● In the telecom sector, Orange is committed to water and energy efficiency by prioritizing closed-loop cooling water systems and free cooling (using outside air) for cooling. The operator considers this approach preferable to that of major U.S. players, who rely on evaporative cooling towers.
● The sector is exploring three main avenues: waste heat recovery, the introduction of an energy label for servers, and the adoption of more energy-efficient AI algorithms.
Electricity. According to the International Energy Agency, global electricity consumption by datacenters is projected to increase by 17% by 2025, and by 50% for those dedicated to AI, against a backdrop of a 3% increase in global electricity demand. By 2030, electricity consumption by datacenters is expected to double, and that of AI-specialized centers is expected to triple. Even in France, a country that benefits from low-carbon nuclear energy, Arcep has observed a 12% increase in datacenter electricity consumption in 2024 compared to 2023.
Heat. In the United States, where awareness of the climate emergency is lower than in Europe, datacenter energy efficiency is a recognized issue: “Some AI-dedicated datacenters create ‘heat islands,’ warming the surrounding land by up to 16 degrees Fahrenheit (9 degrees Celsius),” CNN reported in late March. A study led by Andrea Marinoni of the University of Cambridge’s Centre for Earth Observation specifies that 340 million people worldwide could be affected by these temperature increases. David Nörtershäuser, Energy & Environment Expert at Orange, explains that chips designed to train and run AI currently dissipate colossal amounts of power: “An Nvidia B300 chip that consumes 1,400W dissipates 1,400W as heat… It’s no longer just a matter of electricity consumption; the thermal density in the room has also changed.”
Water. Added to this are the impacts on water availability for datacenters that use cooling towers to cool their infrastructure.
A paradigm shift in IT infrastructure
David Nörtershäuser notes that the vulnerability of existing infrastructure to climate change is a concern for many stakeholders. Vulnerability studies are currently underway. “Another, more political question is emerging: in the event of an extreme heatwave—say, Paris at 50°C—will we need to shed load? Can we afford to reduce service to maintain only essential functions?” The growth in usage therefore calls for reasoned approaches. “There is a commitment,” the researcher emphasizes, “to improve the energy efficiency of datacenters. We are aware of the need to monitor the evolution of this energy consumption.” Telecom operators allocate only 9% of their electricity consumption to datacenters, but remain vigilant in the face of the arrival of AI and GPU hosting in their datacenters.
Orange has succeeded in reducing its footprint through a rationalization policy and upgrade projects. In France, the operator is closing small datacenters and optimizing large ones. In Europe, Africa, and the Middle East, to support growing demand for services, it is replacing older datacenters with more powerful and efficient infrastructure, as seen in the project in Botswana. It is committed to installing solar power in its datacenters and works with its suppliers through performance improvement contracts to reduce the carbon footprint of its servers.
Orange’s choice: closed-loop systems and water conservation
On the issue of water, Orange takes a different stance from that of the major U.S. players. “The GAFAM companies rely heavily on evaporative cooling towers. It’s a cost-effective technique, but it consumes a lot of water,” to the point of causing water supply restrictions in some municipalities where their datacenters are located, such as in Spain. “At Orange, we have made a deliberate choice to prioritize closed-loop systems and free cooling.” In other words, using outside air as a cooling source as long as temperatures permit. Air conditioning is used only when absolutely necessary, without evaporation. The operator is innovating with water cooling integrated directly into the chips, the activation of a standby mode for servers with zero traffic, and the continuous optimization of cooling circuits to accommodate GPUs in its datacenters. Other players, such as OVH, use direct liquid cooling, which allows for handling the densest workloads while significantly reducing air conditioning requirements.
Three approaches to reducing the sector’s carbon footprint
- Waste heat recovery. A European directive, recently transposed into French law, now requires all new datacenter constructions to recover at least 20% of the heat produced for reuse. “This could rise to 40%,” notes David Nörtershäuser.
- Energy labeling for servers. A European standard is currently being drafted to provide servers with a standardized energy label based on benchmarks. This would allow buyers to compare the actual energy consumption of machines for the same computational result. “Until now, the architects who decide on the machines have had no visibility.”
- Energy-efficient algorithms and improved energy efficiency. Less energy-intensive AI models exist and are advancing. But for David Nörtershäuser, “the entire industry must be encouraged to develop more energy-efficient equipment.” On this last point, a research paper titled “Balancing Sustainability And Performance: The Role Of Small-Scale LLMs In Agentic Artificial Intelligence Systems” (published in February 2026 by researchers from Capgemini Invent and Salesforce) examines the energy impact of large language model (LLM) inference when integrated into agent-based AI systems. The study highlights that replacing massive, closed-source proprietary models (such as GPT-4o) with smaller open-source models (or those with specific architectures) offers a significant energy efficiency gain of up to 70%. Along these same lines, teams at Orange are working on ShortLM, an approach aimed at adapting LLM models into lighter, more efficient versions.
Other innovations emerging on the market: TPUs, shared cloud, Edge
- Google’s TPUs (Tensor Processing Units), specialized chips, can drastically reduce the energy consumption of inference tasks, according to this 2025 study (https://arxiv.org/abs/2502.01671).
- The development of shared cloud infrastructure offers an additional avenue for promoting horizontal and shared resource usage, thereby limiting infrastructure duplication.
- The deployment of infrastructure at the Edge (network periphery) requires logistical adjustments such as the integration of appropriate air conditioning systems; the implementation of this equipment can be justified by the viability of the business model and the demonstration of a favorable overall energy balance. To validate the economic and environmental viability of this strategic opportunity, Orange is currently studying this scenario, which includes a feasibility study and an analysis of the total cost of energy (TCO).
Educating to Change Behaviors
Finally, behind the equipment lies behavior. Orange is committed to assessing the carbon footprint of all its purchases and to expanding the sale of recycled devices. A key societal question remains: how can we empower users to take responsibility without resorting to guilt-inducing rhetoric? For the researcher, the answer lies first and foremost in education—starting at a very young age.
This text has been translated by an artificial intelligence.
David Nörtershäuser
