Liquid cooling dominates the headlines, but for the majority of server racks, air remains the logical choice. HPE warns that upgrading existing data centers to liquid cooling is often more expensive and complex than the hype suggests.
Anyone following the news around AI infrastructure might think air cooling is on its deathbed. Nothing could be further from the truth, according to HPE at Discover 2026. For lower-power racks, air cooling remains the most economical solution, and it will stay that way for decades. Think of point-of-sale systems, storage, transaction processing, and other workloads that don’t require high-bin GPUs or CPUs. According to HPE, it is simply nonsensical to cool that kind of hardware with liquid.
The message is remarkably down-to-earth for a manufacturer heavily invested in direct liquid cooling. “We don’t want to sell liquid cooling for the sake of liquid cooling itself,” they say. The reasoning is that only AI, HPC, and other high-density racks truly benefit from the switch. For everything else, air is cheaper, simpler, and more than sufficient.
Density grows, the number of servers per rack shrinks
Yet this doesn’t mean nothing is changing. As chip density increases, so does the power per component. A classic air-cooled rack that used to house twenty servers might only contain ten or even five in the future.
The cooling principle remains the same, but the amount of computing power you can physically fit in decreases. For customers who have been running air-cooled systems for twenty years and don’t want a migration path to liquid, that option will remain available—just with less hardware per cabinet.
Converting existing data centers is rarely the smart move
That is exactly where the problem lies in the upgrade discussion. Retrofitting an existing data center for liquid cooling is particularly unusual, according to HPE, and for good reason. It is invasive and expensive: operators have to take a large part of their installation offline, drill holes in walls, and add reservoirs. Many older data centers, whether they are ten, fifteen, or twenty years old, were never designed with a megawatt per rack in mind.
That’s where physics comes into play. Ten years ago, 100 to 200 kilowatts per rack was the norm, and 400 kilowatts was already considered the limit of what was feasible. Few data centers had the foresight to provide six- to eight-inch water pipes.
Those who want to scale up now run into pipe diameters, water flow rates, and a budget that rarely approaches the billion-dollar investments of the newest players. HPE is therefore trying to provide technology that fits into older infrastructure, rather than forcing customers to tear everything down.
A matter of budget balance
The core of the story is a trade-off between compute and facility. Customers want to maximize their investment in computing power, but must simultaneously modernize their facility. Those who put all their budget into the building infrastructure have nothing left for the actual compute, and vice versa.
For those starting from scratch, liquid cooling is easy to plan. For the large installed base of existing data centers, air cooling remains a viable and often wiser answer. The future is not black and white, but a mix in which old and new cooling methods will coexist for a long time to come.