Sustainable AI: from the race for power to the search for efficiency

Artificial intelligence is settling into companies at speed. Assistants, document search, content generation, task automation, data analysis, autonomous agents: the use cases keep multiplying, and executive committees no longer ask whether the subject concerns them.

That acceleration raises a question projects have long pushed to one side: how much resource do we actually need to get the result we are after?

It is no longer a marginal question. According to the International Energy Agency, global electricity consumption by data centres is expected to roughly double between 2025 and 2030, from around 485 TWh to 950 TWh. AI is one of the main drivers of that growth: consumption by AI-specialised centres rose by about 50 % in 2025 alone.

None of this argues for slowing adoption. It means efficiency now belongs among the design criteria of an AI system, alongside accuracy, cost, security and response time. It almost never features there today.

Not all AI is the same

Talking about "the consumption of AI" as a single quantity is misleading. The amount of computation depends on the model, its size, the infrastructure, the volume of data processed, the number of tokens exchanged — and above all on the nature of the problem being solved.

Classifying a document, finding information in a database, running a multi-step chain of reasoning or producing a video have nothing in common. The International Energy Agency notes that some recent uses — video generation, advanced reasoning, agentic tasks — can consume hundreds, even thousands of times more energy per request than a short text generation.

So the useful question is no longer "are we using AI?" but "how much AI are we mobilising to solve this particular problem?". That is an architectural reflex, and it can be learned.

The most powerful model is not always the best

Since large generative models arrived, a habit has taken hold: calling a highly capable model for tasks of very different natures. It is simple to implement, it saves you an arbitration, and it works. That does not make it rational.

A rules engine is sometimes enough. A conventional search often avoids a call to a large model. On a repetitive, well-defined task, a smaller specialised model gives more stable and faster results. An answer already computed can be served from a cache rather than regenerated identically. The most capable model then keeps its full place — for the cases that genuinely warrant it.

The aim is not to systematically pick the smallest model. It is to pick the sufficient model. The distinction has very concrete effects: on the inference bill, on response times, on the ability to scale without costs following the same curve, and incidentally on the resources consumed.

The stakes move from training to use

Early discussions about AI's footprint focused on training large models. It was the obvious angle: long, costly, spectacular campaigns. It is no longer a sufficient one.

Once deployed, a model is called millions of times, over years. For a company that consumes models rather than trains them, the essential part plays out in daily use: which model is called, how much context is sent to it, how long an answer is requested, how many searches are triggered, how often, and how many calls chain together for a single request.

Take an internal document assistant. Two teams can deliver an almost identical user experience and end up with systems that have nothing in common. The first sends the model every candidate document on every question, in a generous context, with a single powerful model. The second filters upstream, passes only the relevant passages, distinguishes factual questions from requests for analysis and reserves the heavy model for the latter. At comparable answer quality, their cost, latency and consumption profiles differ by an order of magnitude.

The architecture of the system matters as much as the choice of model. That is why the subject belongs to consulting and engineering, not to a licensing decision.

AI agents change the equation again

Agents shift the cursor once more, because they multiply the invisible operations behind a single request. A conversational assistant makes a few calls to the model. An agent analyses the request, builds a plan, searches for information, calls several tools, interprets their answers, revises its strategy and launches fresh operations. A question asked in one sentence can trigger several dozen technical actions.

That depth is precisely what makes agents interesting. It is also what makes their behaviour hard to predict — in cost as in duration. Hence the need to set bounds at design time: a maximum number of steps, a ceiling on calls, an allowed depth of reasoning, a model chosen according to the type of task, explicit stopping conditions, and a point of human validation wherever the decision commits the organisation.

The more autonomous the system becomes, the more you have to control what it consumes in order to produce a decision or an action.

These bounds are not environmental constraints in disguise. They are the same levers that hold cost, latency and reliability. An agent looping without a stopping condition is an operations problem before it is an energy problem.

Technical efficiency will not be enough

A well-documented mechanism deserves a mention here: the rebound effect. When a request becomes ten times cheaper and ten times less demanding, usage does not stay constant. It rises — often faster than the unit gain. The net benefit can then cancel out, or reverse.

That is no reason to give up on efficiency gains: progress on models and hardware is real and necessary. It is a reason not to stop there. A more efficient AI is not automatically a more frugal one if its use becomes unlimited.

Inside an organisation, this translates into simple things: knowing which uses are running, at what volume, for what result. An assistant rolled out across the company without anyone ever looking at what it produces is not an adoption success.

The subject is not limited to electricity

Reducing the question to terawatt-hours would be convenient and incomplete. A data centre consumes electricity, but it also has to be cooled, sometimes with water; building it mobilises land, materials and equipment; connecting it assumes grid capacity that does not exist everywhere.

That connection constraint now weighs on siting decisions. The growth of AI infrastructure is leading operators to look for territories offering more available electricity, more land and acceptable connection lead times — a shift European players have been documenting for months.

The trade-offs there are rarely simple. Depending on the technology chosen, reducing a site's water consumption can increase its electricity use. Optimising one indicator at the expense of the others gives flattering results and questionable decisions. Better to reason in terms of overall impact than with a single environmental indicator.

A more sustainable AI starts with architecture

All of the above stays theoretical until it is translated into design decisions. Five principles carry most of the weight.

Start from the need, not from the model

Before selecting a technology, clarify the problem, the expected volume of use, the data available, the level of quality required and the value being sought. The order looks obvious; it is regularly reversed, with the choice of model preceding the definition of the need.

The question to ask is always the same: which solution offers the best ratio between the value produced and the resources mobilised?

Size the model to the use case

Rather than a single model for everything, route between several: a light model for simple, repetitive requests, a more capable one when complexity warrants it. Classifying the requests can itself be handed to a lightweight component.

Routing is not first and foremost a resource-saving measure. It is a cost and latency measure, whose restraint is a side effect — which explains why it is far easier to get accepted than an environmental argument.

Cut out needless processing

A significant share of an AI system's consumption serves no purpose. Whole documents sent where a few passages would do, contexts inflated out of caution, redundant calls on questions already answered, agentic loops with no ceiling, trivial requests that reach the model when an upstream filter would have resolved them.

None of this is fine-grained optimisation. These are ordinary architectural decisions, taken early or never.

Measure efficiency as well as quality

An AI solution is generally assessed on its accuracy, its relevance and user satisfaction. Those criteria remain necessary. They say nothing about the cost of producing the result.

Add execution time, token count, number of calls, inference cost — and energy consumed where the measurement is available. This instrumentation belongs to AI quality and evaluation, which gains here a dimension the classic grids ignore.

A slightly less sophisticated system that is ten times more efficient may well be the better industrial choice.

Build infrastructure into the decisions

Hosting choices are not neutral: location, energy efficiency of the facilities, share of low-carbon energy where it can be steered, cooling method, pooling of resources across uses, scheduling of non-urgent workloads for hours when the grid is less strained.

Not all of these variables are within the reach of the company using the system, and it would be dishonest to pretend otherwise. Some are: the choice of a hosting region, the decision to batch a workload instead of running it continuously, dropping a redundancy that no longer serves a purpose.

You also have to measure what AI prevents

An article that stopped there would have a flaw: it would look at only one side of the ledger.

Artificial intelligence is also a lever of efficiency for what it helps to steer — energy consumption in buildings, predictive maintenance that avoids interventions and replacements, better organised logistics flows, grid operations, resource sizing. Taken individually these gains are rarely spectacular; they are often lasting.

So the right analysis does not consist only in measuring what AI consumes. It must also assess what it helps to improve or avoid. That is more demanding, because it means documenting the situation beforehand rather than settling for a narrative about expected benefits.

Hence the question that should close any AI project scoping: are the resources mobilised by the system proportionate to the economic, operational or environmental value it produces?

From digital restraint to performance

The discussion about sustainable AI marks, in its way, a form of maturity. The first years of adoption were taken up by a single question: can AI do it? The answer is now yes in a growing number of cases, and it has stopped being discriminating.

The next question is more interesting, and it is the one asked by organisations that are industrialising: what AI does it actually take to do this properly? It engages architecture, governance and delivery far more than the choice of a supplier.

A more sustainable artificial intelligence does not mean giving up on innovation. It means using the right technology, at the right level of power, for the right use. It is also a very concrete way of putting technology at the service of performance.

Sources and references

The figures in this article come from the International Energy Agency's work on energy and artificial intelligence.

• International Energy Agency, Energy and AI, 2025 — global electricity consumption by data centres and the role of AI in its growth. iea.org/reports/energy-and-ai
• International Energy Agency, Key Questions on Energy and AI, 2026 — projection from around 485 TWh in 2025 to close to 950 TWh in 2030, a rise of about 50 % in consumption by AI-specialised centres in 2025, and the gaps in consumption per request between text generation, video, reasoning and agentic uses. iea.org/reports/key-questions-on-energy-and-ai
• Reuters, Europe AI data centres seek cheaper, quicker energy and land, 19 August 2026 — context on the shift of siting projects towards territories offering more energy, land and connection capacity. reuters.com