Hybrid Cooling in a Data Center Leaves 20% to 43% on Air

Upsite Technologies, as quoted by Data Center Knowledge in June 2024, says even the most advanced liquid cooled data centers will probably still need “at least 20% to 30% of air cooling.” That sentence states the case for a hybrid cooling data center, where air and liquid share the same rooms. It holds for cold plate racks fed with cool water. At warmer supply water the one measurement in the record puts the room’s share at 41% to 43%, so the claim works better as a floor than as a typical value.

Cold plate racks leave between 20% and 43% of server heat in the room under load

Direct liquid cooling (DLC) in this piece means cold plates on the processors, in some designs on memory too, with coolant carried to the rack. Whatever the plates do not touch leaves through server fans into room air.

The Open Compute Project white paper on a modular technology cooling system for cloud scale says many liquid cooling solutions capture 70-80 percent of total IT system heat, and that a typical cold plate system captures 70%-80% of the IT equipment heat. Heat the liquid misses needs computer room air handlers, fan walls or other air cooling. The same section adds that a cold plate system paired with a door heat exchanger could capture close to 100% of the IT heat.

Source: Open Compute Project, “Modular Technology Cooling System for Cloud Scale,” section 3.6 (Rev 0 published 14 October 2024; workstream input for Rev 1)

The only measurement in the set comes from Lawrence Berkeley National Laboratory. In report LBNL-6641E, Coles and Greenberg put 38 Cisco servers with Asetek cold plates on CPUs and memory into one rack, with two unmodified servers in the next rack as a baseline. Capture is a percentage of the power drawn at the cords, normalized to a 28°C server air inlet.

The best results ranged from 67 percent to 75 percent, at supply water of 15°C to 20°C, and were not significantly affected by server power level.

With supply water at 30°C, capture fell to between 35 percent at idle and 59 percent at full power.

Source: Lawrence Berkeley National Laboratory, “Direct Liquid Cooling for Electronic Equipment,” LBNL-6641E (March 2014), Project Results

Table 1. Heat left in room air by cold plate cooling, by condition (% of server power)

* 100 minus the liquid share, calculated for this table. LBNL rows: percent of cord power, 38 servers with CPU and memory cold plates, normalized to 28°C server air inlet. OCP rows: guidance values, no test conditions stated.

Source: LBNL, “Direct Liquid Cooling for Electronic Equipment,” LBNL-6641E (March 2014); Open Compute Project, “Modular Technology Cooling System for Cloud Scale” (Rev 0, 14 October 2024)

The OCP range and the LBNL best case line up, and both describe cool supply water. LBNL reports that the fraction captured rises as supply water gets colder, and its abstract notes that supply water for this kind of system can run well above what facility chilled water loops commonly provide, with compressor-free cooling possible.

The efficient plant and the high-capture plant pull in opposite directions. In this prototype, 30°C water moved the room’s share from the 25% to 33% band to 41% to 43% at load.

A liquid row holding a tenth of the load removes only 5.9% to 8% of a hall’s air duty

Phased adoption changes the air load by less than the liquid share suggests, because liquid takes only a fraction of what it touches. Let f be the share of IT load in cold plate racks and c the share of that load the liquid captures. For 1,000 kW of IT load, the air load is 1,000 kW × (1 − f × c).

Two values of c bracket the sources: 0.59, the LBNL capture at 30°C supply water and full power, and 0.80, the top of the OCP range.

At f = 0.10 the air load is 941 kW with c = 0.59 and 920 kW with c = 0.80.

At f = 0.25 it is 852.5 kW and 800 kW.

At f = 0.50 it is 705 kW and 600 kW.

These are scenario values, not measurements. They assume every liquid rack captures the same share and the air-cooled remainder of the hall does not change.

Uptime Institute’s 2024 Cooling Systems Survey shows where operators sit on the f axis. Of 964 survey respondents, 22% of those asked about DLC use (n=453) said their organization makes some use of it, and nearly half of those users said fewer than 10% of their IT racks incorporate it.

Asked to describe current use, 30% of DLC users chose only a few servers and 47% chose some IT racks. 17% said they make substantial use of DLC, and 5% said one or more data centers predominantly use it.

Source: Uptime Institute, “2024 Cooling Systems Survey: Direct Liquid Cooling Results” (May 2024)

The survey counts racks, not kilowatts. Liquid racks usually run denser than the rest of the hall, so load share likely exceeds rack share, but the survey does not report it and the f values above are assumptions. A facility at the low end of the rack distribution sees its air duty fall by single digits, and the chillers, air handlers and containment behind that duty stay in service.

The OCP paper supplies the other end of the line. Its example takes a 1200 A busway at 415 V, three phase, unity power factor, rated at about 860 kW.

Multiplying 1,200 A × 415 V × √3 gives 862.6 kW, consistent with the rounded value.

The loop is sized at 80% of the busway rating, which is 688 kW, and the remaining 172 kW is air even where every rack is liquid cooled.

Across the range from f = 0 to f = 1 the air plant never reaches zero, so a hall in its first liquid phase is hybrid by arithmetic, not by preference.

A 50 kW cold plate rack still hands the room 10 kW to 20.5 kW of heat

Take a 50 kW rack, the size ASHRAE TC 9.9 uses in its fan power example. The heat left to air is 50 kW × (1 − c).

At c = 0.59 that is 20.5 kW. At c = 0.75, the top of LBNL’s best case, it is 12.5 kW. At c = 0.80 it is 10.0 kW.

In the same Uptime survey, 820 respondents were asked at what rack density air cooling becomes so costly, or unable to meet requirements, that DLC becomes necessary. The largest group, 29%, chose 20-29 kW per rack.

A further 19% chose 30-39 kW. The residual from a warm-water cold plate rack, 20.5 kW, lands in the band the largest group names as air’s limit.

That comparison is a judgment, since survey opinion is not a thermal limit. It does show that a liquid rack on warm water is not an air-free rack.

Where a room has no water at the rack, the sources describe stages that keep final heat rejection in air. ASHRAE TC 9.9’s white paper says such sites might use liquid-to-air heat exchangers, in the rack or at the row, for an interim period. They are less efficient than rejecting heat to facility water, but they bridge the gap until facility water headers go in and liquid-to-liquid exchangers replace them. A rear door heat exchanger is one form of that stage.

Uptime Intelligence describes a smaller version, self-contained air-assisted liquid cooling inside the server. It needs no coolant distribution units or facility water, but deploying it in existing space will not expand cooling capacity. Selecting it can reduce IT fan power or support roughly 20% greater TDP than the same server could otherwise carry.

Uptime reports internal HPE testing in which the pumps and fans of a 1U unit used around 40% less power than the server fans of an air-cooled equivalent. It notes that systems outside HPC and AI deployments typically support up to 1.2 kW of load per 1U server.

Uptime also says these systems are not currently available for GPU servers and suit a transitory measure better than a destination.

These stages let a room start without a water retrofit, but its air plant sets the ceiling, because the heat still ends in room air. A site that needs more total cooling than its air plant can reject needs the facility water work described in the retrofit path without rebuilding the building.

ASHRAE also names a half-liquid server as a trap. Cooling the CPU with cold plates while memory stays on air leaves system fan power high, because the plates add airflow impedance, and heat capture to liquid is decreased, eroding the overall TCO benefit. The LBNL result fits from the other side: the reduction in server energy, mostly fan energy, was about 40 percent of the modeled site savings.

The capture data is one 2014 rack of CPU and memory plates, and no source here measures a GPU rack

LBNL’s measurements cover 38 servers in a single rack on a prototype cooling system. The site-level savings came from Romonet models with San Jose, California climate data, so they describe a modeled retrofit, not a metered one. The report lists the heat paths of uncaptured server power and the effect of server air inlet temperature as areas for future study, the two variables a hybrid room depends on.

The OCP figures are guidance: the section read gives 70% to 80% as typical, with no test conditions. Uptime’s survey is self-reported, with 86 to 94 respondents on the DLC-specific questions, and its AALC figures are vendor test results relayed secondhand. The 20% to 30% air share that opens this piece is one Upsite engineer’s statement, as quoted in trade press.

No source above gives measured capture for a GPU server or a rack at today’s densities. No cost data for hybrid builds against single-method builds is in the record either, so nothing here prices the choice. Any design decision on these figures needs qualified engineering review.

The boundary is the air the plates miss, and it keeps the original plant running

A hybrid facility is a liquid loop for what cold plates reach and an air system for everything else in the chassis, the rack and the room. The boundary is the capture fraction, which the sources place between 57% and 80% for cold plates under load and much lower for an idle server on warm water. Each point below 100% is heat the old plant must still reject. The plant is not retired in any phase computed above, so a facility pays for two systems and runs both. LBNL’s models still found savings of 14 to 20 percent in site energy at 50 percent server power, and 16 percent in the case that kept existing chilled water, so in that model carrying two systems was not a net loss.

A floor is how I read the 20% to 30% figure. The evidence supports it at cool supply water, and the same evidence shows the share passing 40% when the water is warm or the servers idle. Operators who size the air plant at 20% to 30% of IT load and then choose warm water for efficiency are sizing for the best capture and operating at the worst.

Hybrid cooling is the answer in practice because partial conversion cannot avoid it and full conversion still leaves a remainder. What stays open is how long the air and liquid sides run together, which depends on supply water temperature, rack density and the air capacity a room has left.

References

• Open Compute Project, “Modular Technology Cooling System for ‘Cloud Scale’: Design & Delivery of Liquid to Rack Distribution Systems,” OCP Data Center Facility group attachment, Rev 0 published 14 October 2024 (workstream input for Rev 1), https://ocp-all.groups.io/g/OCP-DCF/attachment/530/2/CX%20Update%20of%202025%20Modular%20TCS%20%40%20CloudScale%20Design%2C%20Delivery%2C%20Selection%20Guidance%20.pdf
• Coles, H. and Greenberg, S., “Direct Liquid Cooling for Electronic Equipment,” LBNL-6641E, Lawrence Berkeley National Laboratory, March 2014, https://datacenters.lbl.gov/sites/default/files/direct_liquid_cooling.pdf
• Uptime Institute, “2024 Cooling Systems Survey: Direct Liquid Cooling Results,” May 2024, https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2024.Cooling.Survey.Report.pdf
• Uptime Institute, “Self-contained liquid cooling: the low-friction option,” Uptime Institute Blog, 10 September 2025, https://journal.uptimeinstitute.com/self-contained-liquid-cooling-the-low-friction-option/
• ASHRAE Technical Committee 9.9, “Emergence and Expansion of Liquid Cooling in Mainstream Data Centers,” white paper, 2021, https://www.ashrae.org/file%20library/technical%20resources/bookstore/emergence-and-expansion-of-liquid-cooling-in-mainstream-data-centers_wp.pdf
• Data Center Knowledge, “Hybrid Cooling: The Bridge to Full Liquid Cooling in Data Centers,” 3 June 2024, https://www.datacenterknowledge.com/cooling/hybrid-cooling-the-bridge-to-full-liquid-cooling-in-data-centers
• Accessed: 2026-10-09