RTX 5090 Compute Clusters: Air Cooling vs. Liquid Cooling, Compared by PUE

The GeForce RTX 5090 has become the default engine of small AI clusters: 575 W per card, 32 GB of GDDR7, and a price that makes a four-GPU node a realistic purchase for a startup lab, a rendering farm, or an edge inference room^[1]. Stack thirty of those nodes in a room and you are no longer running "workstations" - you are running a small data center, whether you planned for it or not.

At that point the single decision that shapes your electricity bill for the next five years is how you move the heat out of the cards. Air or liquid. The two paths differ in rack density, in how much extra power the facility burns to reject each kilowatt of IT load, and in how much of that penalty shows up inside the racks themselves. This article puts hard numbers on the comparison: what each approach costs you in PUE, what the difference is worth per year on a 1 MW pod, and where each design still makes sense.

Start with what the card actually asks for

NVIDIA rates the RTX 5090 at 575 W total graphics power - the same class as datacenter accelerators of two generations ago - on a dual-slot consumer card with a 32 GB GDDR7 frame buffer running at 1,792 GB/s of memory bandwidth^[1][2]. In practice a card holds its boost clocks by trading fan speed for power: sustained all-core inference or training workloads routinely pull the full 575 W for hours.

Build a realistic node around four of them: 2,300 W of GPU, plus CPU, memory, NICs, storage, and power-conversion losses, and the node lands at roughly 3.0 kW at the wall. A one-megawatt pod of such nodes is 333 nodes and 1,332 GPUs. Those are the reference numbers used for every calculation below.

Know what PUE measures before you compare

Power usage effectiveness is total facility power divided by IT power. A PUE of 1.55 means that for every watt reaching the servers, 0.55 W is spent on cooling, power conversion, and everything else in the building. The industry-wide average has been stuck between 1.54 and 1.58 since 2018 - Uptime Institute's global survey reported 1.56 for 2024^[3] - because the cheap airflow fixes (containment, blanking panels, variable-speed drives) were adopted everywhere years ago. The next step change requires changing the cooling architecture itself.

One subtlety matters for GPU clusters: server fans are inside the IT number. When liquid cooling lets a node drop from high-RPM server fans to a couple of low-speed chassis fans, that saving does not improve PUE at all - it shrinks the IT load. It shows up instead as freed electrical capacity and, downstream, as a smaller facility. PUE measures only the building overhead around the compute. Both effects are real; they just appear on different lines of the bill.

Air cooling runs out of room around 15-20 kW per rack

The average rack in Uptime Institute's survey runs about 8 kW, and the majority of facilities worldwide still deploy no rack above 10 kW^[3]. Against that baseline, a 5090 pod is already dense: at 3.0 kW per node, a rack holding four to six nodes sits at 12-18 kW, which means hot-aisle containment, blanking panels, and carefully managed CRAH airflow just to keep inlet temperatures in spec.

Operators themselves put the crossover point where air stops being workable at 20-50 kW per rack in Uptime's surveying^[4] - a spread, not a magic number, because it depends on climate, ceiling height, and how much airflow the raised floor can actually deliver. The same industry survey shows how uncommon liquid is still: about 75% of operators run perimeter air cooling and only 22% report direct liquid cooling in operation^[5].

The physics behind the ceiling is airflow. Removing 18 kW with air at a realistic 15 K temperature rise requires roughly 1.1 kg/s per rack of continuously driven air, and high-speed server fans to push it through dense GPU heatsinks. Push past 20-25 kW and fan power climbs steeply, hot spots multiply, and every additional rack costs more airflow than the last.

Liquid cooling moves the ceiling to 40 kW and beyond

Direct-to-chip (D2C) liquid cooling mounts cold plates on the GPUs - and usually the CPUs - and routes coolant through rack manifolds to an in-rack CDU that exchanges heat with a facility water loop. In production D2C deployments, 70-85% of the rack's heat is captured to liquid, leaving 15-30% (memory, VRMs, drives, NICs) for a small amount of supplemental air^[6]. Warm-water designs supply the loop at 35-45 °C^[6], which is the detail that unlocks everything else in this article.

With the GPU heat carried by water instead of air, the same 42U cabinet that topped out at 12-18 kW holds 40 kW or more - two to three times the compute per foot of row. At the pod level that difference compounds: fewer racks, fewer breakers, less aisle, less white space.

Where the PUE gap actually comes from

PUE overhead is not one number - it is a stack of subsystems. The table below allocates the overhead of a 1 MW IT pod for each architecture. Treat the split as indicative (real sites vary with climate and plant age), but the totals match the ranges reported for each class: optimized air-cooled facilities land around 1.4-1.8 depending on climate, while warm-water D2C facilities typically measure 1.10-1.25^[6].

The single largest line is heat rejection. A chilled-water plant must produce 7-13 °C water year-round and pays compressor power for every kilowatt. A warm-water D2C loop rejects heat at 35-45 °C, which a dry cooler can reach with outside air alone for most hours of the year in temperate climates^[6] - the compressors simply stop running. That one temperature difference is worth more than every other efficiency measure combined.

Run the numbers on a 1 MW pod

Now put the two architectures side by side on the reference pod: 333 four-GPU nodes, 999 kW of IT load, 1,332 RTX 5090s.

Facility-level overhead comparison for the reference pod. Fan-energy savings inside the IT number are extra - see below.

Three things to notice. First, the saving is continuous: 349 kW of plant that simply never runs, all 8,760 hours. Second, at commercial tariffs from $0.08 to $0.12 per kWh the annual saving spans roughly $245,000-$368,000, so the payback math is dominated by your electricity price. Third, the "every GPU at the meter" row is the number that matters for capacity planning: each 5090 in the air-cooled design costs you 1.16 kW of utility power, and only 0.90 kW in the liquid design - a 22% difference in how much compute a given electrical service can host.

Do not forget the fans inside the IT number

Server fans are the hidden second utility bill. In air-cooled GPU servers, fan power typically runs 10-20% of node power at high densities - designers report up to 20% of total server energy^[8], and production cooling guides put the range at 5-10% for typical loads^[6]. On a 3.0 kW 5090 node, 15% is 450 W of fans; across the pod that is about 150 kW of electrical capacity spent moving air - energy you meter as IT but that produces no computation.

Fan physics also punishes dense racks disproportionately. Fan power follows the cube of speed, so a node whose fans must run 20% faster to hold a hot rack draws about half again the fan power:

Fan power is measured inside the IT load, so these kilowatts come back as sellable compute capacity, not as PUE.

Direct-to-chip does not eliminate fans entirely - the 15-30% of heat that stays on air (memory channels, VRMs, drives, NICs) still needs modest airflow^[6] - but it replaces the high-RPM GPU heatsink fans that dominate the air-cooled figure. Immersion designs remove them altogether; that is the next step beyond this article's scope.

The water-temperature lever decides how low PUE goes

The whole liquid-cooling PUE story hangs on one design choice: the supply temperature of the facility loop. A conventional chilled-water plant delivering 7 °C water still needs compressors most of the year. A D2C loop specified at 35-45 °C^[6] only needs the water to be colder than the chips' allowable inlet - and cold plates are happy with water far warmer than any chilled-water plant. Once the loop runs warm, heat rejection becomes a dry-cooler problem: reject 850 kW (the liquid-side share of the reference pod at 85% capture) at a 15 K loop rise and the CDU circulates about 810 LPM (214 gpm) - a modest flow that modest pumps handle.

How many hours a year your site can reject 35-45 °C water without mechanical chilling depends on local wet-bulb climate - that is a site-selection question, and we treat it in depth in our cooling tower and wet-bulb article. How to share one warm-water loop across both liquid-cooled racks and the residual air handling is an architecture question covered in shared vs. dedicated heat rejection.

What liquid cooling costs you

The comparison is not free. Budget for the other side of the ledger:

• CapEx premium of roughly 25-40% over the air-cooled baseline for cold plates, manifolds, CDUs, and the facility loop^[6] - though liquid-cooled designs buy back part of that in smaller CRAC plants, fewer racks, and less white space.
• Leak risk management. Water inside racks demands dripless quick disconnects, rack and row leak sensing tied to shutdown logic, and pressure-tested commissioning. Our quick disconnect buying guide covers the hardware side.
• A coolant program. Chemistry, filtration, and periodic servicing of CDUs and loops - see the CDU maintenance guide for the actual checklist.
• Operational maturity. Facilities teams and IT teams now share one thermal chain; someone has to own it end to end.

When air cooling still wins

Liquid cooling is not the answer by default. Air remains the right call when:

• Your racks stay under ~15 kW. Four 5090 nodes per rack with containment is firmly air territory; Uptime's fleet data shows most of the industry runs there permanently^[3].
• The pod is small. Below roughly 20 racks, the facility savings may not amortize the liquid plant premium within your holding period.
• The room is a retrofit. Cutting a facility water loop into an operating hall is a construction project; a rear-door heat exchanger or a modest containment refresh may deliver 80% of the benefit for 20% of the disruption.
• Utilization is low or bursty. An inference fleet that averages 25% load will fight its own liquid plant - the part-load penalty erases the design-point PUE advantage.

Decision checklist

• Compute your actual design density: node wall power × nodes per rack. Under 15 kW/rack → air with containment. Over 20 kW/rack → price liquid seriously; operators put the crossover at 20-50 kW^[4].
• Price the loop at 35-45 °C, not chilled water - the free-cooling hours are where the PUE lives^[6].
• Check your local wet-bulb and the hours per year a dry cooler can carry the full load^[6].
• Add the fan dividend: fans inside air-cooled 5090 nodes run 5-20% of node power^[6][8] - that capacity returns to compute with D2C.
• Budget the liquid side honestly: +25-40% CapEx^[6], leak detection, coolant program, and a team that owns the thermal chain.
• Model part-load PUE, not just design-point PUE, if utilization will average below ~40%.

Bottom line

For a 5090 pod at true datacenter density, the arithmetic is blunt: warm-water direct-to-chip liquid cooling turns a ~1.55 PUE air plant into a ~1.15-1.20 one, worth about 3.1 GWh and $306,000 a year on a 1 MW pod at $0.10/kWh - roughly $230 per GPU per year, every year, before counting the ~125-150 kW of fan capacity returned to compute. Below 15 kW per rack, none of that matters more than simplicity: contained air cooling is cheaper to build, easier to run, and entirely adequate. Measure your density first, then choose the plant that matches it.

Sizing a 5090 or datacenter-GPU pod and weighing air against liquid for your own room? Send us your load list - we match cooling architecture, CDU capacity, and heat-rejection plant to the actual density, not to a brochure.

References

1. NVIDIA, "GeForce RTX 5090" - official specifications (Total Graphics Power 575 W; 32 GB GDDR7). nvidia.com
2. NVIDIA, "NVIDIA RTX Blackwell GPU Architecture" whitepaper (GB202, TGP 575 W, 1,792 GB/s memory bandwidth). images.nvidia.com (PDF)
3. Uptime Institute, "Global Data Center Survey 2024" (average rack density ~8 kW; density distribution; industry PUE plateau at 1.56). datacenter.uptimeinstitute.com (PDF)
4. Score Group, "Data Center Cooling Reliability and GPU Density" - citing Uptime Institute surveying on where air cooling becomes too costly or inefficient (crossover concentrated at 20-60 kW/rack). score-grp.com
5. Tele.net.in, "Global Trends: Cooling technology innovations and market dynamics" - Uptime Institute Cooling Systems Survey 2025 (75% perimeter air, 22% direct liquid cooling). tele.net.in
6. RAX Engineering, "Direct Liquid Cooling for GPU Data Centers: Design, Economics, and Deployment Guide" (air PUE 1.4-1.8 vs D2C 1.10-1.25; 70-85% heat to liquid; warm water 35-45 °C; fan power 5-10% of IT; CapEx premium 25-40%). rax.ae
7. NVIDIA, "NVIDIA Driver License Agreement", Section 2.8 - GeForce software "not licensed for datacenter deployment." nvidia.com
8. Winshare Thermal, "How to Choose AI Server Cooling: A Buyer's Framework" (server fans up to 20% of air-cooled server energy; liquid cooling typically necessary at 30-40 kW/rack). winsharethermalloy.com