Executive Summary
- Organisations deploying AI clusters – which require five times more fiber infrastructure density than current data centres – will grapple with power densities projected to reach 1MW and beyond
- Densification, where IT power and computing capacity is increased in a smaller footprint, is noted by Vertiv in its 2026 Frontiers report as the ‘defining macro force whose effects are felt across the entire data centre and technology landscape’
- Liquid cooling is emerging as a key solution for AI workloads thanks to reduced energy usage but with greater planning, such as through Schneider Electric and NVIDIA’s reference designs, even more can potentially be achieved
No less an authority than the UN’s Environment Programme (UNEP) has recently chimed in on how to make AI data centres more sustainable. Hongpeng Lai, chief of the mitigation branch in UNEP’s climate change division, said in January that sustainability should not be seen as a ‘constraint’, but a ‘pathway to resilient, competitive digital infrastructure.’
It’s a high-wire act to balance the inevitable strain of AI performance with sustainability – but new innovations are showing it is possible without falling off the tightrope.
Organisations deploying AI clusters are grappling with rack power densities projected to reach 1MW and beyond, according to a recent release from Motivair by Schneider Electric. This equates to a single rack consuming as much power as an entire – small – data centre did just a decade ago. At the network level, meanwhile, AI clusters require five times more fiber infrastructure density than current data centres.
The staggering 1MW figure was touted by the hyperscalers last year. At the 2025 OCP EMEA Summit, Google Cloud introduced +/- 400 VDC power delivery that can support up to 1MW per rack. This is an inexorable trend for two reasons, as Madhusudan Iyengar and Amber Huffman, principal engineers at Google Cloud noted. Firstly, the need for machine learning to require more than 500 kW per IT rack before 2030, and secondly, the need for densification of each rack, requiring a much higher voltage DC power distribution solution with power components and battery backup outside the IT rack.
Vertiv, in its 2026 Frontiers report, analysing the technology trends shaping the future of the data centre, outlines extreme densification as the ‘defining macro force whose effects are felt across the entire data centre and technology landscape.’  This runs from chip level, to system-level integration, to full facility design.
One interesting trend Vertiv touched upon was understanding the data centre as a single unit of compute. “Power, cooling, and compute must be highly integrated into one architecture – from rack, to row, to site,†the report noted. These factory-built, modular infrastructure blocks which form the gigawatt-scale data centre, combining power, cooling and IT, ‘must be done at speed to meet the compressed timelines of AI-era capacity demands’, Vertiv added.
The options available to operators to cool these next-gen data centres vary, but one thing is becoming increasingly clear: liquid cooling is becoming a must-need. Air cooling is reaching its thermal limits not just in performance but in energy consumption, as Christopher Leonard, writing for Schneider Electric’s blog in November, points out. “Air cooling systems cannot push enough air to dissipate all the heat generated by GPUs,†wrote Leonard. “You would have to place high-density AI racks in a wind tunnel for air cooling to handle that much heat.â€
Take Motivair’s most recent coolant distribution unit (CDU), the MCDU-70, as an example of what the company claims is a liquid cooling design to ‘cool high-density data centres reliably, at scale.’  At 2.5MW per unit, six MCDU-70s can provide 4+2 redundancy for large-scale facilities designs such as the NVIDIA Omniverse DSX Blueprint, where deployments target 10MW to reach gigawatt scale.
While the argument for reduced energy usage and water consumption with liquid cooling are there, from a sustainability perspective greater planning is required. That is the argument Schneider Electric puts forward with its reference designs, announced in partnership with NVIDIA in September, aimed at accelerating time to deployment and aiding operators as they adopt AI-ready infrastructure solutions.
“By providing validated, proven, and documented data centre physical infrastructure designs, Schneider Electric enables operators globally to design and incorporate next-generation power and liquid cooling controls infrastructure before the newest AI infrastructure solutions even arrive, while optimising for cost, efficiency, and reliability,†the company noted.
