As AI racks become denser and more complex, vendors across the data centre industry are moving towards integrated power and cooling — making the rack itself a new infrastructure battleground.
Data centres used to be a fairly simple equation: put servers in racks, connect the power, provide enough cooling and keep everything running, that equation is changing. The problem is AI servers consume more electricity, with more power, more heat and more infrastructure being concentrated into the same physical space – the rack is quickly becoming the point where the data centre’s biggest challenges collide.
This is not just an nVent story – many companies are trying to build around that shift with equipment manufacturers, cooling specialists, electrical suppliers and infrastructure vendors increasingly moving towards the same idea: power, cooling and rack infrastructure needs to be designed as a connected system.
Is the rack is becoming a bottleneck
For operators, higher-density does not just mean buying more powerful servers – a rack drawing dramatically more power needs a different electrical architecture: the heat generated by that equipment may require liquid cooling, the physical space available for cables, power equipment and cooling infrastructure becomes tighter and maintenance becomes more complicated, then it all has to work – together.
AI workloads are pushing rack densities well beyond the levels many facilities were originally designed for, so operators must shift their thinking from whether the building can support the cluster, to whether the infrastructure around the rack can actually support it.
A wider shift across the industry
The convergence of power and cooling is clearly visible across the infrastructure market: Cooling companies are moving deeper into liquid distribution and rack-level systems, electrical infrastructure suppliers are developing higher-capacity power architectures for AI, Rack and enclosure manufacturers are integrating cooling, power and monitoring rather than treating the cabinet as a passive piece of equipment and the reason behind this shift is fairly straightforward: The AI rack is becoming too dense for systems to be designed independently.
This convergence is important for operators as it could ultimately make AI deployments much easier to standardise – instead of engineering every high-density rack as a collection of separate systems, the industry is moving towards more integrated, repeatable infrastructure – which matters.
Power and cooling can no longer be designed in isolation
Electricity and heat are actually two sides of the same problem – every additional kilowatt delivered to an accelerator eventually becomes heat that has to go somewhere. At conventional rack densities, the relationship could often be managed through facility-level cooling but at much higher densities, the rack itself becomes part of the thermal system, driving operators towards liquid cooling, higher-capacity power distribution and increasingly sophisticated rack infrastructure – nVent’s strategy is built around this convergence, their data centre portfolio already spans power distribution, liquid cooling, racks and related infrastructure but the more interesting issue is the problem those systems are intended to address: how do you make a very high-density rack behave like a reliable piece of critical infrastructure?
The retrofit problem is just as important
Not every AI deployment has the luxury of starting with a blank sheet of paper – operators likely have existing footprints with existing electrical connections, established cooling infrastructure and limited floor space, so the challenge becomes finding enough additional capacity without rebuilding the entire facility. This means rack-level infrastructure becomes increasingly important as the objective is not just higher density anymore – It is achieving that higher density without having to turn every deployment into a bespoke engineering project.
The AI rack is becoming a system
The next generation of AI infrastructure requires closer coordination between things that were historically treated as separate disciplines and the rack sits at the centre of these disciplines. Whilst the industry is focused how many GPUs can fit into a data centre, the harder question to ignore is how much AI can the infrastructure around each rack actually support – that’s the challenge operators are facing right now, and it may become just as important as the choice of GPU.
