The hidden cost of AI

A digital revolution… powered by electricity

Generative artificial intelligence is generating as much ink today as it consumes energy.

As it spreads into our work tools, our search engines and our social feeds, it is becoming the symbol of inevitable technological progress: able to generate code, simulate dialogue or produce images worthy of the biggest studios.

On social media, viral trends such as “AI-generated starter packs,” stylized portraits or absurd memes produced in seconds have become cultural expressions of a technology that is now commonplace.

Article overview

That said, behind this creative excitement and apparent lightness lies a material reality that is far heavier to bear.

Contrary to the illusion of immateriality often associated with digital technology, every request made to an AI model (every image generated, every answer produced) draws on a colossal infrastructure of servers, data centres, GPUs and cooling systems, all powered by a resource that could not be more tangible: electricity.

A study by Hugging Face and Carnegie Mellon University shows that generating a single image with a model such as Stable Diffusion XL can use as much energy as fully charging a smartphone.

Worse still, generating 1,000 images would be equivalent to driving 4.1 miles (6.6 km) in an average gasoline-powered car.

And it is not so much the initial training of the model that weighs most heavily on the energy balance (although it too is extremely costly) as its daily use at scale.

According to researcher Sasha Luccioni: “For models like ChatGPT, it only takes a few weeks of massive use for the emissions generated by queries to exceed those of their initial training.”

Herein lies the fundamental paradox of our time: a technology presented as a lever for efficiency (and potentially an ally in the energy transition) can also accelerate our dependence on resources if its use is left unchecked.

In a world where large companies are investing hundreds of billions in AI, where every tool becomes “AI-powered,” and where demand for computing is exploding, it is essential to step back.

If AI is set to play a key role in organizational productivity and innovation, it cannot be adopted without greater awareness of its environmental footprint.

This awareness is not opposition to progress, but a call for strategic intelligence.

At a time when the energy intensity of AI models grows as their power increases, and when promises of efficiency are sometimes contradicted by uncontrolled rebound effects, we urgently need to guide our technology choices with discernment.

This article offers a management perspective on the efficient use of generative AI: how to reconcile innovation and responsibility, how to make the right technology trade-offs, and why it is imperative to treat environmental impact as a strategic criterion in the selection, implementation and governance of AI tools.

The hidden side of our prompts: what AI consumes behind the scenes

It is tempting to believe that a request sent to a conversational assistant, however sophisticated, is a harmless act with no tangible consequence.

After all, it is only a few words typed on a keyboard, processed in an abstract algorithmic cloud.

Yet this apparent immateriality hides a very concrete energy reality: generative artificial intelligence relies on an industrial infrastructure that is hungry for electricity, water, materials and computing resources.

Take a simple example: generating a five-page report with a large language model such as GPT-4o.

According to measurements from the Code Carbon platform, a single request of this kind consumes 439 Wh on average, the equivalent of 4.49 megajoules of primary energy, and emits 268 grams of CO₂ equivalent.

To give a sense of scale, that is equivalent to:

• 2.58 km driven in an electric car;
• 5.37 km of running in terms of human energy expenditure;
• 4.18 hours of video streaming.

In other words, automatically drafting a professional document is far from trivial.

If only 1% of the world’s population (80 million people) performed this type of task every day for a year, the global footprint would be colossal: it would represent the annual output of nearly two nuclear power plants, or the equivalent of the emissions of 44,000 round-trip flights between Paris and New York.

What this demonstration highlights is the silent accumulation of a distributed energy footprint.

Each request is inexpensive on an individual scale, but their multiplication, made possible by the growing accessibility of models and their integration into everyday software (office suites, CRMs, search engines, voice assistants), creates a systemic cost.

In this context, generative AI is no longer a gadget, but a full-fledged player in global digital energy consumption.

According to the International Energy Agency (IEA), data centres already consumed 415 TWh of electricity in 2024, or about 1.5% of global consumption.

This share is expected to more than double by 2030, reaching a level comparable to Japan’s current consumption.

A significant part of this growth is directly linked to the rise of artificial intelligence uses.

This finding overturns a common belief: it is no longer only the training phases of models (admittedly energy-intensive) that are the problem, but their massive and widespread use.

As researcher Sasha Luccioni pointed out earlier, it takes only a few weeks of use by millions of people for the emissions associated with using a popular model like ChatGPT to exceed those of its initial training.

In other words, it is not the models themselves that are unsustainable, but the way we use them, without discernment or strategy.

Digital technology does not escape the fundamental laws of physics: every calculation has an energy cost, every iteration draws on matter, every generation consumes resources.

It therefore becomes essential to take a systemic view of AI’s environmental impact, moving beyond the myth of software magic towards rational, managerial thinking.

The point is not to restrict innovation, but to shed light on the conditions for optimizing it.

Beyond electricity: water, metals and the hardware life cycle

When people discuss the environmental impact of artificial intelligence, most conversations focus on electricity consumption.

Yet this consumption is only the tip of a much larger iceberg.

Behind every request, every model and every data centre lies a considerable water, mineral and material footprint, too often overlooked in strategic thinking.

1. Water, the invisible resource at the heart of AI

Cooling data centres

The ultra-dense servers used to train and deploy AI models generate significant heat, requiring sophisticated cooling systems.

Many of them use evaporative water systems, which pump clean water through radiators to remove heat.

Depending on the technology used and the local climate, an estimated 0.18 to 1.1 litres of water is required for every kWh consumed.

As an indication, GPT-3 (the basis of ChatGPT) reportedly uses about 500 mL of water to process 10 to 50 requests.

And a typical hyperscale data centre can consume up to 2.1 million litres of water per day, the equivalent of a small town’s daily consumption.

Semiconductor manufacturing

Water is also crucial upstream, in the manufacturing of GPU chips.

This process requires “ultrapure water” to clean wafers in repeated cycles of photolithography, etching and deposition: about 60% of manufacturing energy is dedicated to technical processes on the wafers.

The remaining 40% is related to water treatment, ventilation and buildings.

Global giant TSMC consumes 157,000 tonnes of water per day, an increase of more than 70% since 2015, reflecting the growing needs of AI.

Overall water footprint

In 2023, the annual water consumption of data centres was estimated at 560 billion litres, two-thirds of it linked to electricity generation, a quarter to direct cooling, and the rest to semiconductor manufacturing.

If nothing changes, this figure could reach 1,200 billion litres by 2030, the equivalent of 480,000 Olympic swimming pools.

2. Rare metals: critical resources under pressure

AI servers and accelerators contain a myriad of rare and critical metals: copper, silicon, aluminum, tantalum, tin, cobalt, lithium, gallium, germanium and tungsten, to name only the main ones.

These materials require intensive extraction: hundreds of tonnes of ore must be processed to produce a single tonne of refined metal.

Some are classified as “conflict minerals,” notably cobalt and tungsten, whose extraction in unstable areas raises serious ethical and social questions. Security of supply

By 2030, the exponential growth of data centres could account for 2% of global demand for copper and silicon; > 3% for rare earths; and 11% for gallium.

Yet more than 90% of aluminum, silicon and gallium is produced in just three countries, which considerably increases the vulnerability of supply chains.

3. Equipment life cycle: thinking beyond use

Manufacturing vs operation

While the active use of a server (over 5 years) accounts for about 80% of its total energy consumption, manufacturing alone accounts for about 20%, a figure that is far from marginal given the volume of equipment deployed.

Embodied emissions

The study by Luccioni et al. (2023) on the BLOOM model (176 billion parameters) highlights a complex distribution of carbon impact:

• 22.2% of training-related emissions come from GPU manufacturing;
• 48.9% from dynamic consumption (active use);
• 28.9% from idle consumption (cooling, storage, ventilation).

Ultimately, AI’s footprint goes far beyond the electricity meter.

It is part of an extractive, water and material cycle that draws on increasingly scarce resources, in unstable geopolitical contexts.

For AI to be a lever for sustainable development, moderating its use is not enough.

Every link in its value chain must also be managed, from the mine to the data centre, including water purity and intelligent design.

Debunking the myth of “zero marginal cost”: why intelligence is not just electricity

A widely shared idea in technology circles holds that, thanks to the growing automation of data centres, the cost of artificial intelligence should eventually converge with the cost of electricity.

As appealing as it may be, this view stems more from abstract economic reasoning than from rigorous systemic analysis.

Reducing intelligence to simple electricity consumption (for example, 0.34 Wh per ChatGPT query according to Sam Altman) obscures a far more complex reality: every operation relies on a resource-intensive material, logistical and energy ecosystem.

Behind the shine of the technological promise lies an entire value chain, drawing on GPUs, critical metals, ultrapure water, cooling infrastructure, distribution networks and maintenance cycles.

Focusing exclusively on instantaneous energy consumption therefore masks the broader, cumulative footprint of the system.

So it is not AI that is inherently unsustainable, but the way we use it without discernment, wrongly believing that its growth can come with energy neutrality.

In short, true mastery of artificial intelligence will not come from lower unit costs alone, but from our ability to govern how intensively it is used, to calibrate deployments to business objectives, and to factor invisible externalities into every technology decision.

Because intelligence is not free: it depends on our willingness to make it sustainable, not only economically, but ecologically and strategically.

Let’s not use a jackhammer to crack a nut

With the lightning-fast democratization of generative AI tools, a temptation is growing in organizations: to do everything with large models.

Proofreading a text, generating SEO tags, sorting emails, summarizing meeting notes… For each of these cases, the reflex is often to call on ChatGPT, Claude or Gemini, sometimes built right into the office suite.

Automation is just a click away.

But at what cost? The analogy is simple, almost trivial, but it reveals a fundamental tension: you don’t buy an 18-wheeler to pick up a pint of milk.

Using a massive language model to accomplish a simple task (one that could be handled by a specialized tool or a lighter model) amounts to wasting a precious computing resource.

You don’t use a jackhammer to crack a nut. - Sasha Luccioni

This technological “oversizing” is not just a misallocation of resources: it generates unnecessary operating costs (APIs, infrastructure, electricity consumption) and a disproportionate environmental footprint.

For example, classifying movie reviews with a generic LLM (large language model) uses up to 30 times more energy than a specialized model designed for that task.

The choice of model should therefore never be left to chance or to the end user’s convenience. It is a strategic decision, just like choosing software, a supplier or a production tool.

And that choice must include a new dimension: its return on environmental investment (REOI).

Performance should be assessed not only in terms of quality or speed, but also by its energy intensity, its indirect emissions and its impact on material resources.

Technology leaders themselves acknowledge this.

The efficiency gains offered by companies such as NVIDIA, through more powerful GPU architectures, or through model optimization via quantization or distillation, are necessary… but far from sufficient.

As long as use remains massive, frequent or even automatic, these improvements will be swallowed up by a rebound effect: the more efficient a technology becomes, the more it is used, and the more its total consumption rises.

That is why sobriety cannot be purely technological; it must also be behavioural and organizational.

In other words, it calls for active AI governance.

This can include:

• internal guidelines specifying when to use a heavy versus a light model;
• setting “carbon budgets” per team or department for the use of AI tools;
• raising employee awareness of the invisible impacts of their requests;
• energy performance indicators tied to AI use in dashboards.

Because if every task, every draft and every light automation systematically goes through a general-purpose LLM, the organization unknowingly contributes to the explosion of global energy demand, undermining the promised efficiency gains.

In short, choosing the minimum effective model combines strategic intelligence, environmental responsibility and operational rigour.

This is not about slowing innovation, but channelling it.

Because in a world where demand for computing is about to rival that of major heavy industries, true modernity will no longer mean adopting AI at any cost, but doing so at the right cost, and with clear eyes.

The illusion of automatic progress: AI’s rebound effect

One of the great myths of technological progress rests on the idea that efficiency spontaneously leads to sustainability.

Yet the history of technology shows exactly the opposite: as a tool becomes more powerful, accessible and fast, its use becomes commonplace, intensifies… and ends up generating more consumption than it originally saved.

This phenomenon, well known in energy economics as the rebound effect, fully applies to artificial intelligence.

Applied to generative AI, the rebound effect translates into a proliferation of micro-uses: an email summary here, a stylized image there, an automated brainstorm, a prompt for a LinkedIn post… Uses that seem harmless in isolation but which, multiplied by millions of users and machines, represent a massive cumulative environmental cost.

The numbers speak for themselves.

When the BLOOM model was deployed, the use phase (that is, users’ daily queries) generated, in just a few months, a carbon footprint equivalent to that of the model’s entire training phase.

And that is without counting the indirect water consumption for server cooling, or the rare metals used in hardware components.

In other words, the more efficient AI becomes, the more we use it, and the more we consume overall.

The equation is counterintuitive, but formidable.

“And the most worrying part is that these new uses are often unnecessary, created out of opportunism or convenience, with no real added value; like leaving a tap running simply because the water is unlimited.” - Gabriel Tassé

Faced with this dynamic, the response cannot be purely technological.

If we rely solely on more frugal models or greener data centres, we risk ignoring the behavioural root of the problem.

Because even a more efficient AI, used without restraint, will end up deepening our collective footprint.

This is where management comes in, no longer as a user of AI, but as the guardian of its purpose.

It is no longer enough to train teams to use the tools; we must also build a culture of discernment.

This can take several concrete forms:

• Establishing reasoned usage rules (e.g., “prompt light” for routine tasks); training on the invisible environmental impact of digital uses;
• Assessing use cases based on their real usefulness and their marginal energy cost;
• Prioritizing high-ROI automations over conveniences with little strategic value.

From this perspective, the next frontier in steering AI will not only be mass adoption, but organizations’ ability to finely regulate its use so that technical progress does not become an ecological headlong rush.

The real challenge is therefore not to slow innovation, but to intelligently orchestrate how intensively it is used.

Sustainability will not be built by default, but by design, and by governance.

Building a sustainable AI strategy: towards responsible model governance

As generative AI becomes embedded in every layer of the organization (from marketing to finance, HR, production and customer service), it becomes unthinkable to let its use develop opportunistically, without a framework or strategy.

Sustainable AI is not banned or restricted AI: it is AI that is orchestrated, governed and aligned with the company’s values and objectives, including environmental ones.

Like IT security or the protection of personal data, the use of AI should be subject to clear internal policies, monitoring mechanisms and broader performance criteria that include environmental indicators.

This managerial approach does not rest solely on ethical principles; it is also grounded in economic logic: reducing the energy footprint also means reducing infrastructure, server, licensing, API and cooling costs.

Towards a return on environmental investment (REOI)

It is time to complement traditional ROI analysis with a new metric: REOI, or “Return on Environmental Investment.”

For equal performance, the model that consumes less, uses fewer abiotic resources (such as rare metals) or requires fewer server calls should be preferred.

This means including the following in the evaluation of AI projects:

• Energy per task (Wh);
• Associated CO₂eq emissions;
• Water use for data centre cooling (often invisible, but critical);
• Pressure on material resources (metals, rare earths).

This analytical framework becomes all the more relevant with emerging tools such as the AI Energy Score project, which proposes to rank AI models by their energy efficiency per task, making it easier to compare heavy, light and specialized models

Levers for decision-makers

Here are some concrete avenues for organizations wishing to structure a responsible AI strategy:

• Develop an internal taxonomy of AI uses: distinguish critical uses (high value, automation of costly processes) from incidental uses or those that other tools can handle.
• Set up an “energy budget” or “carbon budget” per department: make teams aware of the invisible cost of their uses.
• Develop an energy performance dashboard for AI projects, paired with business KPIs.
• Encourage the adoption of “slim” or specialized models when they are sufficient for the task.

AI is not a “turnkey” product to be adopted passively.

It is strategic infrastructure to be managed over time. And in an era of CSR reports, ESG audits and growing climate commitments, companies can no longer afford to ignore the environmental impact of their technology choices.

TL;DR

Generative artificial intelligence marks a turning point in the history of digital productivity.

It opens up an extraordinary field of optimization, automation and assisted creation.

But like any technological revolution, it requires a critical reassessment of its externalities, not after the fact, but from the moment it is implemented.

The myth of an ethereal, seamless AI with no footprint must give way to a more mature awareness: every prompt has a cost, every generation a consequence.

Whether it is electricity, cooling, storage or distributed computing, requests sent to the most powerful models draw on resources as tangible as an industrial engine or a production line.

Leaders who can integrate this reality into their technology strategy will gain a head start, not only in terms of image or ESG compliance, but in terms of operational efficiency, energy resilience and environmental sustainability.

Adopting AI responsibly does not mean restricting innovation; it means steering it intelligently, maximizing its added value where it is justified, and limiting its use where other solutions (human, analog or conventional digital) are sufficient, or even preferable.

In an era marked by the tension between technological growth and planetary boundaries, the question is no longer whether we should use generative AI, but how we should do it.

Not at any price, but at the right price.

Not by giving in to euphoria, but by choosing with rigour.

And above all, by remembering that true innovation today is no longer just about doing more, but about doing better, with less.

Sources

• AS Luccioni, S Viguier, AL Ligozat. Estimating the carbon footprint of bloom, a 176b parameter language model.
• Making an image with generative AI uses as much energy as charging your phone. MIT Technology Review. S. Luccioni, B. Trevelin, M. Mitchell.
• The Environmental Impacts of Artificial Intelligence: A Primer. Hugging Face.
• Energy and AI: A Growing Relationship with Big Implications. International Energy Agency.
• AI Energy Score: Towards Comparable Energy Efficiency Ratings for AI Models. Hugging Face.