Artificial intelligence is expanding at a scale now measured in megawatts, cubic metres and tonnes. Training large models and serving high volumes of inferences intensify demand for electricity, water and hardware. Europe illustrates both scale and scrutiny, with rapid capacity growth, tighter efficiency targets and mandated transparency. Yet per-task efficiency gains risk being outweighed by aggregate usage. Progress depends on pinpointing where impacts arise, applying proven optimisation across the AI lifecycle and aligning governance with the reality of rising demand so capability advances without amplifying environmental load.
Energy, Water and Materials Pressure
Environmental impact concentrates on the intensive compute for model training and the persistent energy draw of inference. Over time, inference dominates because small per-query costs accumulate at scale. This load drives rising data centre electricity demand, while embodied impacts extend upstream to construction, networking and end-user devices.
Europe highlights the trend and the policy shift. Investment in European data centres is projected to reach €100 billion by 2030, with annual demand rising 15% between 2023 and 2030. Surveyed facilities report 94% of energy from renewables. From 1 January 2025, the Energy Efficiency Directive requires EU data centres to measure power usage effectiveness (PUE) and submit results to a public database, improving comparability and accountability.
Efficiency benchmarks are tightening. European facilities reported an average PUE of 1.48 in 2024, with targets of 1.30 or lower for new sites in cooler climates and below 1.40 in warmer regions. One site in Saint-Ghislain, Belgium, reported 1.08. These gains show what is possible when cooling, power distribution and IT loads are integrated, yet they do not cancel scale effects from accelerating AI use.
Water introduces a second constraint that can diverge from carbon metrics. Shifting workloads to cleaner electricity windows may coincide with hotter hours that drive up cooling water demand, raising the water footprint despite lower emissions. Operators therefore need to optimise timing, location and technology choices across electricity and cooling rather than prioritising a single metric.
What the Data Shows on Water and E-Waste
Recent disclosures sharpen the picture on water intensity. One provider reported that training a large system and 18 months of use consumed 281,000 m³ of water and that a 400-token inference required 45 mL. At the hyperscaler level, combined reporting indicates water withdrawn rose 76.6% between 2020 and 2024, with an average of 68% of withdrawn water consumed. In 2024, 36.6 million m³ was consumed, up from 18.4 million m³ in 2020. Water sourcing risk is uneven, with reported consumption including portions from high- and medium-risk sources and gaps where breakdowns are not disclosed. These patterns show why water stewardship must be assessed alongside carbon, grid mix and thermal conditions across specific regions.
Hardware refresh cycles compound the challenge. Global e-waste reached 62 million tonnes in 2022, with about 20% formally collected and properly recycled. Fast-moving AI workloads accelerate server and accelerator upgrades, shortening useful life and increasing volumes entering waste streams with limited tracking. One lifecycle assessment reported that manufacturing, transportation and end-of-life of server infrastructure accounted for 61% of materials consumed over a model’s lifecycle, underscoring embodied impacts relative to runtime energy. Projections suggest AI expansion could raise global e-waste by 3% by 2030, equivalent to an additional 2.5 million metric tons annually, while only a quarter of operators measure e-waste. Without better measurement, reuse and recovery, valuable and hazardous materials risk being discarded with insufficient oversight.
A paradox follows. Software and hardware advances make individual tasks more efficient, yet aggregate energy demand accelerates as adoption widens. Lower costs and easier access drive more applications, more queries and more models in production. Efficiency alone cannot deliver absolute reductions when total usage expands faster than per-task savings.
Must Read: NHS Ten-Year Plan for Sustainable AI-Enabled Care
Mitigation Pathways with Measurable Impact
Multiple levers now show material reductions at model and system levels. Energy-aware cluster design for inference can reconfigure resources dynamically to meet latency targets while conserving power. One framework reported conserving energy by 53%, cutting operational carbon by 38% and reducing costs by 61%. Model-centric tactics such as quantisation and pruning and using appropriately sized models for the task, can lower energy for individual tasks by up to 90%. For many specialised applications, small language models (SLMs) suit repetitive, domain-specific work, including emerging agentic patterns, delivering sufficient capability with lower overhead than large language models (LLMs).
A sustainable path couples optimisation with governance that manages scale and system constraints. Projections indicate data centre energy demand more than doubling by 2026, driven by newly accessible AI capabilities and widespread deployment, so right-sizing workloads matters as much as squeezing efficiency from each operation.
Operators are also shifting the energy base. Long-term power purchase agreements secure renewable generation. Facilities are reusing unavoidable heat through district schemes. Backup power is evolving as renewable diesel such as hydrotreated vegetable oil replaces conventional diesel. Cooling is becoming smarter, with extensive use of free cooling that leverages ambient air where conditions permit.
Architecture choices can further curb network and compute overhead. Edge AI moves inference closer to where data is generated, reducing traffic to centralised clouds, improving responsiveness and strengthening security and privacy. The combination of SLMs and increasingly capable consumer-grade hardware makes such deployments feasible across a widening set of generative use cases. Beyond data centres, AI can support renewable integration and grid management and can improve irrigation precision in agriculture by combining Earth observation, on-field sensors and predictive analytics to reduce unnecessary water use.
AI’s environmental footprint spans energy, water and materials. Evidence points to rising pressure and to practical levers for change. Progress rests on three pillars: tighten operational efficiency across training and inference, select right-sized models and architectures including SLMs and edge deployments and pair technical gains with governance that manages scale and improves transparency. With renewable procurement, heat reuse, cleaner backup fuels and advanced cooling, data centres can reduce operational load while better measurement and lifecycle strategies address embodied impacts and e-waste. Aligning AI adoption with these measures enables responsible innovation that supports clinical, operational and policy goals without overlooking environmental constraints.
Source: PwC Belgium
Image Credit: iStock
