Artificial intelligence is no longer a niche workstream; it is rapidly becoming the defining driver of digital infrastructure growth. The acceleration of large-scale AI training, generative models and inference workloads has triggered a seismic shift in how data centres are designed and operated, in addition to how they are regulated.
Although this creates many opportunities, more often than not these are accompanied by intense scrutiny. Recent media attention and high-profile coverage on water consumption in hyperscale facilities has heightened public awareness of potential environmental implications of digital growth and questions are surfacing.
What are the real environmental risks posed by the new AI-driven data centre era? How are operators mitigating them in a way that balances innovation with sustainability? What does this mean for me?
At the heart of the discussion lies cooling, the often invisible but vital process that keeps AI chips functioning. AI itself does not consume water, but the servers it runs on require systems that often rely on significant water volumes. The choice of technology, whether traditional air cooling, advanced liquid cooling, immersion systems or hybrid models, has implications for both environmental footprint and operational resilience. The industry must now decisively respond with transparency, proactive risk management and clear stakeholder education to ensure clients, regulators and the public understand both the scale of the challenge and the strategies already underway to address it.
Power and compute growth
The transformation of computing hardware illustrates the scale of the challenge. In the early 2000s, traditional central processing units (CPUs) such as Intel Xeon or AMD Opteron consumed 80 to 150 watts (W) per chip, producing rack densities in the range of five to ten kilowatts (KW). The shift to graphic processing units (GPUs) for high performance computing and early AI workloads during the 2010s raised power consumption to 250 to 400W per chip and lifted rack densities to 15 to 30KW. Operators began adopting hot-aisle and cold-aisle containment strategies, and this is where we saw emerging pilots in liquid cooling.
The current era, marked by the deployment of accelerators such as NVIDIA’s H100 and AMD’s MI300, has pushed power consumption to 700 to 1000W per chip, lifting rack densities to 40 to 80KW, or more. With scale, air cooling has become insufficient and liquid-based approaches are now mainstream. Looking ahead to 2030, next-generation AI and quantum hybrid processors are expected to reach 1.2 to 1.5KW per chip, producing racks of up to 200KW. Beyond 2030, racks could exceed 250 kW, as chips are projected to approach three kilowatts each and the market continues to accelerate based on projections and future roadmaps. At that stage, immersion cooling or advanced dielectric liquid systems will be mandatory, and facilities will begin to resemble energy plants as much as digital hubs.
In the UK, electricity demand today is approximately 325 terawatt-hours per year and is projected to exceed 411 terawatt-hours by 2030. A single gigawatt-class AI facility, equivalent to 500,000 or more GPUs, would alone account for around 3% (estimated) of Europe’s projected 2030 data centre load. A facility like this would act as a major grid anchor, requiring both high-voltage interconnection and dedicated substations.
The scale of AI clusters continues to illustrate this shift. Large hyperscale deployments of around 20,000 GPUs are already hitting loads of 20 to 30MW, with flagship facilities designed to accommodate 100,000 GPUs consuming between 100 and 170MW. At the extreme, future gigawatt-class clusters could reach 600 to 900MW per site. This represents a fundamental departure from the traditional enterprise workloads that once defined the sector and demands a new approach to siting, grid alignment and long-term power sourcing.
Water, cooling and public perceptions
While power is the dominant challenge, water and cooling are both contributors to reputational risk. Media narratives often highlight headline-grabbing figures on water consumption, yet often overlook the technical sophistication of these systems. The key metric here is Water Usage Effectiveness, or WUE, which is beginning to attract the same scrutiny that Power Usage Effectiveness (PUE) has attracted for years. Legacy evaporative cooling systems can use between one and 1.8 litres of water per KW-hour, while modern hybrid systems and liquid-cooled facilities are achieving between 0.1 and 0.3 litres. The most advanced operators are targeting 0.05 litres or less, approaching near-zero consumption. It’s also worth noting that indirect water use, such as in power generation, often dwarfs direct cooling consumption, a point frequently missed in media debates.
Available technology varies massively. Traditional air cooling is no longer viable beyond 30KW per rack. Cold plate liquid cooling has become the dominant approach for modern GPUs, while immersion cooling is emerging as the only feasible option for racks exceeding up to 200KW. Closed loop and waterless designs are being prioritised in regions with acute hydrological stress, reflecting the growing importance of water risk in site planning. The reality is that cooling strategy is now inseparable from corporate responsibility and operators must integrate hydrological conditions, regulatory expectations and public scrutiny into their infrastructure strategy.
Regulation, design and operational risks
As technical risks escalate, the regulatory environment is tightening. In Europe, the Corporate Sustainability Reporting Directive (CSRD) and the Corporate Sustainability Due Diligence Directive (CSDDD) mandate detailed disclosure of energy and water use. At the global level, standards such as ISO 22237 and ISO 50001 are increasingly central to frameworks for resilient and energy-efficient design, alongside the EU Taxonomy and the Energy Efficiency Directive (EED 2023/1791). Compliance is vital as a key driver of reputational and operational risk management.
Site-specific assessments are increasingly necessary, incorporating grid availability, local hydrology, planning rules and infrastructure constraints. At Gleeds, we’re always seeking opportunities to coordinate feasibility reviews, risk mapping and retrofitting strategies to reduce exposure. These measures not only mitigate operational risks but also support transparent communication with clients and regulators, ensuring responsibilities are understood and reputations are protected.
Toward sustainable innovation
The growth of AI is accelerating data centre demand to scales once thought impossible. Facilities in the hundreds of megawatts and even approaching gigawatt-class are no longer theoretical but inevitable. Power and water are not ancillary issues but defining constraints, shaping site strategies and exposing operators to public and regulatory scrutiny. Governments are beginning to respond, supporting recent initiatives such as the US’ AI infrastructure incentive and the UK’s announcement of AI growth zones. Interestingly, the bulk of new infrastructure is being driven by the US, raising more questions around what this means for the market in Europe.
Ultimately, the path forward requires prioritising sustainable innovation, meeting the extraordinary promise of AI while proactively managing regulatory, operational and ever-growing reputational pressures. The data centre sector now stands at the centre of the AI era’s most pressing question: can digital growth be achieved without overwhelming the very resources on which it depends?
