For four decades, air was enough. A server cabinet in the 1980s dissipated two or three kilowatts; when dense racks arrived the figure climbed to five or eight, and the industry solved it with bigger fans, properly separated hot and cold aisles and airflow containment. Air was cheap, familiar and, above all, sufficient. GPUs broke the sequence: Nvidia’s A100, released in 2020, already asked for 400 W per chip, the H100 went up to 700 W, the B200 touches 1,000 W and a Blackwell NVL72 rack —72 GPUs and 36 CPUs in a single cabinet— draws between 120 and 130 kW. Manufacturers rate GB300 variants at 132 to 155 kW depending on configuration, and the successor, Vera Rubin NVL72, is already aiming at the 200 kW class per rack.
The limits of air are not an opinion: they are thresholds the industry handles with names and numbers. Up to roughly 20 kW per rack, air holds. Fit a heat exchanger on the rack’s own rear door and you stretch it to 100 kW. Above 175 kW, the only sensible path is to submerge the hardware in liquid. In between lies the band where the real game is being played: direct-to-chip liquid cooling, between 100 and 175 kW.
The problem is not the heat but how much air it takes to carry it away
The underlying reason sits in a material property. A cubic metre of air at normal pressure stores about 1.2 kilojoules per degree; a cubic metre of water stores roughly 4,180. Water therefore carries almost 3,500 times more heat per unit of volume. Translated into the flow rates of a real rack: removing 100 kW across a 10 K temperature rise requires almost 10 kg of air per second, about 8.3 m³/s, on the order of 30,000 cubic metres per hour through a single cabinet. With water, the same 100 kW and the same 10 K rise are satisfied by some 2.4 kg/s: roughly 8.6 m³/h, a pipe you can hold between two fingers. Accept a larger rise of 15 K and the flow drops below 6 m³/h. The difference is not a matter of degree; it spans three orders of magnitude.
There is a second gap, the one separating hot metal from the fluid cooling it. A fan blowing air across a chip surface achieves a film coefficient around 50 watts per square metre per degree. A coolant flowing through the copper microchannels of a cold plate reaches roughly 15,000: some 300 times more. It is the difference between fanning a hot pan with a magazine or dropping the magazine into ice water. Both move heat; only one keeps up.
From there come the effects you actually see in production. A poorly cooled H100 can drop clock speed within seconds of starting a sustained training run, and in a job spread across thousands of cards a single throttled chip holds back the rest. The fans needed to shift that volume of air consume a sizeable share of the equipment’s own power, make noise and create hard-to-spot hot spots. The US Department of Energy estimates cooling can account for up to 40% of a data centre’s electricity use in a traditional design: a figure that translates directly into the bill and the carbon footprint.
Cold plates: 45 °C water that never touches the electronics
The dominant solution does not submerge anything. It is a cold plate: a copper block with microchannels machined inside, bolted onto the GPU or CPU package, with a layer of thermal interface material —paste, or better an indium sheet— in between to fill the microscopic irregularities between metal and metal. The liquid runs inside the block and never meets the die.
Around the plate there is as much engineering as inside it. The rack loop connects through hoses and spill-free quick-disconnect couplings —the open UQD specification, driven by Intel and adopted by the Open Compute Project— and lands in a CDU, the coolant distribution unit that regulates flow, pressure and temperature and keeps the rack fluid separate from the facility fluid. Along the way there are manifolds, temperature and conductivity sensors, filters, biocides and corrosion inhibitors, plus a leak-detection cable that, on sensing liquid, stops the pumps and closes the main valve in under two seconds. The usual coolant is de-ionised water or 25% propylene glycol, with a service life in a closed loop that can reach ten years.
This year’s shift is temperature. Nvidia has specified that its entire MGX rack architecture accept 45 °C inlet water: the facility supplies around 41 °C to the CDUs, they deliver 45 °C to the racks and the return leaves at about 55 °C. With water that warm, a rooftop dry cooler —a radiator driven by ambient air— is enough to reject the heat for most of the year without switching on compressors. That is where the economics lie: mechanical refrigeration is one of a data centre’s largest loads, and the 45 °C platform keeps it out of the picture much of the time. The result is a PUE —the ratio of total energy to energy delivered to IT equipment— in the region of 1.1, and up to 10% more racks within the same electrical budget. Vera Rubin is already designed as single-phase direct liquid cooling at 45 °C with every compute and networking tray liquid-cooled; GB200 and GB300 remain hybrids, shedding 10 to 15% of their heat to air. Lenovo has been at this since 2012 with its Neptune platform, which for years was an oddity.
Immersion: dunking works, but the fluid has become complicated
The other end of the catalogue is putting the servers inside the liquid. In the single-phase version the hardware bathes in a dielectric hydrocarbon or synthetic ester; an external heat exchanger cools the bath and convection moves the fluid. Servers have to be adapted —no fans, and mechanical drives are not submerged— but efficiency and rack density are hard to match.
The two-phase variant is more elegant on paper: the liquid boils at around 50 °C in contact with the chips, the vapour rises and a coil at the top of the tank condenses it back into liquid. A phase change absorbs a great deal of energy and the system self-regulates by temperature. The problem is not thermal but chemical: the fluids that made this possible are per- and polyfluoroalkyl substances, PFAS, the family known as “forever chemicals”. 3M stopped making its Novec line —production ended in 2025 as part of its exit from all PFAS—, the US Environmental Protection Agency has classified them as hazardous substances and the European Chemicals Agency is preparing a broad restriction. The substitutes arriving on the market add more PFAS to the equation. Against that backdrop, two-phase immersion has stalled just when demand was highest, and the market has pushed towards single-phase cold plates and single-phase immersion.
The heat is no longer thrown away
When a data centre loop leaves at 55 °C, the conversation reverses: that is not waste, it is usable heat. In Milan, Equinix and the utility A2A signed an agreement in July 2026 to feed up to 225 GWh of thermal energy per year from a data centre into the city’s district heating network, equivalent to about 20% more decarbonised heat distributed across the city. Similar examples exist in Stockholm, Helsinki, Paris and Frankfurt, with a caveat: the hotter the water leaves, the easier it is to place in a district network, which is why architectures running at 45 °C are also the easiest to recycle. In Europe the Energy Efficiency Directive pushes operators to assess and report that recovery, and in Germany the energy efficiency act requires it with concrete thresholds. As a bonus, a closed loop evaporates no water: the water consumption that is enormous in traditional cooling towers drops to almost nothing.
What comes next: channels inside the silicon itself
Cold plates have a ceiling too. Cooling a large surface is easy; cooling a one-square-millimetre spot where hundreds of watts concentrate is not. Next-generation chips approach 1,000 W/cm² at their hot spots, and there the thermal interface and the silicon itself become the bottleneck.
The answer under research is to put the coolant inside the chip. Microchannels are etched into the silicon’s backside using the same deep-etch techniques used to build MEMS, and the liquid flows a few microns from the transistors. TIM-free architectures remove the paste layer entirely so the coolant touches a heat sink integrated into the chip. Labs such as imec have demonstrated impingement coolers that raise chip temperature by less than 15 °C for every 100 W/cm² applied, and embedded microchannels have sustained densities above 1,000 W/cm² in the laboratory. What remains unsolved is manufacturing at scale, ten-year reliability and the small detail of every package carrying pressurised liquid inside it. When that arrives, the data centre will look a lot less like a room full of computers and a lot more like precision plumbing: computing has become a heat problem, and heat is solved with water.
