Cold Aisle vs Hot Aisle Containment: Which Is Better for Data Center Cooling?

Containment is one of the cheapest efficiency projects available to an air-cooled data centre: ENERGY STAR reports paybacks measured in weeks to months for cold aisle containment. Cold aisle containment (CAC) and hot aisle containment (HAC) both stop supply air mixing with exhaust air, and both cut cooling energy. They differ in cost, retrofit effort, fire engineering and working conditions. For sites in Germany they now also differ in a new way: how far they move the annual PUE that the Energy Efficiency Act (EnEfG) caps from July 2027.

Key takeaways

• ENERGY STAR puts cooling savings from a hot aisle/cold aisle layout at 10 to 35 percent, and reports an analysis in which hot aisle containment saved 40% more than cold aisle containment.
• In our worked case, a 1 MW site at the 2025 industry-average PUE of 1.54 reaches about 1.46 with containment. That meets the EnEfG 2027 limit of 1.5 but not the 1.3 limit that applies from July 2030.
• Adding direct liquid cooling to about 75% of the load cuts total energy by roughly 10.6% in the same case, yet reported PUE moves only from 1.46 to 1.41, because the server fan power it removes is counted as IT load.
• A June 2026 government bill would relax the existing-site limits to 1.6 and 1.4. It is not law yet, so plans should still be checked against 1.5 and 1.3.

Understanding the Basics

Cold aisle containment encloses the cold aisle with doors and a roof, so the air delivered by the cooling units reaches server intakes instead of spilling over the racks. The rest of the room fills with exhaust air and runs hot. It fits naturally on a raised floor, where perforated tiles already feed the cold aisle.

Hot aisle containment encloses the hot aisle instead and ducts the exhaust back to the cooling units, often through a ceiling plenum. The room stays at supply temperature and the return air reaching the coils is hotter, which is what lets the cooling plant run more efficiently. The price is a very hot enclosed aisle for anyone working behind the racks.

Energy Efficiency and Cost Savings

Separating the air streams lets operators raise supply temperatures and slow fans without creating hot spots. ENERGY STAR’s guidance on containment puts cooling savings from a hot aisle/cold aisle layout at 10 to 35 percent. It adds that cold aisle containment can reduce cooling-related energy costs by as much as 30 percent, because with containment in place CRAC units can slow their fans, CRAH units can use warmer chilled water and economizers can run more often. The same page cites an analysis in which hot aisle containment saved 40% more than cold aisle containment, because hotter return air makes the cooling units more efficient and because an overheated open room under CAC pushes staff to add mechanical cooling.

The one measured retrofit we could verify was more modest than those headline ranges, and much of its payoff came as released capacity rather than kilowatt-hours. When Lawrence Berkeley National Laboratory retrofitted airflow management in its own legacy data centre, the US Department of Energy’s 2010 case study recorded 8% lower fan energy, CRAC setpoints raised by 3°F, and 60 tons of stranded cooling capacity released, which let one 15-ton unit go to standby.

None of these figures is a PUE forecast for your site. PUE depends on the whole overhead, including chillers, heat rejection, UPS losses and lighting, and on how many hours a year the climate allows free cooling. The worked case below shows how a cooling saving translates into PUE, and where it stops.

Worked case: will containment and liquid cooling meet the EnEfG PUE limits?

Germany’s Energy Efficiency Act, section 11, requires data centres that started operating before 1 July 2026 to reach an annual average PUE of 1.5 or lower from 1 July 2027 and 1.3 or lower from 1 July 2030. New data centres must reach 1.2. On 24 June 2026 the federal cabinet approved a bill that would raise the existing-site limits to 1.6 and 1.4 and limit the rules to sites with at least 500 kW of IT load, as cloudmagazin reported. The Bundesrat gave its opinion on 10 July 2026, and the bill is still in parliament. Until it passes, the current limits apply.

The case below is a transparent calculation you can redo with your own meter data, not a site survey. Its inputs:

• IT load 1,000 kW, air cooled, running all year (an assumption chosen for round numbers).
• Starting PUE 1.54, the average reported by operators in the Uptime Institute Global Data Center Survey 2025.
• Cooling is 75% of the 540 kW overhead, or 405 kW. The other 135 kW is power distribution losses and lighting. This split is an assumption; replace it with your own metering.
• Containment cuts cooling energy by 20% in the central case, with 10% and 30% as low and high cases, drawn from the ENERGY STAR figures above (10 to 35 percent for the layout, up to 30 percent for cold aisle containment).
• The liquid cooling step applies the ratios from a Vertiv and NVIDIA analysis (published in full at ASME InterPACK 2022). In that model, moving 74.9% of the load to direct-to-chip cooling cut total power by 10.2% while PUE fell only from 1.38 to 1.34. Vertiv reports an 18.1% cut in facility power; the rounded published PUE and total figures imply about 7.5% less IT power, from slower server fans, and about 17% less overhead, which is what the case applies.

Step	IT power (kW)	Overhead (kW)	Total (kW)	Reported PUE	vs 1.5 (2027)	vs 1.3 (2030)
Baseline, no containment	1,000	540	1,540	1.54	Fails	Fails
Containment, low case (cooling −10%)	1,000	500	1,500	1.50	Meets, no margin	Fails
Containment, central case (cooling −20%)	1,000	459	1,459	1.46	Meets	Fails
Containment, high case (cooling −30%)	1,000	419	1,419	1.42	Meets	Fails
Central containment plus direct liquid cooling on ~75% of load	925	380	1,305	1.41	Meets	Fails

Three things stand out. Containment alone is enough to meet the 2027 limit in the central case, but the low case lands exactly on 1.50, so a site that only achieves the bottom of the range has no margin. Nothing in the table reaches 1.3.

The liquid cooling step is where PUE misleads. Total power falls from 1,459 kW to 1,305 kW, a 10.6% cut worth about 1,350 MWh a year at constant load. Reported PUE falls only 3.3%, from 1.46 to 1.41. The reason is the denominator: server fans are powered by the server, so their energy counts as IT load, and removing it shrinks the denominator along with the numerator. Measured against the 925 kW the servers still draw after the fans slow, the containment-only site would score 1.58, so the like-for-like improvement is from 1.58 to 1.41. The Uptime Institute made the same point in 2021: direct liquid cooling changes where energy is counted, and annual PUE can hide real savings.

To reach 1.3 from the last row, overhead must fall to about 277 kW, roughly 100 kW or 27% below where the case ends. That gap has to come from heat rejection and the power chain, not from airflow. The usual levers are more free-cooling hours, warmer chilled or secondary water (the Vertiv model raised secondary inlet water to 45°C), and lower UPS and distribution losses. Under the proposed 1.4 limit the case would miss by about 10 kW, which is within the uncertainty of the assumptions.

Limits of this case: the cooling share, the containment saving and the liquid cooling ratios are all assumptions or come from other sites. The Vertiv baseline already had hot aisle containment and a chiller with free cooling, in Baltimore, so its ratios may not transfer to your plant or climate. Replace every input with metered annual data before using the result in a compliance plan.

Implementation and Retrofitting Considerations

Cold aisle containment is usually the easier retrofit on a raised floor. The cold aisle is already defined by the perforated tiles, so the work is mostly doors, roof panels, blanking plates and sealing cable cut-outs. Hot aisle containment needs a defined return path, such as a ceiling plenum, chimney cabinets or ducting to the cooling units, which makes it more intrusive in an occupied room.

Costs depend heavily on the room, the rack mix and whether the ceiling can serve as a plenum, so ask for quotes against your own layout rather than relying on per-rack rules of thumb. ENERGY STAR says cold aisle containment typically pays back in a few weeks to a few months, depending on whether utility rebates are available, and that PG&E found typical containment retrofit paybacks of under two years. Budget for rebalancing afterwards: the LBNL project removed 27 perforated tiles and changed the flow rating of 34 others once the aisles were contained.

Working conditions matter for maintenance. With CAC the open room runs hot, which is where technicians spend most of their time. With HAC the room stays cool but the enclosed hot aisle is uncomfortable for work at the back of the racks, so plan short tasks and ventilation for that space.

Fire Safety and Compliance

Any containment roof or panel can block sprinkler discharge and change how smoke reaches detectors, so both designs need a fire engineering review before installation. Depending on where the enclosure sits, detection or suppression may have to be extended into the contained aisle. ENERGY STAR describes one workaround at a certified Kaiser Permanente data centre: above-aisle containment panels that shrink and drop to the floor at high temperature, avoiding the need for a fire suppression system in that design.

Requirements depend on local fire codes, the authority having jurisdiction and your insurer. Involve them early, because a suppression redesign after installation means reworking the enclosure.

Long-Term Cost Analysis

Over a ten-year horizon the comparison comes down to three numbers you can measure: the capital cost of each option in your room, the cooling energy each saves, and the fire and ventilation work each triggers. The evidence above suggests HAC saves more energy when the return path is available, while CAC is cheaper and faster where it is not.

• Cold aisle containment: lower upfront cost and faster installation, with energy savings at the lower end of the published ranges.
• Hot aisle containment: higher cost where a return plenum or ducting must be built, and larger cooling savings from hotter return air.

In Germany, add the regulatory value. If containment is what moves a site under the 2027 limit, its payback includes avoided non-compliance, not just lower bills.

Which Containment Strategy is Best for You?

• An existing raised-floor room that needs a quick, low-disruption improvement: cold aisle containment.
• A new build, a room with a usable ceiling plenum, or rising rack densities: hot aisle containment.
• A German site planning for the 2030 limit: treat containment as the first step, then model heat rejection and power chain losses, since containment and fan removal alone did not reach 1.3 in the worked case.

Best Practices for Deployment

1. Meter IT and facility energy separately for at least a year before and after, so the annual PUE change can be attributed.
2. Model the airflow first with CFD or a wireless temperature survey to find short circuits and hot spots.
3. Seal the basics: blanking plates, cable cut-outs and gaps between racks.
4. Bring fire engineering and the insurer in during design, not after installation.
5. After containment, raise supply temperatures within ASHRAE’s recommended envelope and slow fans. The energy saving comes from these setpoint changes, not from the panels alone.
6. If liquid cooling is planned, report partial PUE or TUE alongside PUE so the fan saving is not hidden.

Containment remains the first efficiency project for most air-cooled rooms, and in the worked case it is enough for the 2027 limit. It is not enough for 2030. For more on what happens when liquid cooling arrives, read how liquid cooling is solving AI’s heat problem, and for the power side, why power has become the bottleneck for AI data centres.

@online{scienceshot-cold-vs-hot-aisle-2025,
  author       = {Piro, Viros},
  title        = {{Cold Aisle vs Hot Aisle Containment: Which Is Better for Data Center Cooling?}},
  organization = {ScienceShot},
  date         = {2025-02-22},
  year         = {2025},
  url          = {https://scienceshot.com/post/cold-vs-hot-aisle-containment-what-is-the-best-cooling-for-data-center}
}