AI and Sustainable Data Centers: The Challenge Can’t Be Met by Cost-Cutting Alone

In the age of artificial intelligence, sustainable data centers aren’t measured solely by energy consumption: GPU density, liquid cooling, heat recovery, and integration with the local community also matter.

Key points on sustainable data centers in the age of artificial intelligence:

No organization can escape technological progress, and artificial intelligence (AI) is its most evident manifestation today. The real challenge is to ensure that this progress is sustainable—a principle at the heart of industrial ecology. Generative AI has, in fact, transformed the nature of the data center: GPUs pack previously unimaginable computing power into just a few square meters, generate far more heat, and require power supplies that the grid struggles to provide. Are AI and sustainable data centers, then, an impossible equation? The answer is clear: reducing energy consumption is necessary, but it is not enough. The key lies in striking a balance between performance, efficiency, and competitiveness, and in the ability to integrate digital infrastructure into a broader system comprising businesses, research, public administration, and the local community.

Sustainable Data Centers: Why AI Is Changing the Energy Equation

Artificial intelligence is shifting the focus of consumption. According to Gartner, electricity demand from data centers will double by 2030, and nearly half of that will be consumed by servers optimized for artificial intelligence.

+119%

La domanda elettrica dei data center passerà da 448 TWh nel 2025 a 980 TWh nel 2030, Fonte: Gartner

40%

Data center AI esistenti che potrebbero essere limitati nella loro operatività entro il 2027Fonte: Gartner

30kw

La densità cresce soprattutto nella fascia 10-30 kW per racFonte: Uptime Institute

The workload generated by AI differs from that of traditional applications. Training and inference require accelerators that operate continuously and at full power. When enterprise AI projects move from the experimental phase to production, the implications for enterprises are twofold. Energy costs become part of the TCO (Total Cost of Ownership) of AI projects, and power availability becomes a risk factor. Globally, a study published in Frontiers in Environmental Science (2026) estimates that data centers account for about 1% of global electricity demand. Projections for 2030 range from 3% to 13%, depending on the level of efficiency achieved.

For enterprise organizations, the impact is concentrated on three fronts:

• Energy costs: Energy expenses account for an increasingly large portion of the cost of computing services and must be negotiated over multi-year time frames.
• Power availability: In areas with a high concentration of data centers, grid capacity is a scarce resource.
• ESG goals: Decarbonization targets must be recalibrated to account for new computing needs.

According to Gartner (Forecast Analysis: Data Center Power Consumption, November 2025), electricity demand from data centers will rise from 448 TWh in 2025 to 980 TWh in 2030, with AI-optimized servers accounting for 21% to 44% of the total. In Europe, their share of electricity consumption will rise from 2.7% to 5%.

How much of a burden does the energy consumption of AI data centers place on the power grid?

The impact is already evident: Gartner predicts that by 2027, 40% of existing AI data centers will be limited in their operations by energy availability. For this reason, Gartner’s (2024) recommendations to organizations include long-term energy contracts and cost plans that account for significant price increases.

GPU rack density: 2 to 30 kW per cabinet

Artificial intelligence multiplies the power requirements of each rack. Power supply, cooling, infrastructure, and procurement lead times all change: a traditional data center cannot become AI-ready without structural modifications.

With the advent of GPUs, the power required by each rack increases by an order of magnitude compared to traditional servers. A single data-center-class GPU can draw hundreds of watts, and a rack houses many of them. A shift of this magnitude affects the entire facility infrastructure: power lines, switchgear, uninterruptible power supplies (UPS), air or fluid distribution, and floor load capacity. Added to this is the volatility of the hardware market: quotes for GPU servers are valid for only a very short time, and delivery times can stretch to several months. Capacity planning therefore becomes a long-term exercise, not a simple catalog purchase.

The increase in density affects three components of the infrastructure:

• Power Supply: Lines and switchgear rated for loads tens of times greater than historical loads.
• Continuity: UPS (Uninterruptible Power Supply) systems and generators rated for high and constant power consumption.
• Heat dissipation: The heat generated increases with power output and requires dedicated systems.

According to the Uptime Institute (Global Data Center Survey, 2025), density is growing primarily in the 10–30 kW per rack range, while facilities exceeding 30 kW remain few. The highest densities, typical of GPU clusters for AI, are therefore still rare in the installed base.

What does it mean to design an AI-ready data center?

An AI-ready data center is designed from the ground up for high-density workloads, with power, liquid cooling, and connectivity scaled for GPU clusters. Electrical distribution, hydraulic circuits, and technical spaces must be planned before construction begins. This is the concept behind Retelit’s new 13.6 MW AI-ready data center in Milan-Corsico, featuring liquid cooling for AI servers and a closed-loop hydraulic system.

Liquid cooling in data centers: from air to closed-loop systems

Above a certain density, air alone is not enough to dissipate the heat generated by GPUs. Liquid cooling brings the fluid close to the components, improves heat transfer, and makes it easier to remove the heat.

AI servers can generate up to 10 times more heat than traditional servers. Air-cooling systems quickly reach their physical limits: the required fan airflow and power consumption would be unsustainable. Liquid cooling solves the problem at its source. The fluid has a much higher heat capacity than air and collects heat directly from the processors or via heat exchangers mounted on the server racks. The result is greater overall efficiency and waste heat at a higher temperature, which is therefore more useful. However, the Uptime Institute (Cooling Systems Survey, 2025) describes a still-gradual adoption, driven precisely by higher densities.

DLC (Direct Liquid Cooling): a cooling technique in which a fluid circulates through plates in direct contact with the CPU and GPU, removing heat at the source before it dissipates into the room's air.

The most commonly used liquid cooling solutions are:

• Direct-to-chip: cold plates on processors—the most common solution for GPU clusters.
• Immersion: servers immersed in dielectric fluids, for extreme densities and specific applications.
• Closed-loop system: Water recirculates without continuous withdrawals from the aquifer.

When is air cooling no longer sufficient?

Air cooling is no longer adequate when the power density exceeds approximately 20–30 kW per cabinet, because airflow rates and fan power consumption become inefficient. In hybrid configurations, air continues to cool storage and networking equipment, while liquid cools the compute nodes.

Heat recovery from data centers: waste becomes a resource

The primary byproduct of a data center is heat, not energy consumption. Recovering that heat and feeding it into the district heating system transforms an environmental cost into useful energy for the city.

Every kilowatt-hour consumed by servers is ultimately converted into heat. In traditional models, this heat is dissipated into the atmosphere by cooling systems. Moreover, a renewable source like solar power can only supply a data center during daylight hours, while the computational load is continuous. For this reason, the most effective approach is not only to produce clean energy but also to reuse the energy that has already been consumed. Connecting the two systems, however, requires an energy partner, a distribution network, and a shared engineering project. In Milan, this model has become a reality, launched by Retelit in partnership with the multi-utility company A2A.

The heat recovery project at the Avalon 3 data center in Milan supplies 2.5 MW of thermal energy to the district heating network —enough to meet the needs of 1,250 households—resulting in annual savings of 1,300 TOE (metric tons of oil equivalent) and 3,300 metric tons of CO2 avoided each year, as also reported by CorCom.

How do you connect a data center to a district heating system?

The connection is made via heat exchangers that collect heat from the cooling circuits and heat pumps that bring it to the temperature required by the network. Three conditions are necessary: a district heating network near the site, an energy provider willing to invest, and a constant source of waste heat. Liquid cooling facilitates the process because it returns warmer fluids. The Avalon 3 project required two years of collaborative work: it was an infrastructure project, not merely a plant engineering add-on.

Regional Sustainability: The European Model for Heat Recovery

In Europe, data centers are becoming part of urban energy networks. Common regulations, sustainability ratings, and examples from Northern Europe and Italy show that AI can grow while returning heat to the local community.

AI should also be developed within infrastructures where organizations retain control over their own data, without handing it over to third parties: a choice in favor of digital sovereignty that must, however, also be sustainable. In September 2026, the European Commission proposed a rating system for data centers exceeding 500 kW, which measures energy and water usage, the reuse of waste heat, and flexibility with respect to the power grid. The first labels are expected in 2027. According to the Commission, reusing about half of the waste heat from European data centers would cover the heating needs of approximately 4 million households. Individual countries are also taking action: in Spain, a draft decree from 2025 mandates heat recovery for new data centers exceeding 1 MW, unless a cost-benefit analysis shows it to be unfeasible.

Several countries in Northern Europe are already turning heat from data centers into a resource for their communities. In Høje-Taastrup, Denmark, excess heat generated by a Microsoft data center will be reused to heat approximately 6,000 local homes, and the company is pursuing other initiatives to supply heat to district heating networks. The most significant project is underway in Finland: once completed, the new facilities are expected to cover about 40% of the district heating needs of Espoo, Kauniainen, and Kirkkonummi. In Sweden, the Stockholm Data Parks initiative integrates data centers into the capital’s district heating network, which aims to completely eliminate the use of fossil fuels by 2040.

Retelit is bringing this model to Italy: sustainable data centers are becoming an energy resource for the community, not just major energy consumers.

From Efficiency to Value: How to Measure a Data Center’s Sustainability

Reducing energy consumption is necessary but not sufficient. The sustainability of an AI data center stems from a balance between performance, efficiency, and competitiveness—not from a race to the lowest price.

An approach focused solely on cost savings risks producing infrastructure that is efficient on paper but unsuitable for real-world workloads. When AI goes into production, the infrastructure must continuously support high power demands while maintaining guaranteed SLAs (Service Level Agreements). Even PUE (Power Usage Effectiveness), the most widely used efficiency metric, tells only part of the story. According to the Uptime Institute, the average value has remained stable for the sixth consecutive year: the potential for improvement in existing facilities is diminishing. Sustainability must therefore also be measured by what the infrastructure gives back to the local community: recovered heat, preserved soil, and protected water resources.

The table compares the two approaches based on the criteria that have the greatest impact on long-term value.

Criterion	Savings-Focused Approach	Balanced approach
Objective	Reduce the cost per kW	Balance performance, efficiency, and competitiveness
Supported density	Suitable for traditional CPU workloads	Designed for 20–30 kW per cabinet
Cooling	Air, optimized for energy efficiency	Liquid cooling for GPU clusters, air cooling for mixed workloads
Waste Heat	Dissipated into the atmosphere	Recovered for district heating
Land use	New development areas	Rehabilitation of brownfield sites
Success indicators	PUE and price	PUE, energy returned, SLA on AI loads
Time Horizon	Short term	Entire infrastructure lifecycle

The land matters, too: the Corsico site is located on a former industrial area that was redeveloped without using any new land.

The question to ask a supplier is not just how much a kW costs, but how much value that energy generates over the entire life cycle of the infrastructure.

Why does the right mix of virtualization and GPUs increase efficiency?

Combining virtualized resources in a private cloud with dedicated GPUs allows you to make the most of your existing hardware. Next-generation data center racks can accommodate servers that offer higher performance and lower power consumption. Virtualization allows these resources to be shared across multiple workloads: it increases server utilization and reduces the number of physical machines that need to be purchased, replaced, and maintained. The GPUs remain dedicated to AI workloads that require their processing power, while supporting applications and services run in virtual environments. The result is a more favorable ratio between investment and expected return, with lower energy consumption per unit of work performed.

The question to ask a supplier is not just how much a kW costs, but how much value that energy generates over the entire life cycle of the infrastructure.

Building a Network: Businesses, Research, and the Public Sector

A data center becomes sustainable when it integrates into the community that hosts it. Partnerships between digital operators, multi-utility companies, research centers, and government agencies are needed to transform energy and heat into shared value.

No single operator can solve the equation between AI and sustainability on its own. Heat recovery only works if a multi-utility company builds the distribution network. The same applies to energy efficiency and land-use planning, which require the involvement of public agencies and research centers. Scientific literature also points to significant potential beyond the plant itself. A systematic review published in the Journal of King Saud University (2026) estimates that software optimization accounts for 15% to 30% of energy efficiency, given the same hardware.

Each participant in the ecosystem has a specific role:

• Companies: They scale AI applications to the actual computing power required.
• Infrastructure operators: design data centers capable of recovering energy and heat.
• Research: Develops more efficient hardware, software, and models.
• Government: Coordinates energy networks and development areas across the region.

Finally, there is the issue of control. Enterprise AI processes confidential data, and hosting it on infrastructure located within the country is also a matter of digital sovereignty.

Criteria for Evaluating a Sustainable and AI-Ready Data Center

The assessment must take into account available density, cooling, energy metrics, heat and water management, and transparency regarding sustainability data, always in relation to the SLAs required by AI workloads.

For organizations bringing AI into production, the choice of infrastructure should be based on measurable and verifiable criteria. The price per kW remains a key metric, but it must be considered alongside the site’s ability to scale up density without interruptions. The Uptime Institute notes that the collection and reporting of key sustainability metrics have not improved by 2025; therefore, requesting documented data is the first screening criterion.

Criteria to consider when selecting a supplier:

• Guaranteed density: kW per cabinet available today and growth roadmap.
• Cooling: availability of liquid cooling and circuit type (open or closed).
• Efficiency: Measured PUE—not just projected—with historical data.
• Energy: Renewable energy share, long-term contracts, and available additional capacity.
• Heat and water: heat recovery projects and groundwater protection.

Sustainable AI is designed from the very beginning

In recent years, artificial intelligence has moved from the experimental stage to everyday use: it supports decision-making, automates processes, and enables services that were unthinkable until recently. However, this acceleration comes at a cost that can no longer be ignored: its energy and environmental impact. Energy costs, grid capacity, and decarbonization goals are factors that must be considered right from the planning phase. Reducing energy consumption, however, is not enough: a balance is needed between performance, efficiency, and competitiveness—one that also makes the most of what the data center produces, starting with recovered heat.

Ultimately, an AI infrastructure is sustainable when it delivers performance today and provides value tomorrow.

Frequently Asked Questions About Sustainable Data Centers for Enterprise Organizations

With contributions from Luca Canova, Sales Manager for the Northeast, Retelit