As the EU invests in AI gigafactories, France is confronting the land, water and energy behind digital growth. The challenge is to use AI where it creates genuine value without treating natural resources as unlimited.

On 10 July 2026, the Grenoble Administrative Court suspended the building permit for an artificial intelligence computer centre near Valence.

The project would require more than 60 MW of electrical power at full capacity. The court found that an environmental impact assessment should have been carried out and identified serious doubts about whether the installation complied with local planning rules. The decision is temporary, pending a final ruling. [1]

Twelve days later, ADEME published a position paper on the environmental effects of generative and agentic AI. It highlighted rising demand for data centres, energy, hardware, land and water, alongside the difficulty of measuring these impacts, given that leading providers disclose so little about their models and infrastructure. [2]

Then, on 30 July, the European Union launched a call for up to seven AI gigafactories. The programme could receive up to €10 billion in European and national public funding and is expected to mobilise at least €20 billion in private investment. Its purpose is to strengthen Europe’s computing capacity and technological sovereignty. [3]

Together, these decisions bring a less visible part of AI growth into focus. Data centres are built in real territories, where land, electricity and water are shared with communities and ecosystems.

Europe has legitimate reasons to seek technological sovereignty, but forecasts of future demand should not be enough on their own to justify additional infrastructure. They should open a wider discussion about which uses create genuine value, where new capacity can be built responsibly and what environmental limits should apply.

Europe’s AI ambition is becoming physical

Technological sovereignty is usually discussed in terms of model ownership, data residency, security, and regulation.

• Can European organisations rely on foreign providers for critical infrastructure?
• Where is their data processed?
• Which jurisdiction controls the service?
• Can European startups compete without depending entirely on American hyperscalers?

These concerns are legitimate, but sovereignty also has a physical layer that receives far less attention.

Advanced AI requires computing capacity. That capacity depends on processors, servers, electricity, cooling systems, suitable land and network connections.

The EU’s gigafactory programme is designed to expand this foundation. The selected sites will combine advanced processors, cloud technologies, high-speed connectivity and energy-efficient data centres. They will support the training, fine-tuning and inference of advanced models for startups, SMEs, researchers, industry and public authorities. [3]

Companies cannot build competitive AI products without access to infrastructure, but bringing more computing capacity into Europe also brings its constraints closer to European territories.

Why France can still be a strategic location

There is also a legitimate environmental argument for locating some capacity in France. RTE notes that French electricity production is more than 95% decarbonised, meaning that workloads hosted in France may have a lower operational carbon footprint than equivalent activity powered by more carbon-intensive grids. The distinction, however, is whether new French capacity replaces activity elsewhere or simply adds to total global demand. [16]

The sovereignty question therefore becomes more demanding: Can Europe expand AI capacity without treating energy, water and land as invisible inputs?

The answer will shape more than environmental policy. It will influence where infrastructure can be built, how quickly capacity becomes available, what providers charge and which companies can afford to compete.

The territorial cost of computing capacity

Land cannot be treated as a neutral surface available for digital expansion. The Office français de la biodiversité notes that soils host close to 60% of terrestrial biodiversity and that artificialisation damages their biological, hydrological and climatic functions, while destroying and fragmenting habitats.

New infrastructure should therefore prioritise existing industrial and already artificialised sites before consuming natural, agricultural or forest land. [9]

Water reveals the measurement gap

AI’s water footprint has become one of the most visible parts of the infrastructure debate.

Data centres may use water directly for cooling. Water is also involved indirectly through electricity generation and the manufacture of processors, servers and other equipment.

But the numbers are easy to misuse because they often measure different things. These categories are not interchangeable:

• Water withdrawn from a local supply
• Water consumed and not returned to the same system
• Potable water used for cooling
• Indirect water associated with electricity production
• Training or inference
• Operational use or the full lifecycle

ARCEP reported that the French data centres covered by its environmental survey withdrew 681,000 cubic metres of water in 2023, primarily drinking water. That represented a 19% annual increase, although ARCEP notes that the total remains modest compared with withdrawals for other activities. [4]

The French competition authority estimates that the electricity consumed by data centres is associated with more than 5.2 million cubic metres of indirect water withdrawals each year. [5]

National averages can hide local pressure

National totals can appear modest while local pressure remains significant. Water availability varies by season and river basin, and withdrawals during dry periods occur when rivers, wetlands, soils and vegetation may already be under stress. Water scarcity affected 28% of EU territory and 32% of its population during at least one season in 2023. [10]

Data centres are not currently France’s largest water or electricity users. ARCEP describes their direct water withdrawals as modest compared with other activities, while RTE estimates that data centres account for around 2% of French electricity consumption today. The concern is therefore not that they represent the country’s dominant environmental pressure, but that their resource demand is rising quickly, new facilities are becoming considerably larger and local impacts can be concentrated in particular territories and river basins. [4][16]

Why AI-specific figures remain unavailable

The two figures describe different parts of the system. More importantly, neither represents AI alone.

Data centres also support cloud services, websites, financial systems, streaming, storage and conventional enterprise software. Public facility-level reporting cannot currently isolate the water used by one model, product or marketing workflow.

This is why simple statements such as “AI consumed this much water in France” are unreliable.

Why litres per prompt can mislead

Turning an AI interaction into a bottle or glass of water makes an invisible system easier to imagine.

It can also create false precision.

The footprint of a response depends on factors including:

• The model and its size
• Length and complexity of the task
• Number of model calls
• Hardware
• Data-centre location
• Cooling system
• Electricity mix
• Boundaries of the assessment

Mistral offers one of the few public, model-specific examples.

Its lifecycle assessment estimates that a 400-token response from Le Chat, excluding the user’s device, is associated with 1.14 grams of CO₂ equivalent and 45 millilitres of water consumption. [6]

The figure is meaningful within the specific model, infrastructure and methodology assessed by Mistral, but it cannot be generalised to ChatGPT, Claude, Gemini or every Le Chat interaction.

Mistral also acknowledges that environmental accounting for large language models remains incomplete. Hardware data is limited, methodologies are not yet fully standardised, and providers do not all measure the same boundaries.

Without context, environmental efficiency risks becoming another vague product badge.

Growth teams influence which demand gets scaled

Growth teams may not control data-centre cooling or processor manufacturing, but their decisions influence the demand placed on that infrastructure.

They help decide:

• Which tools enter the stack
• Which processes are automated
• How many model calls a workflow makes
• Whether every visitor receives live personalisation
• How much content is generated
• How widely automated outreach is deployed
• Whether complex models are used for simple tasks
• Whether an underperforming experiment continues to run

These choices affect both commercial efficiency and infrastructure demand.

The French competition authority estimates that electricity represents approximately 30% to 50% of a data centre’s operating costs. It also argues that AI frugality could become a competitive factor by enabling providers, including smaller companies, to offer efficient services at lower cost. [5]

AI changes the economics of software usage

Once built, a conventional software feature may cost very little each time it is used. An AI feature creates variable costs whenever a model is called, particularly when it uses a large model, generates rich media or chains several steps together.

Growth teams already decide which activities receive budget, automation and scale. That same discipline can reduce unnecessary ecological demand by limiting duplicated outputs, excessive model calls, real-time personalisation with no proven benefit and large models used for simple tasks.

Commercial efficiency still matters, but it should support a wider objective: creating useful outcomes without treating infrastructure as limitless.

When efficiency increases consumption

Generative AI has lowered the time and monetary cost of producing many marketing outputs.

A team can create more campaign concepts, email variations, advertisements and landing-page drafts than it could produce manually.

This can improve experimentation and execution, but it can also encourage teams to scale output before proving its value.

ADEME specifically warns that AI assessments must account for indirect effects, including rebound effects. When a process becomes more efficient and accessible, total usage may rise enough to offset part of the original saving. [2]

Growth has seen this pattern before: more advertising inventory did not guarantee better acquisition, automation did not necessarily improve nurturing, and a larger volume of content did not automatically create authority or organic demand. AI can amplify the same mistakes at much greater speed.

Content

Producing 100 articles instead of ten is not a growth result.

If the additional content repeats existing ideas, competes for the same search intent or attracts irrelevant traffic, production has scaled while demand has not.

Outbound

AI can research accounts and generate personalised messages at considerable volume.

If targeting is poor, it simply makes irrelevant outreach look more polished. The business creates more activity, model usage and reputational risk without producing more qualified conversations.

Agents

One customer request may trigger planning, retrieval, reasoning, verification and generation across several model calls.

What matters is not the number of actions performed, but the proportion of tasks completed correctly and with proportionate resources.

When production becomes cheaper, the risk is not simply that companies create more. It is that they create more without becoming more useful.

Efficiency needs an ecological boundary

Improving energy efficiency per request will not be enough if the total number of requests, features and facilities continues to rise. A more efficient system can still increase overall pressure when it is deployed everywhere, including where it adds little value.

Growth teams therefore need to assess necessity before optimisation. Before asking how cheaply or efficiently a workflow can run, they should ask whether it needs AI, whether a simpler system can perform the task, and whether the outcome justifies the additional demand.

From cost per token to cost per useful outcome

Tokens are valuable for technical monitoring and billing.

A cheaper model call is not automatically a better decision if the result requires extensive correction, creates a poor customer experience or fails to influence revenue.

Commercial cost per useful outcome

Cost per Useful Outcome offers a more useful starting point by connecting the full operational cost of a workflow to a verified result. The principle is not new. It applies familiar unit-economics thinking to AI workflows, while tracking resource intensity alongside financial and operational costs.

It can improve commercial discipline, but it should sit within clear ecological limits. A workflow may perform well financially and still be unacceptable if it depends on avoidable infrastructure, unnecessary land artificialisation or additional withdrawals in a water-stressed territory.

The calculation should include: Platform and API costs + implementation + human review + correction and failure costs

divided by: Verified successful outcomes

The successful outcome depends on the use case. It could be:

• Qualified meeting
• Incremental conversion
• Activated user
• Retained customer
• Accurately resolved request
• Completed internal task
• Creative asset that improved campaign performance

The outcome must be defined before the workflow is expanded.

Resource intensity per useful outcome

Environmental and technical indicators should be tracked alongside the commercial calculation, not mixed into the same unit.

Depending on the data available, these could include:

• Model calls per successful outcome
• Tokens or compute per outcome
• Energy per outcome
• Water estimates supplied by the provider
• Percentage of work routed through smaller models
• Human corrections per outcome
• Quality of the vendor’s environmental disclosure

The objective is to make a better commercial decision about whether each use case merits continued investment.

What resource-aware growth looks like

Resource-aware growth requires the same discipline already expected in acquisition, experimentation and revenue operations.

Begin with a defined problem

“We need an AI agent” is not a business case.

“We need to reduce the time Sales spends researching accounts without lowering research quality” is one.

A clear problem makes it possible to compare the AI-supported workflow with the existing process.

Establish a baseline

Before implementation, measure:

• Time required
• Current cost
• Output quality
• Error rate
• Conversion or completion rate

Without a baseline, adoption is easily confused with improvement.

Prove incrementality

Use control groups, phased deployment or before-and-after analysis to determine whether AI changed the outcome.

A generated landing-page variation that performs no better than the original has not created incremental growth.

Match the model to the task

Not every activity requires the largest available model.

Classification, extraction, routing and structured rewriting may be handled effectively by smaller or specialised systems.

A more resource-intensive model is justified when it materially improves accuracy, commercial performance or risk management.

Build stopping rules

AI workflows can expand automatically, which makes operational limits essential.

These may include:

• Usage budgets
• Maximum calls per task
• Model-routing rules
• Confidence thresholds
• Human review requirements
• Content-volume limits
• Performance thresholds below which an experiment stops

Ask vendors better questions

Buyers may not receive complete environmental data today, but they can still ask:

• Where are workloads processed?
• Can the model or processing region be selected?
• Are smaller models available?
• Does the system use model routing?
• Is usage visible by workflow?
• Has the provider published lifecycle or inference data?
• What does its methodology include?
• Which impacts remain estimated?

The ability to answer clearly can become part of vendor trust.

What should Europe require before scaling further?

The EU already requires data centres with installed IT power demand of at least 500 kW to report energy and water indicators to a European database.

The framework covers total energy and water use, Power Usage Effectiveness, Water Usage Effectiveness and waste-heat reuse. Public information is currently published at aggregated Member State and EU levels. [7][8]

The Commission is also preparing a wider Data Centre Energy Efficiency Package, including a common European rating scheme and future minimum performance standards. [7]

This does not yet provide workload-level transparency. A buyer cannot use the database to identify the water footprint of its chatbot or campaign.

Energy and water efficiency are moving from an internal infrastructure concern towards reporting, procurement and competition.

From reporting to responsible infrastructure decisions

Europe now has an opportunity to define the conditions under which additional capacity is acceptable. Rather than treating every forecast of AI demand as a reason to build, policymakers, providers and buyers should ask:

• Can an existing industrial or artificialised site be reused?
• Is the proposed location already exposed to water stress?
• Has a complete environmental assessment been carried out?
• Would the facility replace existing foreign capacity or simply add to total demand?
• How will habitats, soils and ecological corridors be protected?
• Could smaller models or fewer calls deliver the same outcome?
• Can users disable generative features they do not need?
• Are energy, water, hardware and lifecycle impacts reported transparently?

A credible provider should be able to explain which models it uses, where workloads run, how tasks are routed, what its assessment includes and how efficiency changes across use cases.

The French competition authority has already identified frugality as a potential competitive parameter that can influence price, quality and innovation. [5]

These conditions could help Europe develop an AI model shaped by greater transparency, resource restraint and territorial responsibility, rather than measuring success only through capacity.

FAQ

Does every AI prompt consume the same amount of water?

No. The result depends on the model, task, hardware, location, cooling technology, electricity source and assessment methodology. A number calculated for one provider should not be generalised to all AI interactions.

Did the French court suspend the Alixan project because of water consumption?

No. The court found that the project required an environmental impact assessment and identified serious doubts about compliance with planning rules. The decision is an interim suspension, not a final cancellation. [1]

Are France’s data-centre water figures specific to AI?

No. Data centres support many digital activities. Current public reporting does not isolate AI workloads from the rest of a facility’s operations. [4][5]

Should companies use less AI?

The better goal is selective and measurable use. Companies should identify where AI creates meaningful value, choose an appropriate model and stop expanding workflows that produce activity without useful results.

Why should growth teams care?

Growth teams influence which workflows are deployed, how frequently models are called and whether those systems continue to scale. These decisions affect margins, customer experience and infrastructure demand.

Can efficiency become a competitive advantage for European AI?

Potentially. Efficient models, regional infrastructure, transparent methodologies and auditable reporting can support lower costs, stronger procurement cases and greater trust. European origin alone is not enough.

Conclusion

July 2026 has made the tension surrounding Europe’s AI ambitions unusually visible. A French court suspended an AI infrastructure permit because the required environmental assessment had not been carried out, ADEME called for greater transparency and lifecycle measurement, and the EU launched its call for up to seven AI gigafactories, backed by up to €10 billion in potential public funding and at least €20 billion in expected private investment. [1][2][3]

These developments are unfolding within a wider ecological reality. Earth Overshoot Day fell on 30 July, while France was experiencing its third official heatwave of the summer and nearly 495,000 hectares had been mapped as burned across the EU in 2026. These events were not caused by AI infrastructure, but they show how heavily land, water and energy systems are already being tested. Additional data-centre capacity can increase that pressure, particularly in water-scarce regions and during periods of extreme heat. [11][12][13]

AI can support research, medicine, accessibility, conservation and everyday work. Its usefulness should encourage more careful decisions about which applications deserve to scale and what each new facility asks of land, water and energy.

Water must remain a public and ecological priority

Water is both a human right and a shared ecological resource. French law gives priority to essential human needs while requiring the protection of aquatic ecosystems and natural water systems. These principles should define the conditions under which industrial and digital uses are authorised, particularly during droughts and heatwaves. [14][15]

Europe still has an opportunity to regulate before rising demand becomes difficult to challenge. This means careful siting, full environmental assessments, transparent lifecycle information, smaller or specialised models where they are sufficient, and the ability to refuse projects whose public value does not justify their local ecological cost.

Where growth teams can make a difference

Growth teams influence which applications become normal, which experiments continue and which automated processes multiply. Measuring useful outcomes can help distinguish applications that genuinely improve people’s work or lives from those that mainly produce more volume.

AI is already part of our economic and social systems, which makes strong regulation essential. European leadership should be judged not only by the computing capacity it builds, but by the care with which it decides what deserves scarce land, water and energy.

💌 If you want more essays on ethical growth, trust-led marketing and practical B2B growth systems, you can join Conscious Growth Dispatch here.

Sources

[1] Grenoble Administrative Court, Suspension du projet de Datacenter à Alixan, 10 July 2026. The court temporarily suspended the permit for an AI computer centre requiring more than 60 MW at full capacity.

[2] ADEME, IA générative, comment quantifier les impacts ?, 22 July 2026. ADEME reviews the direct and indirect environmental effects of generative and agentic AI and recommends standardised measurement, greater transparency and attention to rebound effects.

[3] European Commission Representation in France, L’UE lance un appel d’offres pour les giga-fabriques d’IA, 30 July 2026. The call supports up to seven gigafactories with up to €10 billion in public funding and at least €20 billion in expected private investment.

[4] ARCEP, Enquête annuelle Pour un numérique soutenable, 17 April 2025. ARCEP reports that the French data centres covered by its survey withdrew 681,000 cubic metres of water in 2023, primarily drinking water.

[5] Autorité de la concurrence, Étude sur les enjeux concurrentiels liés à l’impact énergétique et environnemental de l’intelligence artificielle, 17 December 2025. The study covers energy costs, direct and indirect water use, environmental transparency and AI frugality as a competitive parameter.

[6] Mistral AI, Our contribution to a global environmental standard for AI, 22 July 2025. Mistral reports the lifecycle footprint of Mistral Large 2 and the estimated marginal impact of a 400-token Le Chat response.

[7] European Commission, Energy performance of data centres, accessed 31 July 2026. The page explains the European reporting database, its water and energy indicators and the planned Data Centre Energy Efficiency Package.

[8] European Union, Commission Delegated Regulation (EU) 2024/1364, 14 March 2024. The regulation establishes reporting requirements and sustainability indicators for data centres with installed IT power demand of at least 500 kW.

[9] Office français de la biodiversité, Lutter contre l’artificialisation des sols. The OFB explains that soils host close to 60% of terrestrial biodiversity and that artificialisation damages their biological, hydrological and climatic functions.

[10] European Environment Agency, Water scarcity conditions in Europe, updated 2026. Water scarcity affected 28% of EU territory and 32% of its population during at least one season in 2023.

[11] Global Footprint Network, Earth Overshoot Day 2026 falls on July 30, 5 June 2026. The calculation estimates that humanity is using ecological resources and services 73% faster than Earth can regenerate them, equivalent to 1.73 Earths.

[12] Météo-France, Orages et fortes chaleurs à l’est, 30 July 2026. Météo-France confirms that the heatwave beginning on 28 July was the third national heatwave of summer 2026. It followed two earlier summer heatwaves and an exceptionally early May heat episode that did not meet the national heatwave threshold.

[13] European Forest Fire Information System, EFFIS Statistics Portal, accessed 31 July 2026. EFFIS had mapped approximately 495,000 hectares burned across EU countries during 2026, including 91,057 hectares in France. The figures are updated regularly and principally cover fires of approximately 30 hectares or more.

[14] UN-Water, Human Rights to Water and Sanitation. The United Nations recognises access to sufficient, safe, acceptable, physically accessible and affordable water for personal and domestic use as a human right.

[15] Légifrance, French Environmental Code, Articles L210-1 and L211-1. French law defines water as part of the nation’s common heritage, recognises access to drinking water for essential uses and requires the preservation of aquatic ecosystems, wetlands and natural balances.

[16] RTE, Les data centers en chiffres clés, 2026. RTE estimates that data centres currently consume around 10 TWh, approximately 2% of France’s annual electricity consumption. This could rise to between 23 and 28 TWh, around 4% of national consumption, by 2035.

The post The Hidden Cost of AI Growth appeared first on SAMIN GHIASI | GROWTH MARKETING CONSULTANT | FR/EN.