Cooling is part of
the computer now.

Every calculation consumes electricity, and essentially all of that electricity ends up as heat. For thirty years the answer was a heat sink, a fan, and a room kept cold enough to carry the load away. That works when a rack draws single-digit kilowatts. It stops working — as physics, not as engineering preference — somewhere in the tens.

ASHRAE classifies anything above 20 kW per rack as high density and points operators toward water. With containment, careful architecture and a tolerance for noise, air can be pushed to roughly 50 kW. Beyond that the fans start consuming a meaningful share of the facility's power just to move air, and past that point the better engineering move is to stop pushing air altogether.

Meanwhile the average rack has gone from about 8 kW to 17 kW in two years and is heading toward 30 kW by 2027. And the AI cabinets are not on that curve at all. A GB200 NVL72 sits around 130 kW; the Blackwell Ultra generation around 140. Vera Rubin is estimated at 190 to 230 kW — Supermicro's direct-liquid-cooled design for it is sized at 227 kW per rack. That is four to five times what air can carry — and liquid clears it in a single step.

A traditional datacenter holds thousands of separate servers that talk over a network. Each is still its own computer. That arrangement is fine for web serving and badly suited to a large AI model, where hundreds or thousands of processors must exchange intermediate results constantly. If they cannot talk fast enough, they wait — and idle silicon is the most expensive thing in the building.

It is the difference between seventy-two brilliant engineers in one room and the same seventy-two mailing each other letters.

Sixth-generation NVLink is the room. Each Rubin GPU moves data at 3.6 terabytes per second, with 260 TB/s of aggregate switch bandwidth across an all-to-all fabric, so any GPU reaches any other directly rather than queueing through a central bottleneck. NVIDIA rates the resulting cabinet at 3,600 PFLOPS of NVFP4 inference, with 20.7 TB of HBM4 at 1,580 TB/s.

Two things worth knowing. Those headline FLOPS are low-precision inference numbers, so they are not comparable with general-purpose computing performance. And the density exists because proximity is bandwidth: 3.6 TB/s per GPU is not a figure you can achieve across a room. The interconnect forces the packing, the packing forces the heat, and the heat forces the liquid. Each step follows from the one before — which is why solving the heat is what unlocked the whole design.

The mechanism is easier to hold onto if you stop thinking about “cooling” and simply follow the heat. A metal cold plate sits directly on top of each major processor. Liquid runs through fine channels inside that plate, picks the heat up at source, and carries it out of the rack. Because liquid transports heat far more effectively than air, the same cabinet can hold far more computing without the processors overheating and throttling.

The coolant then reaches a coolant distribution unit, which pumps and conditions it, and a heat exchanger that hands the heat across to the building's own water loop. The building finally releases it outdoors through dry coolers, chillers, or some combination. The same fluid circulates round and round; it is not consumed.

NVIDIA's Rubin-generation servers accept coolant arriving at up to 45°C — 113°F. That sounds like the opposite of cooling. It is the most important number in the system.

Heat only moves from hot to cold. If your coolant must be chilled to 20°C, then every hour the outdoor air is warmer than 20°C you need a refrigeration machine to make that cold — which is most hours in most places. If the coolant can be 45°C, outdoor air is almost always colder than the loop, and a dry cooler — effectively a very large car radiator — can dump the heat straight into the atmosphere with nothing but fans. Running the loop hotter is what buys the free cooling.

The leverage is steep. By NVIDIA's figures, raising chiller plant temperatures by a single degree cuts cooling energy costs by roughly 4%, and a 50 MW facility can save over $4 million a year in combined cooling energy and water costs. Note what that implies about scale: on a 50 MW site, cooling strategy alone is worth millions annually before anyone argues about the price of a GPU.

How far this goes depends on where you build, and NVIDIA is admirably specific about it. Its framing is that 45°C moves operators close to a chiller-less ideal where chillers may run just a few days a year — with an explicit caveat of roughly 1% of the year in some climates, and a blunt observation that a datacenter in the Scottish Highlands and one in Phoenix face very different realities. Climate is now a design input — and in the right climate, this is close to free performance.

This is the part of the story that deserves more attention than it gets. A closed-loop, dry-cooled facility can run on essentially no water at all — NVIDIA puts the reduction as high as 100% in favorable geographies. For an industry that has spent a decade defending its draw on local supply, that is not a marginal efficiency. It is the removal of a constraint.

The mechanism is just the closed loop doing what it says. Traditional cooling towers reject heat by evaporating water into the atmosphere, and every evaporated gallon has to be replaced — on NVIDIA's figures, roughly 2.6 million gallons per megawatt per year. Recirculate the same fluid instead and reject the heat through radiator coils, and that entire standing draw disappears. The liquid stays in the pipes; what goes to zero is the water consumed.

Read at campus scale the number is striking. A 100 MW site on evaporative cooling is drinking on the order of 260 million gallons a year. The same campus, closed-loop and dry-cooled, approaches nothing — while moving several times the heat, on less electricity, than the hall it replaced.

The trade is a modest one, and worth naming: more air-side equipment, more outdoor plot, and a stronger dependence on climate. But it is a favorable trade nearly everywhere anyone wants to build, and it turns water from a permitting risk and a standing operating cost into something close to a non-issue. More heat moved, less power to move it, and almost no water — an improvement with no obvious loser.

“Chiller-less” headlines read like bad news for cooling suppliers. The opposite is closer to the truth: the shopping list gets longer. Computer-room air handlers give way to coolant distribution units, pumps, manifolds, cold plates, liquid-to-liquid heat exchangers, leak detection and fluid controls, dry coolers — and, in hot climates, higher-temperature or backup chillers that still have to be bought and installed even if they run rarely.

The content is material. Reported liquid-cooling components run around $49,860 per GB300 NVL72 rack, rising about 17% to roughly $55,710 for the next generation. That is a per-cabinet line item on a product being deployed by the thousand.

The disclosed numbers show the demand arriving. Modine, whose Airedale business supplies coolant distribution units, heat exchangers and chillers, reported data-center revenue up 73% to $1.1 billion in fiscal 2026 — inside total company sales of $3,181.1 million — with the data-center line growing 158% in the fourth quarter alone. In May 2026 it announced an agreement reserving capacity for more than $4 billion of Airedale products across calendar 2027-2029, with the customer paying $165 million upfront to fund the necessary expansion.

Two details strengthen that picture rather than complicate it. First, an upfront payment to reserve capacity is a signal about supply scarcity as much as about demand — the customer is paying to be sure the factory exists. Second, the same period that produced the chiller-less messaging also produced multi-billion-dollar chiller commitments, because “rarely” is not “never” and peak-day design still governs what gets installed.

The gains do concentrate, though. Designs that cut chiller hours genuinely do pressure suppliers anchored to the old configuration, and rapid factory expansion carries its own execution and margin risk. The structural advantage sits with whoever can deliver the whole heat-removal chain rather than one conventional box in the middle of it.

The GPUs do the arithmetic. The CPUs organize the work. NVLink lets the processors behave as one machine. The networking joins rack to rack. The software manages failures and smooths the power swings that synchronized AI workloads inflict on a grid. And the liquid loop is what makes the physical arrangement possible at all. NVIDIA calls this extreme co-design, and the phrase is accurate: the pieces were engineered against each other, not assembled after the fact.

The modern AI computer is no longer a processor on a board. It is the rack, the pipes, the pumps, the switchgear and, increasingly, the building around them.

Which sets up the decade fairly clearly. The next generation of AI will not be won by whoever has the fastest chip. It will be won by whoever can move data, electricity and heat through an entire facility as though it were one machine — and by whoever owns the facilities where that is physically possible.