본문 바로가기
Data Center Technology Lab Data Center Technology Lab

Delta T Data Center Cooling Sets the True Capacity Limit

NVIDIA’s own reference design lists the GB200 NVL72 rack at 120 kW; deployed builds have been reported drawing 130 to 132 kW. Air cooling stops being the economical choice somewhere between 20 and 25 kW per rack, according to Uptime Institute’s operator surveys. The gap between those numbers is not a materials problem or a fan problem on its own — it is a temperature-difference problem. The same governing equation that sizes a server fan also sizes a chiller plant, and delta T is the variable both machines share.

The Capacity Equation Has Two Knobs, Not One

Cooling capacity is not a fixed property of a rack, a CRAH unit, or a coolant loop. It is the product of how much fluid moves and how much its temperature is allowed to rise on the way through. The governing relationship is sensible heat transfer: Q = ṁ × cp × ΔT, where Q is the heat removed, ṁ is mass flow rate, cp is the fluid’s specific heat, and ΔT is the temperature rise across the load. For air-cooled IT equipment at sea level, this reduces to a widely used field shorthand: ΔT(°F) = 3.1 × watts ÷ CFM, or equivalently CFM = 3.1 × watts ÷ ΔT(°F). In practice, the temperature rise through commercial rack-mounted servers typically runs 20°F to 35°F, with blade servers producing a higher rise than typical 2U pizza-box units.

Source: Upsite Technologies, Ian Seaton, “The 4 Delta T’s of Data Center Cooling: What You’re Missing” (2015; updated 2019)

A rack rated for 20 kW at a 20°F rise and a rack rated for 20 kW at a 40°F rise do not need the same airflow — one needs roughly twice as much as the other. The wattage rating on a spec sheet says nothing about which fan, which floor tile, or which coolant distribution unit (CDU, the pump-and-heat-exchanger skid that supplies liquid-cooled racks) is required until a delta T is attached to it.

ASHRAE Sets the Ceiling and Floor on Delta T, Not the Wattage

ASHRAE’s thermal guidelines constrain a temperature range, not a power number, and the achievable delta T is squeezed between a recommended supply floor and an allowable exhaust ceiling. The fifth edition of Thermal Guidelines for Data Processing Environments sets a recommended server-inlet dry-bulb range of 18°C to 27°C, and this recommended range is identical across every air-cooled equipment class. What differs by class is the allowable range: Class A1 extends to 15°C–32°C, A2 to 10°C–35°C, A3 to 5°C–40°C, and A4 to 5°C–45°C.

Source: ASHRAE Technical Committee 9.9, Thermal Guidelines for Data Processing Environments, 5th Edition (2021), summarized in ASHRAE TC 9.9’s 2016 power white paper and multiple TC 9.9 explainer briefs

Table 1. ASHRAE recommended vs. allowable inlet temperature°C, dry bulb

ClassRecommendedAllowable
A118–2715–32
A218–2710–35
A318–275–40
A418–275–45

* Recommended range is identical for every class; only the allowable ceiling and floor change

Source: ASHRAE TC 9.9, Thermal Guidelines for Data Processing Environments, 5th Edition (2021)

The 27°C recommended ceiling is routinely mistaken for a hard limit on the industry. It is a target for long-term operation and energy efficiency; the allowable classes exist specifically to describe how far above the recommended range equipment can run before reliability, not function, is what degrades.

Why 20 to 25 kW Is Where the Air-Side Arithmetic Runs Out

The widely cited 20 to 25 kW air cooling ceiling per rack falls out of the same equation rather than describing a separate physical limit on air. Holding delta T at the 20°F figure that Upsite describes as representative of common rack-mount servers, a 20 kW rack needs about 3,100 CFM. A 120 kW rack at the same 20°F rise needs about 18,600 CFM — six times the air moving through the same floor opening and the same rack footprint. Fan power does not scale linearly with airflow; it scales with the cube of the flow rate, so a sixfold increase in CFM implies roughly two orders of magnitude more fan power to hit that flow, before counting the added noise and the space the ductwork or fan wall consumes.

Source: fan power modeling in U.S. Patent 10034417, “System and methods for simulation-based optimization of data center cooling equipment,” describing fan power as proportional to the cube of airflow rate

Uptime Institute’s operator surveys independently confirm the crossover: most facilities report their highest-density racks in the 10–19 kW range, and Uptime notes that above roughly 20 to 25 kW, direct liquid cooling and precision air cooling become more economical than conventional room-level air cooling.

Source: Uptime Institute Journal, “Rack Density is Rising” (2020)

The 20 to 25 kW figure describes a floor tile and a fan curve, not a property of air. Raising fan speed solves the equation on paper and produces a rack nobody can service, hold a conversation near, or fit inside a normal ceiling plenum.

Liquid Cooling Does Not Change the Equation, It Changes the Coefficient

The same Q = ṁ × cp × ΔT still governs a liquid-cooled rack. What changes is the fluid: water’s heat capacity per unit volume is roughly 3,500 times greater than air’s, so the same heat load at the same temperature rise needs dramatically less flow.

Source: U.S. Patent 7342789, “Method and apparatus for cooling an equipment enclosure through closed-loop, liquid-assisted air cooling in combination with direct liquid cooling,” citing air’s heat capacity per unit volume as roughly 3,500 times smaller than water’s

Running the numbers for a 120 kW load at a 10°C rise: for water (density approximately 1,000 kg/m³, specific heat 4,186 J/(kg·K)), the required mass flow is 120,000 ÷ (4,186 × 10) ≈ 2.87 kg/s, or about 172 liters per minute. For air (density approximately 1.2 kg/m³, specific heat 1,006 J/(kg·K)) at the same 10°C rise, the required volumetric flow is 120,000 ÷ (1.2 × 1,006 × 10) ≈ 9.94 m³/s, or roughly 21,060 CFM. The ratio between those two flow rates is about 3,470 to 1 — matching the volumetric heat capacity ratio cited above, and this is an upper-bound style comparison assuming the two systems remove identical heat at an identical rise rather than a measurement of any single installed product.

Table 2. Flow required to remove 120 kW at a 10°C riseair vs. water

MediumΔTFlow rate
Air10°C~21,060 CFM (~596 m³/min)
Water10°C~172 L/min

* Calculated from Q = ṁ·cp·ΔT; not a specific product spec — air: ρ 1.2 kg/m³, cp 1,006 J/(kg·K); water: ρ 1,000 kg/m³, cp 4,186 J/(kg·K)

Source: Author calculation using heat transfer fundamentals; cross-checked against Alliance Chemical, “GPU Thermal Density & Coolant Flow Specs” (2026)

A coolant vendor’s published guide for GB200-class racks lists a flow requirement of roughly 170 to 195 liters per minute at a 10°C inlet-to-outlet rise, which brackets the 172 liters per minute this calculation produces.

Source: Alliance Chemical, “GPU Thermal Density & Coolant Flow Specs: B200, GB200, MI300 (2026)”

ASHRAE’s own liquid-cooling classes were renamed in the fifth edition from W1 through W5 to W17, W27, W32, W40, W45, and W+, with the number now denoting the maximum facility supply water temperature in Celsius rather than an arbitrary label.

Source: Upsite Technologies, “What You Need to Know About ASHRAE’s Fifth Edition of Thermal Guidelines” (2021); HPAC Engineering, “ASHRAE Introduces Thermal Guidelines for Liquid-Cooled Data-Processing Environments” (2011)

The liquid cooling transition reads less like a materials upgrade than a unit conversion: the industry swapped a fluid with poor volumetric heat capacity for one with abundant capacity and kept the same arithmetic running underneath.

The GB200 Case Shows Nominal and Measured Are Not the Same Number

NVIDIA’s reference architecture lists the GB200 NVL72 rack at approximately 120 kW, while deployed builds have been reported drawing 130 to 132 kW under full load; one OEM’s published specification splits this as 115 kW liquid-cooled plus 17 kW for air-cooled supporting systems.

Source: Pantheon, “GB200 & GB300 NVL72 Power and Cooling Requirements”; Spheron Blog, “GB200 NVL72: Rack Specs and Cloud Pricing”

That 10 to 12 kW gap is not measurement noise. It is the difference between a nominal design figure for the compute tray population and the measured draw of the whole cabinet, support electronics included. Treating a vendor’s headline kilowatt figure as the number to size a cooling plant against repeats the same error as treating the ASHRAE recommended envelope as an allowable one: the nameplate number and the number that shows up on the power meter are answering two different questions, and a facility plan built on the lower one runs short.

Where These Numbers Stop Holding

This piece works from published class boundaries, a field-standard airflow formula, and a from-first-principles heat balance; it does not verify site-specific recirculation, containment effectiveness, or altitude corrections to the 3.1 coefficient, all of which shift the airflow a real rack needs at a given delta T. The 170 to 195 liters-per-minute GB200 coolant figure comes from a coolant vendor’s technical guide rather than a NVIDIA data sheet, and it is not stated there as a measured value versus a design target — treat it as a planning figure to confirm against the CDU or server OEM’s own documentation. The fan-power cube law is an idealized model; real fan curves, variable-frequency-drive staging, and mechanical losses mean actual power draw at high CFM can differ from a clean cubic extrapolation. Finally, ASHRAE’s 5th edition guidelines predate the current generation of GPU-dense hardware, and a future revision may adjust the H1 high-density class or the W-class boundaries referenced here.

Sizing From Measured Load, Not Nameplate Delta T

  • Confirm whether a vendor’s published rack kW figure is a nominal design number or a measured full-load draw, and size electrical and cooling capacity to the higher of the two.
  • Calculate the CFM or liquid flow rate your intended power level actually requires using Q = ṁ × cp × ΔT before assuming an existing CRAH or CDU plant covers a density increase.
  • Treat the ASHRAE recommended envelope (18°C–27°C) as the design target for continuous operation, and reserve the allowable range for short excursions during equipment or utility failures.
  • Weigh fan power at the cube of any planned airflow increase, not linearly, before choosing an air-side retrofit over a liquid conversion.
  • Check which ASHRAE liquid-cooling W-class, by maximum supply water temperature, your CDU or dry cooler plant is designed around before assuming a given number of free-cooling hours.
  • Verify site-specific containment and recirculation losses against the nameplate delta T of the equipment; the number on a spec sheet assumes a clean supply path that a real room may not deliver.

References

Heonsu Nam
Heonsu Nam runs Data Center Technology Lab. His day job is running a small career-education business in South Korea; this site is where he studies data center cooling in public. He is not a mechanical engineer, so every article starts from the primary documents (ASHRAE, Open Compute Project, vendor data sheets) and shows the arithmetic so readers can check it. Spot an error? Please use the Contact page.
함께 보면 좋은 글
Overhead close view of a steel chilled water pipe elbow with condensation droplets beaded on its surface
Dew Point Liquid Cooling Sets a Floor 2°C Above the Room
10월 27일
Copper microchannel cold plate beside a stack of aluminum heat sink fins under raking light
Chip Heat Flux W/cm² Puts the H100 at Air Cooling’s Edge
10월 26일
Overhead macro of a glass sight tube on a steel refrigerant fitting, with a thin band of vapor bubbles rising through the liquid
Latent Heat Cooling Pays Off Until Vapor Quality Nears 0.7
10월 15일
Close view of a passive GPU heat sink fin stack with one warm-tinted fin edge near the base
H100 Thermal Throttling Starts at a Margin, Not at 87 °C
10월 13일

댓글 0

첫 댓글을 남겨보세요.