Industries / Technology / Software and SaaS / Cloud and data centers
Cooling
This is the work of carrying heat away from the chips and out of the building, by air, water or other liquids, so that the machines stay within safe temperatures.
7
Milestones achieved, from 1990s to 2024.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1990s
Air wins, for a while
IBM's big computers had sent heat to the customer's building water since the System/360 Model 91 of about 1967, through a coolant distribution unit, a pump, heat exchanger and tank set between the building's water and the machine's own loop. Through the 1980s its 3081, 3090 and ES/9000 mainframes held their chips in water-cooled "thermal conduction modules", as heat per square centimetre of module rose from 0.3 W on the System/360 to 11.8 W on the ES/9000. In the mid-1990s IBM moved to CMOS chips, which used far less power, and its new machines were cooled by air alone; Computerworld reported that rising chip power brought water back as an option on the zEnterprise 196 in 2010.
LessonCooling follows chip power, and a liquid system dropped when a new chip technology cuts power returns when power climbs again.
Credit
IBM; the account is by Richard C. Chu, IBM Fellow
1992
Hot aisles and cold aisles
Racks are set out in rows that face each other in pairs, so that machines draw cool air from a shared "cold aisle" and blow hot air into a shared "hot aisle", instead of each row breathing in the exhaust of the row in front. The US Environmental Protection Agency's ENERGY STAR program dates the layout to IBM in 1992 and recommends it, citing cooling savings of 10 to 35% (Siemens's figures) from keeping supply and exhaust air apart; a 2021 history by Ian Seaton of Upsite Technologies places Robert Sullivan's first deliberate layout in the mid-1990s.
Credit
Robert F. Sullivan, IBM
2010
Servers in oil
Green Revolution Cooling, an Austin start-up, built a rack laid on its back and filled with a mineral-oil mixture that does not conduct electricity, and in April 2010 the Texas Advanced Computing Center (TACC) installed a pre-production unit. TACC submerged standard Dell servers with their fans removed and their disk drives sealed; its deputy director Dan Stanzione told Computerworld in 2011 that the oil, kept at about 105°F (41°C), "greatly reduces the amount of power", and that the prototype cooled a full 42U rack with about 250 W.
Credit
Green Revolution Cooling; Texas Advanced Computing Center
2011
Wider limits for server air
ASHRAE, the engineering society for heating and cooling, published its first recommended range for air entering data center equipment in 2004 (20 to 25°C), widened it to 18 to 27°C in 2008, and in 2011 added classes of allowable conditions; class A3 allows inlet air from 5 to 40°C (41 to 104°F). The same 2011 guidance set classes for liquid-cooled equipment, from W1, building water of 2 to 17°C that usually needs chillers, to W4, 2 to 45°C, which cooling towers or dry coolers can supply alone, so many sites could run most of the year without chillers.
LessonIn a 2016 study of nine hyperscale data centers by Rutgers University, Microsoft and GoDaddy, humidity mattered more than heat: disk failures rose in humid outside air, though spare copies of data absorbed them and the energy saved outweighed their cost.
Credit
ASHRAE Technical Committee 9.9
2018
AI runs the cooling plant
Google's system took a snapshot of each data center's cooling plant from thousands of sensors every five minutes, chose the settings that used least energy within safety limits and, from 2018, set them itself, with operators able to take back control at any time; Google said savings rose from about 12% at launch to about 30% on average after nine months. A peer-reviewed paper at NeurIPS 2018 by Google researchers described a related controller that learned to run the fans and water valves on a server floor after a few hours of safe trial moves and did better than the existing controllers.
Credit
Google DeepMind (Chris Gamble, Jim Gao) and Google Research (Nevena Lazic and others). The savings figures are Google's own
2022
A cooling fluid withdrawn
On 20 December 2022 3M said it would stop making PFAS, the long-lasting fluorinated chemicals, by the end of 2025, including its fluorinated fluids. Daniel Bizo of the Uptime Institute, a data center research and certification firm, wrote in 2024 that this "marked the beginning of the end" for 3M's Novec fluids used in liquid cooling, with major implications for operators committed to two-phase immersion, in which servers sit in a fluid that boils off the chips; interest in two-phase cooling has since moved to cold plates on the chip.
LessonA cooling design that rests on one maker's chemical carries that chemical's legal and regulatory risk.
Credit
3M (the decision); Uptime Institute (the assessment)
2024
AI racks need liquid
NVIDIA's GB200 NVL72, a rack of 72 GPUs (the chips that run AI models) that works as one machine, is built with liquid-cooled compute and switch trays fed through manifolds and quick-disconnect couplings. NVIDIA puts the cooling the rack needs at 120 kW, and in October 2024 it gave the rack and tray designs to the Open Compute Project, an industry group that shares hardware designs. Outside engineers now plan around it: LiquidStack's chief executive says Blackwell systems take inlet water of up to 40°C, and Compass Datacenters designs new buildings for 80% liquid and 20% air cooling.
Credit
NVIDIA
Watch for
Coolant inside the chip
Microsoft says microfluidic cooling, in which coolant flows through tiny channels etched into the back of the chip itself, removed heat up to three times better than cold plates in lab tests and cut the maximum temperature rise of the silicon in a GPU by 65%; the test server ran core services for a simulated Teams meeting, and the channel pattern was designed with AI help by the Swiss start-up Corintis. Microsoft has set no date for production use.
Achieved whenan outside lab or customer publishes measurements, or chips with etched channels run in production data centers.
Claimed by
Microsoft, September 2025
AI halls without chillers
NVIDIA's chief executive Jensen Huang said at CES in January 2026 that its Vera Rubin systems run on 45°C water and "At 45 degrees Celsius, the data center doesn't need a chiller". In June 2026 NVIDIA described fanless Rubin racks cooled entirely by a mix of 75% water and 25% propylene glycol entering at 45°C and leaving at about 55°C, with chillers needed perhaps 1% of the year in some climates. Outside engineers disagree on how much changes: LiquidStack's Joe Capes says chillers may still be needed in Rubin halls, and Vertiv found that over 90% of North America's top data center markets see design temperatures of 40°C or more.
Achieved whenan operator outside NVIDIA publishes a year of data from a Rubin hall run without chillers, with the hours of any chiller use reported.
Claimed by
NVIDIA, January and June 2026
Liquid added to air halls
AWS says it designed its own direct-to-chip liquid cooling, a closed loop with a cold plate on each chip, because no vendor system it evaluated fit its needs. It went from whiteboard to prototype in four months and to the first production unit in 11 months, and the system can be added to existing air-cooled data centers where it is needed without raising water use.
Achieved whenan outside party reports measured cooling capacity and energy use for halls that mix it with air cooling.
Claimed by
AWS, June 2025
Open
01
Cold plates on one rig
A public test of at least six commercial direct-to-chip cold plates, the metal blocks with internal channels that clamp onto a chip and carry coolant over it, all run on one thermal test vehicle (a heater block the size of a processor) under the Open Compute Project's new cold plate test specification, publishing thermal resistance and pressure-drop curves for each.
The briefCredit · unclaimed02
Quick couplings, drips counted
An independent endurance test of the hand-operated quick-disconnect couplings that join servers to liquid-cooling manifolds, built to the Open Compute Project's Universal Quick Disconnect (UQD) form, from at least five makers: thousands of connect and disconnect cycles each, with the liquid spilled at each disconnection weighed, and plugs of one maker mated with sockets of another.
The briefCredit · unclaimed
Open 01 · the brief
Cold plates on one rig
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of at least six commercial direct-to-chip cold plates, the metal blocks with internal channels that clamp onto a chip and carry coolant over it, all run on one thermal test vehicle (a heater block the size of a processor) under the Open Compute Project's new cold plate test specification, publishing thermal resistance and pressure-drop curves for each.
Why it is still open
The specification, effective 3 September 2026, says an operator "currently has no reliable basis for determining whether two cold plates were held to equivalent standards". A 2024 ASME study found that much of the scatter in reported thermal resistance comes from the test vehicles themselves. The nearest comparison, JetCool's March 2025 "third-party" test on NVIDIA H100 GPUs, was presented by the maker, does not name the rival plates and reports the whole stack including the thermal paste.
Built with what exists
A resistive heater block with a copper top, built after the standard test vehicle of Martinez, Caceres and Ortega and kept under 500 W; a benchtop recirculating bath holding water at 45°C, the specification's typical test temperature; a turbine or Coriolis flow meter; a differential pressure sensor; type-T thermocouples; a data logger such as a LabJack; Python for the curve fits.
First step
Read the specification's thermal and pressure-drop sections, build the heater block, and measure one consumer CPU water block to shake down the rig.
Done looks like
Six cold plates, each tested at six or more flow rates from 25% to 150% of rated flow and mounted three separate times, with measurement uncertainty stated, raw data on Zenodo, and at least one plate retested by a second lab within that uncertainty.
Open 02 · the brief
Quick couplings, drips counted
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An independent endurance test of the hand-operated quick-disconnect couplings that join servers to liquid-cooling manifolds, built to the Open Compute Project's Universal Quick Disconnect (UQD) form, from at least five makers: thousands of connect and disconnect cycles each, with the liquid spilled at each disconnection weighed, and plugs of one maker mated with sockets of another.
Why it is still open
The UQD specification, version 2.0.0 of 22 September 2026, sets a minimum of 5,000 coupling cycles and a maximum fluid loss of 0.02 to 0.07 mL per cycle, depending on size, and requires each supplier to cross-test its couplings with at least three other suppliers. The cross-test reports are made "available to customers upon request", not published, and spillage figures in public come from the makers: Danfoss's datasheet promises "no spillage" and 100% helium-leak testing in its own factory.
Built with what exists
An electric linear actuator with a load cell, driven by an Arduino or Raspberry Pi; water with propylene glycol at room temperature and about 2 bar; a 0.001 g laboratory balance and absorbent pads; a camera filming each run; Python for logging.
First step
Buy one pair of UQD couplings, build the actuator rig and run 500 cycles with water, weighing the drips.
Done looks like
Five coupling models, three samples each, 5,000 cycles each, spillage weighed every 500 cycles and set against the specification's limit, connection force logged, at least one mixed pair from two makers run the same way, and all data and video published, with any coupling that leaks named with its cycle count.