Data centers built for the cloud era are running into a wall that fans and cold air cannot fix. Nvidia's B200 chip dissipates up to 1,200 watts of heat on its own.
A full GB200 NVL72 rack pulls 120 to 132 kilowatts, roughly seven to nine times what a traditional air-cooled rack was designed to handle.
Liquid cooling, a technology data centers largely phased out decades ago in favor of cheap airflow, is now coming back because physics leaves no other option.
Why Air Cooling Hits a Physical Wall With AI Chips
Air has always cooled data centers by moving cold air across server components and exhausting the heated air elsewhere. That approach worked fine when racks drew 5 to 10 kilowatts.
Conventional raised-floor cooling with hot and cold aisle containment tops out around 20 to 30 kilowatts per rack before hot spots form even with the best airflow engineering.
Modern AI training racks blow past that ceiling by a wide margin. The jump comes from chip-level power draw, not just server count. Nvidia's A100 pulled 400 watts per chip in 2020, the H100 pushed that to 700 watts, and the B200 Blackwell chip now hits 1,000 to 1,200 watts.
ASHRAE's technical committee on data center equipment now recommends direct liquid cooling above 20 kilowatts per rack, a threshold set by heat-flux physics rather than vendor preference.
Water carries heat away roughly 3,500 times more efficiently by volume than air, and its thermal conductivity runs about 23 times higher. That gap is why a cold plate pressed against a chip removes heat at a rate air simply cannot match, regardless of how much airflow gets thrown at the problem.
How Direct-to-Chip Cooling Actually Works
Direct-to-chip cooling replaces the idea of cooling a room with the idea of cooling a component. A metal cold plate sits directly on top of the GPU or CPU, and coolant flows through microchannels etched into that plate, absorbing heat right at the source before it ever spreads into the surrounding air.
A coolant distribution unit, or CDU, manages the temperature, pressure, and flow rate of that liquid across an entire row of racks. On a typical Nvidia GB200 deployment, coolant enters around 25°C and exits near 45°C after pulling heat from dozens of chips in sequence.
Sensors throughout the loop track flow rate and pressure to catch leaks before they become a facility-wide problem.
This is the same rack-scale approach Nvidia ships by default on GB200 NVL72 systems, which pack 72 Blackwell GPUs and 36 Grace CPUs into a single liquid-cooled cabinet. There is no air-only configuration for that hardware, because the chips were designed around the cooling method rather than adapted to it afterward.
Immersion Cooling for the Densest AI Clusters
A smaller but growing share of operators skip cold plates entirely and submerge whole servers in a bath of non-conductive dielectric fluid. Immersion cooling handles extremely uniform, extremely dense fleets well, since every component touching the fluid gets cooled at once rather than only the chips with cold plates attached.
Direct-to-chip still commands the larger share of new deployments because it slots into existing rack architecture with fewer structural changes, while immersion tanks require a different physical layout entirely. Most operators now run hybrid setups, using hyperscale data centres as testbeds for whichever combination proves cheapest to retrofit at scale.
Google, Microsoft, and Nvidia Are Racing to Go Liquid
TrendForce projects liquid cooling penetration among AI chips will rise from about 33 percent in 2025 to 53 percent in 2026, approaching 60 percent in 2027. Google already runs liquid cooling across more than 80 percent of its AI servers, built on years of experience customizing infrastructure around its own TPU accelerators.
Microsoft began fleet-wide direct-to-chip deployment across Azure campuses in mid-2025 and is now testing microfluidic channels etched straight into future chip packages. That approach, sometimes described as cooling from the silicon up, aims to remove heat before it can even reach the surface of the chip package.
Nvidia's own hardware roadmap has effectively settled the debate for the rest of the industry. The company's next-generation Nvidia chip architecture, code-named Rubin, is built assuming liquid cooling from the start, and running it in an air-cooled facility would not just be inefficient. It would not work.
The Efficiency Gains and Water Tradeoffs Nobody Advertises
Liquid cooling is not just about keeping chips from throttling. The US Department of Energy estimates cooling already accounts for up to 40 percent of total data center energy use, so shifting to a more efficient heat-removal method directly cuts operating costs alongside preventing hardware failure.
The tradeoff that gets less attention is water. Closed-loop liquid systems recirculate coolant rather than consuming it, but the chillers that reject heat from that loop into the outside environment often still rely on evaporative cooling towers that do use water.
Newer membrane-based systems in humid, tropical climates are being designed around near-zero water loss to avoid straining local supplies, a response to years of criticism over how thirsty chip manufacturing and the data centers running those chips have become.
What Comes After Blackwell and Rubin
Engineers are already looking past today's single-phase liquid cooling toward two-phase direct-to-chip systems, where coolant partially boils at the chip surface and absorbs far more heat during that phase change than a liquid staying in one state ever could. Industry analysts expect early two-phase announcements before the end of 2026.
Next-generation platforms are also pushing rack density toward 500 kilowatts, a scale where even today's direct-to-chip cold plates will need redesigning. Meeting that kind of sustained power draw eventually runs into a separate constraint entirely.
That is why some operators are now watching next-generation fusion reactor output as a longer-term answer to keep pace with AI's appetite for both electricity and cooling capacity.
Liquid cooling's return is not a retro engineering trend. It is a direct consequence of chip power density outrunning what moving air can physically remove.
As GPUs keep climbing past 1,000 watts each, the cold plate and the coolant distribution unit are becoming as fundamental to data center design as the server rack itself, with two-phase systems and denser architectures already queued up behind them.
Frequently Asked Questions
Why is liquid cooling coming back to data centers now?
AI chips like Nvidia's B200 generate up to 1,200 watts of heat each, far beyond the 20 to 30 kilowatt-per-rack ceiling air cooling can handle. Liquid removes heat far more efficiently than air, so it has become mandatory rather than optional.
What is direct-to-chip liquid cooling?
It is a method where a cold plate sits directly on a GPU or CPU and coolant flows through it, pulling heat away at the source. A coolant distribution unit manages temperature and flow across the whole rack.
Is immersion cooling different from direct-to-chip cooling?
Yes, immersion cooling submerges entire servers in dielectric fluid while direct-to-chip only cools individual components through cold plates. Direct-to-chip currently has broader adoption because it fits existing rack designs more easily.
Does liquid cooling use more water than air cooling?
Not necessarily, since closed-loop coolant recirculates rather than being consumed, though the chillers rejecting that heat outdoors sometimes still rely on evaporative towers. Newer membrane-based designs are cutting that water use toward zero.
Which companies have adopted liquid cooling the fastest?
Google leads among major cloud providers, with liquid cooling across more than 80 percent of its AI servers. Microsoft began fleet-wide Azure deployment in mid-2025, and Nvidia now ships liquid cooling by default on its top-end racks.
What happens after today's liquid cooling systems reach their limit?
Engineers are developing two-phase direct-to-chip systems, where coolant boils at the chip surface to absorb more heat during that phase change. Early two-phase announcements are expected before the end of 2026.