altman himself gave in 2025 and which fits estimates derived from Google searches. Generating an image runs more expensive, around 2.9 Wh, according to the work of Sasha Luccioni and colleagues in Power Hungry Processing (2023). A single query frightens nobody, and it's right not to. Multiplied by the billions launched every day, it stops being an anecdote and becomes heavy infrastructure.
Two curves that chase each other without catching up
The IEA's projection describes two curves advancing at incompatible speeds, and the mismatch between them is the real problem, not either one on its own.
The demand curve runs. No need to speculate: it's enough to add up what the companies themselves put in writing in a single quarter. Microsoft announced on 3 January 2025, through Brad Smith, a forecast of 80 billion dollars in AI data centers for its fiscal year 2025. Stargate —OpenAI's venture with SoftBank, Oracle and MGX, presented on 21 January 2025— set a goal of investing half a trillion dollars in infrastructure over four years. Meta, in its late-January 2025 results report, put its planned annual capital investment between 60 and 65 billion. Three announcements, one month, and the tally already passes half a trillion before adding Google's or Amazon's.
The electrical capacity curve drags its feet, and it drags them for physical reasons, not negligence. A new plant —nuclear, wind, solar with storage or gas— takes between five and fifteen years from the decision to the delivery of its first megawatt. The regions hoping to host large data centers already hit the ceiling: northern Virginia, Texas, Ohio, Ireland, Aragon, Singapore. In Virginia, Dominion Energy went so far in 2024 as to impose waits of up to seven years for new connections in the most saturated zones, though it later resumed part of them as soon as it could. In Ireland, the regulator CRU, together with grid operator EirGrid, has maintained a de facto moratorium on new data centers in the Dublin area since 2021.
Money moves at the speed of a press release. Copper, concrete and permits, at the speed of a civil works project with neighbors who object. That difference in pace isn't a scheduling setback to be solved by waiting: it's the structural reason why the investments committed for 2026-2030 comfortably exceed the electricity the grids will be able to supply with the infrastructure that's planned to be built.
Where the energy is really burned
The spending doesn't concentrate in a single point that could be blamed, which makes it convenient to hold no one responsible. It's spread along a chain, and it's worth walking the whole thing, because the computation itself is only one stretch of the balance.
Training is intense and one-off: weeks or months with thousands of GPUs operating at the limit, an enormous, concentrated consumption peak that each generation of model repeats with a bigger scaling factor than the last. It's the easy headline, the one that illustrates the reports with photos of cold rooms. It isn't, however, what weighs most when you add up the time.
What really accumulates is inference. Every query from every user pulls on electricity, and a model deployed at mass-consumption scale gets queried without pause for months. I don't have a closed accounting that signs the comparison company by company, but the arithmetic points in a single direction: a one-off cost, however enormous, is eventually reached and overtaken by a small cost repeated relentlessly millions of times a day. If training is the bill for putting up the building, inference is the rent paid every day, long after the report about the construction has stopped interesting anyone.
And there's cooling, which doesn't appear in the speeches about algorithmic efficiency because there's nothing elegant about boasting of air conditioning. Dissipating the heat costs electricity when it's done by air, and water when it's done by evaporation. Google's 2024 Environmental Report declared a total water consumption of some 24 billion liters, attributable in the great majority —around 95%— to its data centers. Microsoft, in its own environmental data, recorded a 34% growth in its water consumption between 2021 and 2022, blamed largely on AI workloads. Water doesn't evaporate in a comfortable metaphor. It evaporates in a specific tower, in a specific town, over an aquifer that specific people drink from.
The asymmetry someone designed on purpose
The chatbot seems ethereal, and not by accident. The interface of ChatGPT, Claude or Gemini is text on a white background, without the slightest trace of its physical substrate. The data center that holds it up, by contrast, is heavy industry of the dirtiest kind: kilometers of cable, diesel generators awaiting the next blackout, towers releasing plumes of vapor over the countryside.
Two imaginaries that don't fit coexist in the user's head. The product's suggests lightness; the infrastructure's, the dead weight of concrete. Whoever designed the experience worked precisely so the second never showed while the first is in use, and they succeeded: nobody typing a question thinks of the gas that burns or the water pumped so they get answered in two seconds.
This has a consequence that's political before it's aesthetic. If the user doesn't perceive the cost, they don't factor it into the decision to use or not use. The sensitivity of consumption to its real cost stays at zero, not because the user is insensitive, but because that cost has been carefully kept out of sight.
The footprint that never reaches your bill
Here I put my own reading, even if it leans on national accounts that are only just beginning to register the phenomenon. The energy bill for asking the chatbot doesn't evaporate with the answer. It changes address.
Part of it is borne by the company that hosts the model, which pays it and passes it on where it can: in the price of the service, in the advertising it serves you, in a subscription that goes up a few euros. So far the market does what's expected of it. The problem starts afterward. The host country absorbs the grid strains, the costs of expanding the infrastructure, the water that comes out of its aquifer and the emissions not charged to any specific customer; that portion nobody signs for. And there's what gets pushed toward the future: the CO₂ of the fuels that burn today to feed a grid that will still be running when the user of this query is no longer around to argue whether it was worth it.
If each answer arrived with its spending label stuck to its side, I suspect something in behavior would change. The invisibility, however, isn't a design oversight. It is the design. The aggregate rises query by query precisely because nobody notices the weight of their own.
Thermodynamics doesn't sit down to negotiate
The debate over the sustainability of AI tends to split into two opposing stories, and neither of the two is entirely honest when left on its own.
There's the story that AI accelerates decarbonization: it optimizes grids, anticipates demand, designs materials, helps look for drugs. It has partial basis, and it's worth acknowledging so as not to fall into the opposite caricature. DeepMind reported in 2016 a reduction of close to 40% in the cooling energy of Google's centers by applying its models, equivalent to an improvement of around 15% in the overall efficiency indicator, and in 2018 it handed autonomous control of the system to those algorithms. There are grids already operating with AI-backed optimizers. The indirect benefits exist and I have no interest in denying them.
And there's the opposite story, that AI is an energy disaster because its consumption grows faster than any indirect benefit that could offset it. This one has a firmer basis, and the uncomfortable part is that the companies themselves supply it in their reports. Microsoft acknowledged a 29.1% increase in its total emissions relative to its 2020 base year. Google reported a 48% rise since 2019. Both maintain public commitments to neutrality while their absolute emissions keep climbing. It's not a contradiction they hide in a footnote: they publish it themselves, with their logo on top.
Looked at head-on, the discussion is less moral than physical, and the laws of thermodynamics don't report to the marketing department. The two curves cross or don't according to measurable numbers, and according to the IEA's, Stanford HAI's and the regional electricity regulators', they cross. What's left open isn't the crossing, but what a society does when it arrives: limit the use, expand capacity at forced marches, take the consequences or shift the cost toward whoever notices it least. None of those exits is painless, and the one finally chosen won't be dictated by physics, but by whoever holds the upper hand when the time comes to decide.
The arithmetic of the climate, meanwhile, doesn't ease off on its own. The net-zero doctrine the IEA defends calls for drastic cuts to the power sector this decade to stay on the 1.5 °C path —I don't have a single, well-closed figure for that cut at hand, so I leave it as a general requirement and not an exact percentage—. The growth in electricity consumption driven by AI and data centers pushes, in exactly these years, in the opposite direction to what that path demands.
Definitions
TWh: terawatt-hour, a unit of energy equivalent to a trillion (10¹²) watt-hours. It serves to measure the consumption of countries or entire sectors.
Inference: the execution of an already-trained model on new data to produce an answer. Unlike training, it scales with the number of users and is continuous.
PUE (Power Usage Effectiveness): the ratio between the total energy a data center consumes and the energy consumed by computation alone. A PUE of 1.0 would be the unreachable ideal; modern centers move between 1.2 and 1.5.
WUE (Water Usage Effectiveness): liters of water consumed per kWh of computation. The metric analogous to PUE for the water footprint.
H100 GPU: Nvidia's graphics processor designed for training and inference of large models, the usual unit of account for AI computation.
References
IEA — Energy and AI (April 2025). Source of the figures for data center consumption (≈415 TWh in 2024), the projection to 945 TWh for 2030 and the growth rates of accelerated servers (≈30% a year) against conventional ones (≈9%). Executive summary at https://www.iea.org/reports/energy-and-ai/executive-summary
Stanford HAI — AI Index Report 2026 (April 2026). Source of the figure for installed capacity for AI (≈29.6 GW at the end of 2025, equivalent to New York's demand peak) and of its more-than-tripling per year since 2022. https://hai.stanford.edu/ai-index/2026-ai-index-report
Epoch AI — Compute Trends Across Three Eras of Machine Learning and 2024-2025 updates. Source of the estimates of compute cost of the frontier models.
Meta AI — Introducing Llama 3.1 (2024). Source of the scale of the Llama 3.1 405B training (a cluster of around 16,000 H100 GPUs). https://ai.meta.com/blog/meta-llama-3-1/
Sam Altman — The Gentle Singularity (2025), reported by DataCenterDynamics. Source of the figure of ≈0.34 Wh per ChatGPT query. https://www.datacenterdynamics.com/en/news/sam-altman-chatgpt-queries-consume-034-watt-hours-of-electricity-and-0000085-gallons-of-water/
Luccioni, S., Jernite, Y. & Strubell, E. — Power Hungry Processing (2023). Source of the estimated consumption per image generation (≈2.9 Wh). https://huggingface.co/papers/2311.16863
Google — Environmental Report 2024. Source of the total water consumption (≈24 billion liters, around 95% attributable to data centers) and of the 48% rise in emissions since 2019. https://blog.google/company-news/outreach-and-initiatives/sustainability/2024-environmental-report/
Microsoft — Environmental Sustainability Report 2024, with coverage from ITPro. Source of the 29.1% increase in total emissions relative to the 2020 base year. https://www.itpro.com/infrastructure/data-centres/microsofts-ai-fueled-data-center-rush-caused-carbon-emissions-to-surge-by-29-in-2023
Microsoft — water footprint data 2021-2022, reported by The Register (September 2023). Source of the 34% increase in the company's water consumption between 2021 and 2022, attributed largely to AI workloads. https://www.theregister.com/2023/09/11/microsofts_ai_investments_skyrocketed_in/
OpenAI — Announcing The Stargate Project (21 January 2025). Source of the half-trillion-dollar four-year investment goal of OpenAI's venture with SoftBank, Oracle and MGX. https://openai.com/index/announcing-the-stargate-project/
Microsoft — The Golden Opportunity for American AI (Brad Smith, 3 January 2025) and coverage from DataCenterKnowledge. Source of the forecast of 80 billion dollars in AI data centers for fiscal year 2025. https://www.datacenterknowledge.com/data-center-construction/microsoft-to-spend-80-on-ai-data-centers-this-years
Meta — Fourth-quarter and full-year 2024 results (late January 2025), with coverage from Bloomberg. Source of the forecast of capital investment between 60 and 65 billion dollars for 2025. https://www.bloomberg.com/news/articles/2025-01-24/zuckerberg-warns-of-higher-than-expected-capex-at-meta-in-2025
DeepMind — DeepMind AI Reduces Google Data Centre Cooling Bill by 40% (2016). Source of the ≈40% reduction in cooling energy (≈15% in overall PUE) and of the handover of autonomous control of the system in 2018. https://deepmind.google/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-by-40/
Dominion Energy — Public statements and coverage on data center connection limits in Virginia (2024); waits of up to seven years in the most saturated zones, partially resumed afterward. CRU / EirGrid — De facto moratorium on new data centers in the Dublin area since 2021. https://www.datacenterfrontier.com/energy/article/11436951/dominion-resumes-new-connections-but-loudoun-faces-lengthy-power-constraints
IEA — Net Zero Emissions by 2050 Scenario. General framework of the 1.5 °C path that demands drastic cuts to the power sector this decade. https://www.iea.org/reports/global-energy-and-climate-model/net-zero-emissions-by-2050-scenario-nze
Further reading
Crawford, K. — Atlas of AI (Yale UP, 2021). A material map of AI: mines, water, labor.
Hogan, M. — Big Data Ecologies: Thinking Critically about Data Infrastructures (Routledge, 2023).
Strubell, E., Ganesh, A. & McCallum, A. — Energy and Policy Considerations for Deep Learning in NLP (ACL, 2019).
Greenpeace — Clicking Clean (2017 and 2019 reports). A chronology of the activism on clean data centers.
Comments0
No comments yet.
Leave a comment