Modest, stable density
General-purpose compute, storage, and network racks at conventional densities are designed for air, and the fan-power penalty stays small there.
CMP-02COMPARE
Air cooling moves server heat with fans and room airflow; liquid cooling moves it through a fluid circulating in a cold plate on the processor itself. This page sets out the criteria, the assumptions behind them, and where each side wins.
PUBLISHED LAST VERIFIED BY JOSEF ELIMELECHREVIEWED PODOS AI ENGINEERING
The short answer
The comparison
Ten criteria, each with the input that decides it. The fourth column is the useful one: most disagreements about cooling are really disagreements about which input is binding.
| # | Criterion | Air cooling | Liquid (direct-to-chip) | What decides it |
|---|---|---|---|---|
| TH-01 | Density ceiling | Falls away as density rises; fan power scales steeply with flow. | Carries far more heat per unit volume, so density stops being binding. | Sustained kW per rack across the refresh horizon, not day one. |
| TH-02 | Thermal path | Die to sink to room air to coil — several steps, each adding temperature rise. | Die to cold plate to coolant — short path, small rise, so the loop runs warm. | Thermal margin the processor needs at rated power. |
| TH-03 | Facility plant | CRAH units, containment, raised floor or ducting, chilled-water or DX plant. | CDUs, manifolds, quick disconnects, a facility water loop, plus a retained air plant. | Whether the plant already exists or is designed from scratch. |
| TH-04 | Parasitic energy | Server and CRAH fans become a rising share of overhead as density climbs. | Pumps replace most fan energy and typically draw less per unit of heat moved. | Density, and how warm the supply temperature may run. |
| TH-05 | Free-cooling hours | Bounded by the inlet temperature the IT accepts and the local wet-bulb. | Warm water classes widen the dry-cooling window, often to most of the year. | Climate, and the warmest water class the hardware accepts. |
| TH-06 | Water use | Zero if rejected dry; substantial with evaporative or adiabatic assist. | Same rule; warmer loops make dry rejection viable in more climates. | The rejection stage and local water rights — not the cooling method. |
| TH-07 | Retrofit path | Incremental: containment, blanking panels, floor pressure, higher setpoints. | Staged: rear-door exchangers, then in-row CDUs, then a facility loop. | Floor loading, pipe routing, spare capacity in the rejection plant. |
| TH-08 | Serviceability | Pull the server; failure modes are fans, filters, and bypass airflow. | Disconnect couplings; adds coolant chemistry, filtration, and pressure testing. | Staffing depth and appetite for a new maintenance discipline. |
| TH-09 | Failure profile | Graceful: a failed fan degrades slowly and hot spots show up early. | Low probability, high consequence: leaks need detection and rehearsed isolation. | Risk appetite and the maturity of operating procedure. |
| TH-10 | Heat reuse | Exhaust heat is low grade, so recovery is rarely economic. | Higher return-water temperature makes recovery practical. | Whether an adjacent heat consumer exists, and at what temperature. |
Side A
Air cooling is not a legacy technology to apologise for. Each case below is one where adding liquid buys risk rather than performance.
General-purpose compute, storage, and network racks at conventional densities are designed for air, and the fan-power penalty stays small there.
An existing hall with adequate containment and cooling capacity, where tuning the room is cheaper than plumbing it.
A lease with a few years left rarely justifies plant that pays back over a decade.
Thin operations staffing, since coolant chemistry, filtration, and leak procedure need people and drills.
Cold plates are package-specific, so mixed fleets keep re-qualifying interfaces.
Cool climates with long economiser seasons, where the energy argument for liquid is weakest.
Side B
Six conditions, matched one for one against the six above. Where none of them holds, the loop is an expense without an argument.
The hardware ships liquid-cooled, so the question is not whether to plumb but where the CDU sits.
Liquid converts an airflow problem into a plumbing problem — usually the cheaper of the two.
Warm climates where free cooling matters, because warm water classes allow dry rejection across more of the year.
Scarce water or contested permits: a closed loop rejected through dry coolers consumes no water in operation.
A nearby heat consumer, since higher return-water temperature is what makes recovery worthwhile.
Expensive floor area, because removing airflow volume lets the same load occupy far less space.
Reading the matrix
The market is mid-transition, not post-transition. Uptime Institute's 2025 survey of more than 800 operators shows rack densities climbing into the 10–30 kW band while industry-average PUE has been flat for roughly six years: air-side tuning has run out of headroom, yet most halls remain air-cooled.[3] A site running both is the normal outcome, not a compromise.
Liquid cooling does not transform facility efficiency by itself. The published ceiling for excellent conventional plants is already high — Google reports a fleet-wide trailing-twelve-month PUE of 1.09, and Microsoft publishes a design PUE of 1.12 with water-use effectiveness of 0.30 L/kWh.[8][9] Against numbers like those, the overhead liquid can remove is real but bounded. What it changes decisively is the density at which those numbers stay achievable at all.
Water is the other misread. A closed loop consumes nothing; consumption belongs to the rejection stage, where evaporative towers trade water for lower temperatures and dry coolers trade temperature for zero water — in either architecture. Liquid loops tolerate warmer supply water, so the dry option stays viable in more climates,[1][7] and that same warm return water is what makes heat recovery economic.[10]
Most operators are not choosing an architecture; they are choosing how far up a ladder to climb inside a building they already own. Rung one is air-side tuning. Rung two is rear-door heat exchangers, capturing rack exhaust into water without touching the servers. Rung three is in-rack or in-row CDUs, which carry direct-to-chip racks with no facility water plant. Rung four is a facility loop with plant-scale CDUs — a construction project. Federal-lab guidance covers piping and integration for the middle rungs,[7] and the Open Compute Project publishes vendor-neutral cold-plate and disconnect requirements that keep a loop open to more than one vendor.[5][6]
A site running both is the normal outcome, not a compromise.
10
Criteria, and the input that decides each
Method
Cooling comparisons go wrong when the assumptions stay hidden. These are ours; change one and rows in the matrix move.
| # | What the comparison assumes |
|---|---|
| 01 | Liquid here means single-phase direct-to-chip cold plates, the mainstream implementation. Immersion and two-phase fluids move the serviceability and cost rows. |
| 02 | Air means a competently run hall: aisle containment, blanking panels, and setpoints raised toward the top of the accepted class. Comparing against a badly tuned room flatters liquid. |
| 03 | No single kW-per-rack crossover is published here. That threshold is a site calculation, not an industry constant. |
| 04 | Capital cost is directional only, because pricing depends on scale, region, and whether the plant already exists. |
| 05 | Both architectures are held to the same availability target, so the comparison isolates the thermal question. |
| 06 | Retrofit and greenfield are treated as different problems; the same criterion often resolves in opposite directions. |
Honest limits
In the product
PODOS resolves the tradeoff at the factory rather than on the floor. Each PODOS Pod is designed as a standardized 1 MW building block and designed for 128 GPUs, with closed-loop direct-to-chip liquid cooling specified as part of the enclosure instead of added to a room. Density, loop, power architecture, and heat-rejection interface are designed together and tested before shipment — one reason PODOS targets a 90-day window from order to commissioning for a standard unit.
That is a constraint as much as an advantage: a factory-integrated thermal design is the wrong answer for an operator who needs more capacity out of an existing air-cooled hall. The wider build-versus-manufacture question is covered in modular vs traditional AI data centers, the delivery model under deployment, and unfamiliar terms in the AI infrastructure glossary.
QUESTIONS
There is no universal threshold. Uptime Institute's Global Data Center Survey 2025 reports fleet densities climbing into the 10-30 kW band, and rack-scale AI systems such as NVIDIA's GB200 NVL72 ship liquid-cooled because no air-cooled version is offered. For a given hall the crossover depends on inlet temperature, containment quality, floor pressure, and how much fan energy the operator will tolerate.
Yes. For general-purpose compute, storage, and networking at modest density, in an existing hall with working containment and no liquid infrastructure, air is simpler, cheaper to maintain, and staffed by people who already know it. Adding a loop there buys risk, not performance.
Not inherently. A closed loop circulates a fixed coolant charge and consumes nothing. Site water use is set by the rejection stage: evaporative towers consume water in either architecture, dry coolers consume none. Because liquid loops accept warmer supply water, they can reject heat dry more often.
Almost every liquid-cooled facility already does. Cold plates capture heat only from the components they touch, so regulators, drives, and power supplies still need an air path. Mixed halls are the normal case, not an edge case.
Rack load, hall condition, climate, and remaining lease. Engineering will say where the crossover sits for your site — including when it does not sit anywhere yet.