Cooling a 1 MW Rack

For most of the history of the data centre, "free cooling" was treated as a luxury. It was the sort of thing an operator installed to feel virtuous, or to trim a little off the electricity bill in a mild winter. That era is over. As Alejandro Thomas, an application engineer at Vertiv, one of the industry's larger suppliers of critical infrastructure, put it in a recent ATA Insights webinar: "Before, free cooling was seen more as a caprice. Now it is seen as a necessity and a requirement for every project."

The reason for the change can be summed up in three words that Mr Thomas heard, with some alarm, at the Madrid Tech Show a couple of weeks before the webinar: one-megawatt rack. NVIDIA is building racks whose power density is climbing so fast that the machines cannot simply be cooled by blowing cold air past them. The question is no longer whether to bother with efficient cooling. It is whether the cooling can keep a rack alive at all.

What free cooling actually is

Strip away the jargon and free cooling is a modest idea: cool the building using the weather outside, so you do not have to run a compressor. Compressors are the thirsty part of any air-conditioning system; a mechanical chiller spends most of its electricity spinning them. If the air outside is already cold enough to do the job, running a compressor to make cold air is a bit like buying bottled water while standing in the rain.

Mr Thomas divides the approach into two flavours. Direct free cooling pushes filtered outdoor air straight into the server hall. It is cheap and simple, and it works best on the direct-expansion units that guzzle the most power. But it has limits at both ends of the thermometer. Too warm outside and there is no benefit. Too cold — below about five degrees, or in an Icelandic winter — and the air can damage the equipment, so the system flips back to compressor mode. Indirect free cooling, by contrast, uses cold outdoor air to chill a fluid through a heat exchanger. It never lets outside air into the hall, which makes it usable across a far wider temperature range, at the cost of being slightly less efficient because a heat exchange always loses something.

None of this is new. What is new is the heat.

The heat is winning

The old rule of thumb for data-centre water loops was to run them cold — seven to fifteen degrees Celsius, in Mr Thomas's telling. Cold water cools reliably, but it is expensive to make, because the outside air is rarely colder than that, so the compressor runs almost all the time. The industry has spent years quietly nudging those temperatures up: fifteen to twenty degrees, then twenty to thirty, and for the newest AI projects, thirty to forty.

This sounds counterintuitive — surely hotter water cools worse? — but it is the whole trick. The warmer you are willing to let the water run, the more hours of the year the outside air is colder than that water, and the more hours you can cool for free. "The higher those temperatures are," Mr Thomas explained, "the more beneficial it is," because it lets the chiller lean on the weather instead of the compressor.

There was a catch, and it was technological. Chillers that could run free cooling at high water temperatures did not exist. On the chart Mr Thomas showed — outdoor design temperature up one axis, water temperature along the other — the top-right corner, hot air and hot water, was simply blank. No product lived there. That happens to be exactly the corner into which AI is now dragging everyone. Vertiv, he said, saw the gap coming and spent the past couple of years developing a chiller — internally a "free-cooling chiller," with a magnetic-levitation compressor rather than the older screw type — designed to work across that whole range. It is not yet on sale, but it is being specified into projects that expect it within months.

A tale of two circuits

To show what the extra headroom is worth, Mr Thomas walked through a worked example, run with colleagues: a new 12 MW data centre in Lisbon — large, but representative of the projects now being planned, and, being coastal and humid, a deliberately unflattering climate for free cooling.

The design splits the heat load roughly 75/25 between liquid cooling and air. That ratio matters, and it is a useful corrective to the fantasy that liquid cooling makes air conditioning obsolete. Even in a dense AI hall, Mr Thomas noted, a quarter of the heat cannot be captured by the cold plates — either because of how the plates are sized, or because other kit in the room keeps radiating heat regardless. The air-cooling problem does not go away; it merely shrinks.

The first design uses a single circuit: one conventional chiller, capped at around 20-30°C, feeding both the liquid- and air-cooling systems. The second uses two circuits — a high-temperature free-cooling chiller dedicated to liquid cooling, running at temperatures as extreme as 37°C to 47°C, alongside a normal free-cooling chiller for the air side. On paper the difference in overall PUE looks small. But, as Mr Thomas was careful to point out, "the one is a mathematical limit, not a physical one" — you cannot push power usage effectiveness below 1.0, so small-looking gains near the floor represent large real savings.

The savings are large. Looking only at the chillers — the single biggest energy consumer in the plant — the dual-circuit design cut consumption by 66%. "It doesn't mean we save on the CRACs, or the liquid cooling, or anything else," he cautioned. "This is chillers only. But it's a very powerful figure, because at the end of the day the equipment that consumes the most is the chillers." And that 66% is the humid-Lisbon number. In a dry climate such as Madrid, at similar temperatures, the figure climbs higher still, because humidity is free cooling's other great enemy: coastal, salty, moist air can force a unit to fire up its compressor just to dehumidify, undoing the point of the exercise.

Let the machine decide

An audience question got at a practical worry: who actually runs the thing? Can an operator force free cooling on when it looks cold enough outside? Yes, said Mr Thomas — and no, do not. "It's not recommended, because you might be forcing the machine to operate in a regime" of high humidity or unsuitable conditions that it was never meant to handle. Vertiv's preference is to bolt the free-cooling module onto the unit itself and let the unit's own controller decide, using three probes — one sensing outdoor temperature and humidity, two more on the return and supply air — to work out, moment by moment, how much free cooling to allow. If the incoming air is too damp, the controller closes the flap, runs the compressor for a set period, and tries again later. The alternative — a free-cooling system wired to a separate controller on the far wall, blind to what the main unit is doing — is a recipe for the two halves fighting each other.

There is an honest limit to how far this goes, and Mr Thomas did not pretend otherwise. Asked whether Vertiv's new chiller had been installed anywhere yet, he was blunt: "I can't tell you it's been applied anywhere, because it really hasn't." It goes on sale in the coming months. Asked about the price premium of a free-cooling chiller over a plain one, he declined to invent a figure, offering only that it "might be around 10%" and that in the Iberian interior the payback is reliable enough that "we always fit it, because we'll always recoup that investment." On the coast — Málaga, the Canaries, Barcelona, Lisbon — it is a case-by-case sum.

The one-second problem

The webinar kept circling back to why any of this suddenly matters so much, and the answer is the AI rack. NVIDIA's densities have leapt from roughly 40 kW a rack in the H100 generation to 120 kW with today's Blackwell systems, and its Rubin Ultra "Kyber" racks, due in 2027, are rated at 600 kW and climbing toward the megawatt that so struck Mr Thomas. Vertiv is one of a score of infrastructure firms — alongside Schneider, Siemens, ABB and others — building the plumbing and switchgear for that world.

At those densities the engineering problem changes character. AI loads are not steady; they can swing from 20% to 100% in moments. That volatility, not the average heat, is the real challenge, because the cooling has to stay stable through the swings. And the stakes of instability are no longer measured in wasted kilowatt-hours. "If there's a second, or half a second, where the cooling equipment doesn't work" on a one-megawatt rack, Mr Thomas said, "is the rack going to blow up? We don't even know what will happen, because a megawatt is such a huge power density we can't even imagine it." Hence the industry's growing interest in precise, predictive control — algorithms that get ahead of a fault rather than mopping up after it.

Which is the quiet irony running under the whole hour. The heat that is breaking the old cooling economics comes from artificial intelligence; the tool being reached for to keep the new machines from cooking themselves is, increasingly, artificial intelligence too.

Will free cooling become the European default, or stay a climate-dependent niche? Mr Thomas would not claim every project, "because there's always some project that gets away." But the direction seems settled to him — driven not only by technology but by decarbonisation rules and energy-efficiency certificates that now help fund the upgrades. "More and more it's going to be an almost obligatory requirement." For an idea that spent decades dismissed as a caprice, that is quite a promotion.

Alejandro Thomas's full presentation — including the Lisbon case study and the free-cooling charts — is available on demand from ATA Insights. The thread runs straight into RENMAD Datacenters, whose next edition takes place in Zaragoza, where cooling, power supply and efficiency for the AI-era data centre will be centre stage.