Most data centers standing today were built for a world where a rack of servers drew a modest, predictable amount of power. AI has changed that. The same floor space now has to carry far more computing, and the building's power feeds and cooling were never sized for it. For an operator, that shows up as space that is technically vacant but can't be leased.
So we read Data Center Knowledge's piece on rising rack power with interest. The numbers are stark. The average rack drew about 6 kW in 2016 and draws about 26 kW today, and the figure more than doubled in the last two years alone. A single modern AI chip can draw up to 700 W. Google has published plans for racks of up to 1 MW, which we estimate is roughly forty times today's average.
Power in becomes heat out. Every watt a rack draws has to be carried away, and the way most rooms do that, with cold air blown across the servers, has a ceiling. Air holds little heat for the space it takes up, so past a certain rack density the fans and floor space needed to move enough of it stop making sense. According to Google, water carries about 4,000 times more heat than the same volume of air. That is why the densest racks are moving to liquid cooling, where coolant runs to the chips through pipes.
We agree with the article's direction: facilities that upgrade power and cooling ahead of demand will be the ones that can take the next wave of work. Our question is what to size for.
The average is a moving target
The average rack went from 6 kW to 26 kW in ten years, and from 12 kW to 26 kW in two. A cooling loop sized for today's average is undersized in two years. A loop sized for a 1 MW rack costs more up front, and an operator pays for capacity that may sit idle. Somewhere between those two, someone has to pick a number, and the article doesn't say which.
Where we come in
A liquid loop doesn't end at the rack. The coolant carries the heat to a coolant distribution unit, the cabinet that pumps coolant out to the racks, and a heat exchanger inside that cabinet hands the heat over to the building's water. That one component sees the full flow from every rack it serves, so its size sets how much rack load the loop can grow into.
Most heat exchangers are still built from stamped metal plates, which limits how much heat they can move in a given footprint. We 3D print ours, which lets us put more heat transfer surface in the same box. Our 1 MW heat exchanger, built for the Open Compute Project's Deschutes cooling design, is sized for one of the racks the article describes. Printing also lets us shape each channel for low pressure drop as well as heat transfer, so the added surface doesn't have to come with a bigger pump.
What operators should ask for
Ask a cooling vendor what rack load the loop is sized for, and what it costs to double it later. The heat exchanger is the cheapest place to buy that headroom now and the most expensive place to add it after the fact, and we're building ours with that in mind.