The share of energy devoted to inference in the lifecycle of large AI models has progressed significantly in recent years, to the point where it now exceeds that of training. In other words, running AI now costs more energy than building it. And this reversal occurs precisely at the moment when each individual request becomes less resource-intensive.
This paradox is not a bug. It is the heart of the energy debate around AI, and it explains why two camps clash using the same figures to say the opposite. Optimists show that efficiency per task improves by at least an order of magnitude per year. Alarmists respond that the sector’s total consumption advances sharply each year — the International Energy Agency estimates this growth at approximately 17% for all data centers in 2025, and approximately 50% for data centers focused on AI for that same year. They are all right. They measure different things.
The Essentials
- Total energy consumption related to AI progresses steadily each year, according to the IEA, even though efficiency per task improves by at least an order of magnitude each year.
- The share of lifecycle energy for frontier models consumed at inference is now dominant and in strong progression, having significantly exceeded that of training in recent years.
- “Reasoning” queries can consume hundreds or even thousands of times more energy than simple text generation tasks according to the IEA, making it a potential governance lever.
- The rebound effect exceeds efficiency gains because deployment accelerates faster than optimizations, a phenomenon well documented since Stanley Jevons’s work on coal consumption in the nineteenth century.
- Confusion between efficiency per query and total consumption disarms regulators, who struggle to choose the right indicator.
When Efficiency Gains Do Not Prevent Rising Costs
The history of technology energy consumption is a long series of paradoxes of the same type. Watt’s steam engines consumed less coal per unit of horsepower produced than those of Newcomen. Result: many more were installed, and coal consumption exploded. Stanley Jevons describes this mechanism in 1865. One hundred sixty years later, it applies to data centers perfectly.
The IEA documents the phenomenon with precision in its report Energy and AI published in April 2025. The energy efficiency per task of AI models improves at a remarkable pace, estimated at at least an order of magnitude per year. A model that performs an image classification task today consumes ten to one hundred times less energy than a comparable model from two years ago for equivalent results. Teams from Google DeepMind, Anthropic, and OpenAI regularly publish benchmarks that confirm this trajectory.
But usage grows faster than efficiency progresses. Each efficiency gain makes AI cheaper to deploy, which attracts new uses, new users, new applications. The overall envelope inflates. This is exactly Jevons’s paradox: making a resource more efficient increases its total consumption because usage spirals. At the growth rate of approximately 15% per year taken as the baseline case by the IEA for data centers through 2030, the sector’s total consumption would double in approximately five years.
The Silent Reversal: Inference Surpasses Training
Until 2023, the energy debate around AI focused on model training. GPT-3, Llama, Gemini: each new frontier model mobilized thousands of GPUs for weeks and consumed quantities of energy that made headlines. It was the visible moment, the costly and well-defined sprint.
Inference, meanwhile, was perceived as cheap. Querying an already-trained model seemed negligible compared to the cost of its construction.
This calculation is outdated. The share of inference in total energy consumed over the lifecycle of a frontier model is now overwhelmingly dominant, and this shift occurred in just a few years. It is explained by simple arithmetic: a model trains once, then responds to billions of queries over months or years. As deployment scales up, the inference phase ultimately always dwarfs the training phase in total energy balance.
This reversal has an immediate practical consequence: optimizing training is now a secondary variable in AI’s carbon balance. What matters is what happens with each query, each API call, each interaction. And it is here that the type of query becomes decisive.
Reasoning Changes the Physics of the Problem
Not all AI models are alike from an energy perspective. A standard query, such as “translate this sentence” or “summarize this text,” mobilizes the model for a few milliseconds. A “reasoning” query, which asks the model to break down a complex problem, verify its own reasoning steps, and explore multiple hypotheses before responding, can last several seconds to several minutes and mobilize computational resources without comparison.
The IEA, in its report Key Questions on Energy and AI published in April 2026, quantifies the gap strikingly: reasoning and agentic tasks can consume hundreds or even thousands of times more energy per query than simple text generation tasks. OpenAI o1, Google Gemini Thinking, the latest versions of Anthropic’s Claude, are all built around this extended reasoning capability. They are also, precisely, the models whose professional and scientific uses are developing fastest.
The Stanford AI Index 2025 confirms this: large enterprises and research institutions are migrating massively to reasoning models for complex tasks, legal analysis, financial modeling, molecular discovery, code generation. The most useful applications are also the most energy-intensive.
This is where the governance variable becomes clearly visible. Unlike the consumption of an entire data center, difficult to regulate without slowing all activity, the type of query is an actionable lever. A policy that distinguishes reasoning uses from standard uses has a precise, measurable, and calibratable instrument. It is not yet in regulatory texts, neither in Europe nor elsewhere, but the technical parameters exist.
Confusion of Indicators Disarms Regulation
The political problem is here: for the past two years, public debate has oscillated between two narratives that do not speak to each other. On one side, AI advocates cite efficiency curves. On the other, critics cite total consumption projections. Both have solid figures. They measure orthogonal quantities.
This confusion is not insignificant. It paralyzes regulators. A decision-maker who does not distinguish efficiency per query from total consumption cannot choose the right indicator, and therefore cannot design the right policy. Should efficiency standards be imposed on models? Consumption caps on data centers? Labels by query type? Carbon reporting obligations by use case? All these options exist, but they do not act on the same variable.
The European Union, with the AI Act and ongoing discussions on the energy component of the Data Act, is beginning to pose the questions but without answering them with precision yet. American regulators, under pressure from major technology companies, remain in the background. China, whose data centers consume a growing share of national electricity, is piloting efficiency plans by data center without distinguishing types of uses.
The question of the carbon intensity of this consumption adds another layer of complexity. Offshore wind and renewables are progressing but at a pace that has not yet compensated for the increase in data center demand. In 2024, global data center electricity demand exceeded 400 TWh according to the IEA. By 2030, central projections place this demand between 700 and 1,000 TWh depending on the pace of AI deployment. The gap between these scenarios depends essentially on the governance policies adopted in the next three years.
The Actors Working on the Problem
The idea that the industry is ignoring the question is inaccurate. Several dynamics are underway, driven by different actors.
Google has published its energy consumption data by type of operation and committed to a trajectory of 24/7 carbon-free energy supply by 2030, which means no longer an annual offset but hour-by-hour correspondence between consumption and clean production in each region. This is a considerable technical ambition. Microsoft has concluded nuclear energy supply agreements, notably with Constellation Energy for the recommissioning of Three Mile Island, precisely to meet the firm and permanent demand of data centers. Amazon Web Services is investing in small modular reactors for the same reasons.
On the models side, teams like those at Mistral AI in France or DeepSeek in China have shown that it is possible to build very competitive models with lighter architectures and less energy-intensive designs. DeepSeek-R1, released in early 2025, caused a moment of astonishment in the sector by achieving performance close to American models with a fraction of the compute. This is a practical demonstration that efficiency is not capped.
The IEA and several academic laboratories, including Emma Strubell’s group at Carnegie Mellon, are working on standardized methods for measuring the energy footprint of models. Without measurement, no regulation is possible. This data infrastructure is a prerequisite for any serious framework.
AI Could Also Accelerate the Transition It Complicates
It would be incomplete to treat AI solely as an energy problem. It is also, potentially, a tool for solving part of the problems it contributes to creating.
Grid optimization models developed by companies like DeepMind (now Google DeepMind) have shown gains of 10 to 15% in managing data centers themselves. Next-generation weather forecasting models, such as GraphCast, enable better integration of intermittent energy sources into networks. Material discovery tools accelerate research on new-generation batteries and solar panels.
These applications are real and measurable. They do not mechanically offset the energy cost of AI, and no one yet has reliable consolidated accounting on the net balance. But they indicate that the relationship between AI and energy transition is more complex than a simple consumption equation. Growing skills around AI in sectors like energy and engineering also accelerates the diffusion of these tools in contexts where they can have a multiplier effect.
The open question remains one of timing. The climate benefits of AI are mostly potentials to be realized over ten to twenty years. The energy cost is real and growing today. If data centers are powered by gas or coal over the next five years while waiting for renewables and nuclear to ramp up, the intermediate balance is negative. The pace of the power sector’s decarbonization is not a fixed variable: it depends on national energy policies, some of which remain poorly aligned with the demand trajectories that AI creates.
What Regulators Can Do Now
None of the problems identified here is without solution. The constraint is real; it is not insurmountable.
A first lever is standardization of measurement. Without a common standard for declaring energy consumption by query type, companies publish incomparable figures. The IEA, IEEE, and European standardization bodies have all proposed frameworks. Getting them to converge toward a single format is a coordination task, not a technological breakthrough.
A second lever is transparency by use case. Requiring major AI service providers to declare the consumption of their APIs by category, standard query versus reasoning, would allow client companies to manage their own footprint. This is the same principle as carbon reporting for supply chains, extended to computing consumption.
A third lever is alignment of price signals. Today, the cost of a reasoning query for the user partially reflects the cost of computation, but not the full energy cost nor its carbon intensity. Pricing that would integrate these externalities, in the manner of what carbon tax attempts to do in other sectors, would incentivize more precise allocation between uses that merit extended reasoning and those that do not.
None of this slows down AI. It directs it. The distinction that the IEA makes between efficiency per query and total consumption is not merely an exercise in conceptual clarity. It is the map without which any policy remains blind.
Sources
- IEA — Energy and AI (April 2025): https://www.iea.org/reports/energy-and-ai/executive-summary
- IEA — Key Questions on Energy and AI (April 2026): https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary
- Stanford University — AI Index Report 2025: https://aiindex.stanford.edu/report/
- Stanford HAI — AI Index 2025: https://hai.stanford.edu/ai-index/2025-ai-index-report
- Stanley Jevons, The Coal Question (1865) — historical reference, public domain
- Wikipedia — Jevons Paradox: https://en.wikipedia.org/wiki/Jevons_paradox
- White & Case — Energy Efficiency Requirements under the EU AI Act: https://www.whitecase.com/insight-alert/energy-efficiency-requirements-under-eu-ai-act
- Luccioni et al. (2025) — Jevons Paradox and AI — ACM FAccT: https://arxiv.org/abs/2501.16548