Rack Density by Cooling Type and PUE by Method, Compared

Uptime Intelligence’s February 2025 report works one case: an Nvidia GPU rack of 130 kW with 70% of its heat handled by cold plates leaves 130 kW × 30% = 39 kW for air. A rack that a comparison chart files under cold plate cooling therefore also sits in the air cooling column, and the two ceilings have to be read together. This piece gathers the stated loads for nine configurations from four documents, sets them beside the conditions printed with them, and sorts out what each method asks of an existing room, with the PUE labels examined separately. Cost, water use and failure rates are left out because they need like-for-like sources of their own.

Air cooling tops out near 25 kW at the room edge and near 50 kW once the coil sits at the rack

Uptime’s report rates perimeter cooling as generally effective up to 20 kW to 25 kW per rack with optimized airflow, and 10 kW to 15 kW in older systems. Close-coupled cooling, meaning in-row units or rear doors, typically reaches 50 kW.

ASHRAE TC9.9’s 2021 white paper supplies numbers that explain the lower limit. It gives a best-in-class floor tile as 1,900 cfm and a 40 kW to 50 kW rack as needing up to 5,000 cfm. If airflow scales with load at a fixed air temperature rise, one tile feeds 1,900 ÷ 5,000 × (40 to 50 kW) = 15.2 kW to 19 kW. This rough estimate falls between Uptime’s older-system range of 10 kW to 15 kW and its optimized range of 20 kW to 25 kW.

The two 50 kW figures in the literature are different claims. A 2022 review by Haghshenas and colleagues states that air-cooled racks reach at most 50 kW, citing earlier work. Uptime’s 50 kW belongs to close-coupled cooling, and its perimeter figure is half of that. A chart with a single air cooling row hides a factor of two, and which factor applies depends on whether a coil sits at the rack.

Table 1. Stated load by cooling configuration, with the condition each source prints (kW per rack or tank)

* Ordered by the upper end of each range; sources use different rack heights and no common test method.

** pPUE is the sheet’s own label; the sheet does not define the term.

Source: LiquidStack, “DataTank 48U data sheet” (Mar 2021); Uptime Institute Intelligence, “AI and cooling: methods and capacities” (Feb 2025); Vertiv, “Deploying liquid cooling in the data center” (2024); Haghshenas et al., “Enough Hot Air: The Role of Immersion Cooling” (arXiv, May 2022)

Several ratings in Table 1 carry a condition that decides the number. Vertiv’s active rear door range depends on application parameters, and its passive door range assumes server fans push the air through the coil.

Uptime’s perimeter range assumes optimized airflow, and LiquidStack’s 252 kW appears next to no water temperature at all. A kilowatt rating without its conditions is a range, and a chart sorted by it ranks the missing conditions as much as the hardware.

Cold plates leave 25% to 30% of rack heat in the air, which caps a rack at 67 to 100 kW in a perimeter-cooled room

Vertiv’s 2024 white paper says cold plates, single-phase or two-phase, remove about 70% to 75% of a rack’s heat and leave 25% to 30% for air systems. Its 1 MW example makes the residual concrete: 20 racks × 50 kW × 25% = 250 kW to air, drawn at 20 × 1,875 cfm = 37,500 cfm.

Uptime puts the air share of cold plate systems at 5% to 30%, and in some cases 50%. Its two worked cases are 70 kW × 30% = 21 kW for a 70 kW rack and 130 kW × 30% = 39 kW for a 130 kW rack, the second of which Uptime says calls for rear door heat exchangers. Those doors carry ratings of their own, treated in what a rear door heat exchanger can carry.

A June 2025 Uptime analysis of the Nvidia NVL72 rack gives a second split: cold plates handle around 100 kW of its 132 kW, so 100 kW ÷ 132 kW = 76%, and the air exhaust exceeds 25 kW. The February case assumes 70% capture and the June figures imply 76%, six points apart on nearly the same rack class, and both leave more than the top of the perimeter range in the room.

The ceiling is a division: rack ceiling = air limit ÷ air share. Taking Uptime’s 20 kW to 25 kW optimized perimeter range as the air limit for one rack position and Vertiv’s 25% to 30% as the air share, the low case is 20 ÷ 0.30 = 66.7 kW and the high case is 25 ÷ 0.25 = 100 kW. This is an upper bound. It assumes the residual air meets no other load, and neighboring racks drawing on the same coil capacity pull it down.

Uptime’s cold plate sidecars with liquid-to-air coolant distribution units (CDUs) are rated for 40 kW to 70 kW and release their heat as warm air into the hot aisle, so the room’s air handling still receives the whole rack load and the capture share does not lower it.

A cold plate row printed with one rack load therefore needs the capture share beside it. Without that share the figure can mean 67 kW in a perimeter-cooled hall or more than 150 kW with extended plates, a spread of more than two to one.

A tank’s 1.02 to 1.03 and a survey’s 1.58 are not the same ratio

LiquidStack’s 2021 data sheet prints pPUE of 1.02 to 1.03 for its 48U two-phase tank and does not define the term.

ASHRAE’s glossary defines a partial PUE as a breakdown of the overhead into one function such as cooling, and the 2021 white paper reports a cooling partial PUE of less than 0.07 for the SuperMUC-NG system, which uses warm-water cooling and absorption chillers. A value below 0.07 and a value above 1 cannot sit on the same scale. If the sheet means 1 plus cooling energy ÷ IT energy, then 1.02 to 1.03 is a cooling overhead of 0.02 to 0.03, a reading the sheet does not confirm.

Facility PUE counts everything else. Uptime’s survey average was 1.58 per respondent in 2023 and 1.47 weighted by provisioned IT capacity, and ASHRAE’s glossary lists power conversion, lighting and office equipment among the overhead beside cooling.

Uptime adds that PUE was introduced to track one facility over time and not to compare facilities. Its regional averages of about 1.5 in North America and Europe and above 1.7 in the Middle East, Africa and Latin America reflect climate and the share of small sites using direct expansion units.

The Haghshenas review collects air-cooled values from 1.1 to 2.9 against immersion values from 1.02 to 1.17, and notes that the 1.17 came from an experiment aimed at maximum cooling capacity without efficiency in mind.

The same review turns PUE into computing efficiency through 1 ÷ PUE. Against best-practice air at 1.12, a tank at 1.02 gains 1.12 ÷ 1.02 = 1.098, or 9.8%. Against standard air at 2, the gain is 2 ÷ 1.02 = 1.961, or 96.1%. The same tank looks about ten times better depending on the air baseline.

Removing server fans moves the other side of the ratio. ASHRAE reports fan power of 10% to 20% in some denser servers, and at least 5 kW in a 50 kW rack.

Take 1,000 kW of IT load including 100 kW of fans, with 300 kW of overhead held fixed so only the denominator changes. Before: PUE = (1,000 + 300) ÷ 1,000 = 1.300. After the fans are gone for the same compute: PUE = (900 + 300) ÷ 900 = 1.333. Total power falls by 100 kW, 7.7%, while PUE rises by 0.033. The values are assumed for illustration.

PUE by method hides three choices: what the numerator counts, where the IT boundary sits, and which air facility is the baseline. PUE measures the building and misses the server entirely takes up the boundary problem. For comparing methods, the defensible figure is total kW at the meter for identical compute, and none of the documents read here prints it.

Three doubts that survive the table: tank ratings, floor area and the capture share

Is 252 kW the load of one rack or of one tank?

LiquidStack’s sheet rates one 48U tank, of which 42U is accessible, at up to 252 kW of IT cooling capacity, which is 252 ÷ 42 = 6 kW per accessible U. The rating belongs to the enclosure, so it carries over to a rack only if the IT inside is built to draw that much per U.

Does ranking by kilowatts per unit survive a switch to kilowatts per square meter?

Uptime’s immersion guideline assumes vats of about 3 m × 1.5 m × 1.5 m, and its rack figures assume 19-inch 48U racks. Taking the two larger vat dimensions as the footprint, 150 kW over 4.5 m² is 33 kW/m², while LiquidStack’s 252 kW over a 2.494 m × 1.158 m footprint is 87 kW/m². The unit has to be named before a chart sorts on it, since Uptime’s 150 kW is a floor and LiquidStack’s 252 kW a maximum.

At what capture share does the residual air stop limiting the room?

Using the same division, a 150 kW rack against a 20 kW air limit needs 1 − 20 ÷ 150 = 87% of its heat in liquid. Uptime says extended cold plates take more than 90%, which leaves 150 kW × 10% = 15 kW or less for air. That agrees with Uptime’s statement that racks near 150 kW or more need total or close to total liquid cooling.

Retrofit effort climbs in the same order as the ceiling: a coil on the rack, plates and a loop, then a tank

Vertiv’s 2024 white paper notes that many organizations start with rear doors because they work with existing air cooling and need no significant structural change to the white space, though each rack gains up to 330 mm of depth. Uptime adds that liquid-to-air sidecars need no external piping but occupy rack positions. Cold plate systems with liquid-to-liquid CDUs bring piping into the white space, CDUs on UPS power and blurred service boundaries, and each IT model must be confirmed for plates or immersion. The Haghshenas review judges retrofitting any air-cooled data center with liquid cooling inconvenient in most cases, and Uptime reports that two-phase availability fell after 3M decided to stop producing PFAS coolants by 2025.

Weight separates the last step. Vertiv tells teams to plan each rack for 3,000 lb of static weight, which is 3,000 × 0.4536 = 1,361 kg. LiquidStack lists 2,500 kg for a tank with a 2.494 m × 1.158 m footprint, so 2,500 ÷ (2.494 × 1.158) = 866 kg/m² if the load spreads evenly, and the sheet does not say dry or filled. Per kilowatt of rating that is 866 ÷ 252 = 3.4 kg/m² for the 252 kW tank.

The ladder has no cheap high step. Rear doors and sidecars change the room least and top out near 70 kW. Beyond 100 kW the air share has to fall below 20% to 25%, which means extending plates to memory and storage or leaving the air path altogether, and both bring new piping and new floor loads. Qualified engineering review is required before any figure here sizes a real room. A chart that serves a facility engineer carries four numbers per row: the stated load, the stated condition behind it, the air share left in the room, and the weight on the floor.

References

• Emergence and Expansion of Liquid Cooling in Mainstream Data Centers / ASHRAE Technical Committee 9.9 / 2021 / https://www.ashrae.org/file%20library/technical%20resources/bookstore/emergence-and-expansion-of-liquid-cooling-in-mainstream-data-centers_wp.pdf
• AI and cooling: methods and capacities (UII Update 329) / Uptime Institute Intelligence / February 2025 / https://intelligence.uptimeinstitute.com/resource/ai-and-cooling-methods-and-capacities
• Density choices for AI training are increasingly complex / Uptime Institute Journal / 25 June 2025 / https://journal.uptimeinstitute.com/density-choices-for-ai-training-are-increasingly-complex/
• Large data centers are mostly more efficient, analysis confirms / Uptime Institute Journal / 7 February 2024 / https://journal.uptimeinstitute.com/large-data-centers-are-mostly-more-efficient-analysis-confirms/
• Global PUEs: are they going anywhere? / Uptime Institute Journal / 4 December 2023 / https://journal.uptimeinstitute.com/global-pues-are-they-going-anywhere/
• Enough Hot Air: The Role of Immersion Cooling (Haghshenas, Setz, Bloch, Aiello) / arXiv:2205.04257 / 9 May 2022 / https://arxiv.org/pdf/2205.04257
• Deploying liquid cooling in the data center: A guide to high-density cooling / Vertiv white paper / 2024 / https://www.vertiv.com/495b39/globalassets/documents/white-papers/liquid-cooling/deploying-liquid-cooling-in-the-data-center-a-guide-to-high-density-cooling-white-paper.pdf
• DataTank 48U, Up to 252kW, 2-phase Immersion Cooling data sheet / LiquidStack / March 2021 / https://liquidstack.com/content/uploads/2021/02/LQS_Data-Sheet-Template-48U_03102021.pdf