A Different Kind of Power Drain
Not all data centres are created equal. While traditional facilities are built for general-purpose computing using Central Processing Units (CPUs), AI workloads require a different beast altogether. The complex calculations needed to train and run large
AI models depend on vast clusters of Graphics Processing Units (GPUs) and other accelerators. These components are far more power-hungry. A traditional server rack might consume 5 to 15 kilowatts (kW) of power, but a rack packed with high-performance GPUs for AI can demand anywhere from 30 kW to over 100 kW. This extreme power density is the core of the efficiency challenge; concentrating so much energy in a small space generates an immense amount of heat that must be managed.
The Energy Bill Comes Due
The scale of this new demand is staggering. According to the International Energy Agency, global electricity consumption from data centres is projected to roughly double by 2030, with AI workloads driving most of that growth. Some forecasts suggest data centres could consume nearly as much energy as a country like Japan by the end of the decade. For years, the industry has measured efficiency using a metric called Power Usage Effectiveness (PUE). Calculated by dividing the total power entering a facility by the power used for IT equipment, a perfect PUE score of 1.0 means no energy is wasted on overheads like cooling or lighting. While the industry has made great strides in lowering PUE, the sheer intensity of AI is pushing the limits of this traditional benchmark.
When Air Conditioning Is Not Enough
The single biggest energy consumer in a data centre, after the servers themselves, is cooling. Traditional air-cooling systems, which circulate chilled air through server aisles, are hitting a physical limit. They become inefficient and impractical when trying to cool racks that exceed 30-50 kW of power. As a result, the industry is rapidly pivoting to liquid cooling. This can involve either direct-to-chip cooling, where fluid is piped directly to hot components, or immersion cooling, where entire servers are submerged in a non-conductive dielectric fluid. Liquid is vastly more effective at transferring heat than air, allowing data centres to cool racks generating 100 kW or more while often using less energy and water to do so.
Rethinking 'Efficiency' in the AI Era
The intense power needs of AI are also forcing a debate about what efficiency truly means. While a low PUE is good, the metric alone doesn't tell the whole story. It measures the efficiency of the facility's overhead but says nothing about how effectively the computers are working. An AI data centre might perform trillions more calculations per second than a traditional one, meaning its output-per-watt could be far higher, even if its PUE is not perfect. As such, experts argue that new metrics are needed that account for the useful work performed, water consumption, and carbon emissions to provide a more holistic view of sustainability in the AI era.
A System-Wide Race for Solutions
Solving AI's energy problem requires more than just better cooling. The explosive growth in demand is straining local power grids, with connection delays becoming a major bottleneck for building new facilities. In response, the industry is pursuing solutions across the entire ecosystem. This includes designing more energy-efficient chips, optimising software to reduce idle GPU time, and building data centres in locations with access to renewable energy. Other innovations focus on reusing the tremendous amount of waste heat generated by servers, potentially for heating nearby buildings or industrial processes. It's a race to innovate infrastructure as fast as the AI models themselves are evolving.
