Air vs Water Heat Transfer: Why the Gap Is 3,400x, Not 4x

Water’s specific heat is about four times air’s, a figure that shows up in nearly every explainer on this topic. Multiply in the density difference and the gap that actually sizes a cooling system, volumetric heat capacity, runs to roughly 3,400 times. A 120 kW GPU rack makes the consequence concrete: removing that heat with air takes close to 14,000 cubic feet per minute, more than seven raised-floor tiles can deliver, while the rack’s own liquid loop does it with on the order of 120 liters per minute. The four-times number is true and nearly useless for sizing anything.

Specific heat differs by four times, and that is the smallest gap in this comparison

The number people reach for first is also the least relevant to equipment sizing. Air’s isobaric specific heat is 1.006 kJ/kg·K at 0°C and standard pressure. Water’s is 4.186 kJ/kg·K at 25°C. That puts water at roughly 4.2 times air on a per-kilogram basis. Source: thermophysical property tables, Engineering ToolBox, “Air Properties” (2025).

A mass-basis comparison describes what happens if a facilities engineer weighs out a kilogram of each fluid and heats it, which is not how either fluid moves through a data center. Air and water are supplied by volume, through ducts and pipes, at a flow rate measured in cubic feet per minute or liters per minute. A property expressed per kilogram cannot size a fan or a pump without first converting through density, and that conversion is where the real gap opens up.

Density turns the four-times gap into roughly 3,400 times

Once density enters the calculation, the comparison that actually governs equipment sizing looks nothing like the mass-basis number. Air at 20°C and one atmosphere has a density of 1.205 kg/m³. Water at 25°C is about 997 kg/m³, roughly 827 times denser. Source: Engineering ToolBox, “Air Density, Specific Weight, and Thermal Expansion Coefficients” (2025).

Volumetric heat capacity is the product of density and specific heat, and it is what determines how much heat a given volume of fluid can absorb for a given temperature rise. Air works out to about 1.21 kJ/(m³·K); water works out to about 4,173 kJ/(m³·K). Divide one by the other and the ratio lands at roughly 3,443, a number close to the 3,200-times figure sometimes quoted for water and air at 77°F. The two figures differ because they are taken at different reference temperatures, not because the underlying physics disagrees. Table 1. Air vs water: three thermophysical propertiesratio = water ÷ air

* Air at 20°C (density) / 0°C (specific heat), 1 atm; water at 25°C, 1 atm. Volumetric heat capacity = density × specific heat; small temperature mismatches between sources do not change the order of magnitude.

Source: Engineering ToolBox, thermophysical property tables (2025)

This is the number that belongs in a sizing calculation, and it is an order of magnitude larger than the number most explanations stop at. A duct sized on the four-times figure would be undersized by two orders of magnitude the moment someone tried to apply it to volumetric flow.

The convective heat transfer coefficient adds a further gap on top of the fluid property difference

Even after matching volumetric flow, water still moves heat off a surface faster than air does, because the convective heat transfer coefficient itself runs higher for liquids. Engineering reference tables put forced-convection air in the range of 10 to 1,000 W/(m²·K) and forced-convection water in the range of 50 to 10,000 W/(m²·K). Source: Engineering ToolBox, “Understanding Convective Heat Transfer” (2025).

The two ranges overlap, so this is not a fixed multiplier the way the volumetric heat capacity ratio is. What the ranges show is that a well-designed water-side heat exchanger can reach coefficients an order of magnitude beyond what any air-side surface reaches, because the diffusivity and viscosity that set the boundary-layer behavior work in water’s favor at ordinary flow velocities. This is the property that makes a cold plate practical at chip-level heat flux and a heat sink impractical at the same flux: a die running at 500 to 600 W/cm² is 40 to 100 times beyond what an air-cooled heatsink’s typical 5 to 15 W/cm² effective flux ceiling can carry, and the convective coefficient gap is most of why. Source: ToneCooling, “NVIDIA GB200 NVL72 Cooling Requirements” (2026).

Running the numbers on a 120 kW rack turns the property gap into a flow-rate problem

The three property ratios above predict a specific outcome at rack scale, and an NVIDIA GB200 NVL72 rack is a clean test case because its power figure is well documented: a nominal 120 to 132 kW, with a peak draw near 192 kW that sets busway sizing. Source: SimScale, “ASHRAE 90.4 Data Center Cooling Design: 2026 Compliance Guide” (2026).

The governing equation for a single-phase cooling flow is Q = ρ · cp · V̇ · ΔT, rearranged to solve for the volumetric flow rate V̇ = Q / (ρ · cp · ΔT). Assume Q = 120,000 W, a supply-to-return air temperature rise of 15°C (within ASHRAE TC 9.9’s recommended rack envelope), air density of 1.205 kg/m³, and air specific heat of 1,006 J/kg·K. That gives V̇ = 120,000 / (1.205 × 1,006 × 15) ≈ 6.6 m³/s, or about 14,000 cfm. This is an upper-bound estimate: it assumes all 120 kW is removed by air at a single average ΔT, which overstates the air-cooled share of a liquid-cooled rack that ships with most of its heat going to liquid.

Run the same equation for water at a temperature rise of about 14°C, using water’s density of 997 kg/m³ and specific heat of 4,186 J/kg·K, and V̇ works out to roughly 0.002 m³/s, or 2 liters per second, about 120 liters per minute. That figure is not a projection; it matches NVIDIA’s own cited coolant flow of around 2 L/s into the cabinet. Source: FiberMall, “NVIDIA GB200 Superchip Guide” (2026).

Divide the two flow rates and the ratio is close to 3,300, within a few percent of the volumetric heat capacity ratio calculated above. That agreement is the check: two independent paths, a governing equation applied to published constants and a vendor’s own cited flow figure, land on the same order-of-magnitude gap. Table 2. Removing 120 kW: air path vs liquid pathbasis: Q = ρ·cp·V̇·ΔT

* Air figure is calculated, treated as an upper bound; liquid figure is NVIDIA’s own cited rack-level value, not independently measured. Ratio of the two flows (~3,300×) tracks the volumetric heat capacity ratio in Table 1.

Source: calculated from thermophysical property tables; FiberMall, “NVIDIA GB200 Superchip Guide” (2026) I read this agreement as confirmation that the volumetric heat capacity ratio, not the specific heat ratio, is the number that predicts real equipment sizing.

Vendor-cited coolant flow rates for the same rack disagree by nearly tenfold, and the equation shows which ones hold up

The 120 L/min figure above is not the only number in circulation for this rack class. A cooling-hardware integrator’s technical brief lists 130 L/min of coolant flow for a GB200 NVL72 loop. A coolant-chemistry supplier’s technical article cites 170 to 195 L/min at a 10°C rise. A buyer’s guide states flatly that the rack runs “20 L/min coolant flow at sub-30°C inlet.” Sources: QCT, “QOOLRACK STAND-ALONE Advanced Liquid Cooling for NVIDIA GB200 NVL72 Systems” (2025); Alliance Chemical, “GPU Thermal Density & Coolant Flow Specs” (2026); GPUaaS.com, “GB200 NVL72: What Enterprise Buyers Need to Know” (2026).

Substituting each figure into Q = ρ · cp · V̇ · ΔT and solving for ΔT at a fixed 125 kW load shows which numbers are physically consistent with a full-rack loop. The 120 to 195 L/min figures imply a ΔT between roughly 10°C and 14.4°C, a tight and plausible band for a technology cooling loop. The 20 L/min figure implies a ΔT above 89°C for the same 125 kW, which is not a plausible cold-plate loop temperature rise. The 20 L/min number almost certainly describes a different measurement point, a single compute tray rather than the 72-GPU rack, or the facility-side makeup flow rather than the technology loop, and the source does not specify which.

This is not a case of one publisher being wrong and the rest right. It illustrates that “coolant flow rate” is not a single defined quantity in vendor literature; it depends on which loop, which ΔT, and which unit of hardware is being measured, and a headline figure without those three details cannot be checked or compared.

The industry’s own “air cooling limit” figures disagree by more than 2x, and the disagreement is not really about physics

The rack power at which air cooling stops working is quoted differently across sources that all claim to be describing the same physical constraint. A power-and-cooling guide puts the ceiling at roughly 25 to 30 kW per rack. A chiller manufacturer’s guide describes 30 to 80 kW as the range where temperature control becomes tight enough to matter. ASHRAE TC 9.9’s own cfm-based argument, as summarized in a 2026 compliance guide, puts the practical ceiling at 40 to 50 kW, where the rack needs up to 5,000 cfm against a raised floor’s roughly 1,900 cfm per tile. And a cooling-chemistry executive quoted in a 2026 trade article describes the thermal density assumptions baked into ASHRAE TC 9.9’s 2021 H1 class as “roughly 40 to 60 kW per rack.” Sources: SimScale (2026); ASHRAE TC 9.9, Thermal Guidelines for Data Processing Environments, 5th edition (2021), as cited via SimScale.

None of these figures is a thermodynamic limit in the sense that the volumetric heat capacity ratio is. Each describes a different combination of containment quality, floor-tile placement, ceiling height and return-air path, and the spread between 25 kW and 80 kW is the visible signature of that variation rather than four sources disagreeing about physics. The 20 to 25 kW to 40 to 50 kW range is a supply-path limit set by how much air a given floor and containment design can deliver to one rack footprint, not a property of air itself. Most halls never approach it: Uptime Institute’s 2025 survey puts the industry’s most common rack density at 4 to 5 kW and finds 82% of facilities peaking below 30 kW even in their densest racks. Source: Uptime Institute, Global Data Center Survey 2025, as cited via SimScale (2026).

Where the textbook constants fall short

The constants used here are standard reference values at specific temperatures and one atmosphere; real air and water properties shift with temperature, humidity, altitude, and, in GB200-class loops, the presence of a glycol-water mixture rather than pure water, which changes both density and specific heat by a few percent. The vendor-cited coolant flow figures compared above were not independently measured; each is a published specification, and none of the four sources states explicitly whether it describes the facility water loop, the technology cooling loop, a single compute tray, or the full 72-GPU rack, which is the gap this piece flags rather than resolves. The “air cooling limit” figures are drawn from secondary compilations of ASHRAE TC 9.9’s guidance and from vendor and trade-press estimates, not from a single controlled test, and they depend on containment design choices this piece does not model. This analysis does not evaluate immersion cooling, rear-door heat exchangers, two-phase cooling, or the economic trade-offs between air and liquid infrastructure; engineering review is required before any of these numbers are applied to an actual facility design.

Where air still works and where the cfm runs out

• If your densest rack stays below roughly 25 to 30 kW, air cooling with real hot-aisle/cold-aisle containment, not just an aisle arrangement, covers the load with margin under ASHRAE TC 9.9’s own airflow argument.
• Above roughly 40 to 50 kW per rack, check the required cfm against your actual floor-tile or containment air path capacity, not against a nameplate power figure; a 1,900 cfm tile cannot supply a 5,000 cfm rack no matter how good the containment is.
• Before citing a vendor’s coolant flow figure, confirm which loop it describes, facility water or technology loop, and at what temperature rise; a flow rate without a stated ΔT and loop location cannot be checked against the rack’s power figure.
• Use the volumetric heat capacity ratio, not the specific heat ratio, when comparing air and water for sizing purposes; the mass-basis 4x figure understates the real gap by nearly three orders of magnitude.
• Treat published “air cooling limit” numbers as descriptions of a specific containment and floor design, not as a property of air, and verify against your own room’s tile capacity and return-air path before assuming a figure from one source applies to your facility.
• Route any final sizing decision through mechanical and thermal engineering review; this piece establishes scale and checks vendor claims for physical consistency, not a design.

References

• ASHRAE TC 9.9, Thermal Guidelines for Data Processing Environments, 5th edition, 2021 (as cited via SimScale, 2026)
• ASHRAE TC 9.9, Emergence and Expansion of Liquid Cooling in Mainstream Data Centers, 2021 (as cited via SimScale, 2026)
• Uptime Institute, Global Data Center Survey 2025 (as cited via SimScale, 2026)
• Engineering ToolBox, “Understanding Convective Heat Transfer: Coefficients, Formulas & Examples,” https://www.engineeringtoolbox.com/convective-heat-transfer-d_430.html
• Engineering ToolBox, “Air Properties – Density, Viscosity, Heat Capacity, Thermal Conductivity, and More,” https://www.engineeringtoolbox.com/air-properties-d_156.html
• Engineering ToolBox, “Air Density, Specific Weight, and Thermal Expansion Coefficients at Varying Temperatures and Pressures,” https://www.engineeringtoolbox.com/air-density-specific-weight-d_600.html
• SimScale, “ASHRAE 90.4 Data Center Cooling Design: 2026 Compliance Guide,” last updated September 1, 2026, https://www.simscale.com/blog/ashrae-90-4-data-center-standards/
• FiberMall, “NVIDIA GB200 Superchip Guide: Liquid-Cooled Racks and Servers,” updated September 2026, https://www.fibermall.com/blog/nvidia-gb200-superchip.htm
• QCT, “QOOLRACK STAND-ALONE Advanced Liquid Cooling for NVIDIA GB200 NVL72 Systems,” 2025, https://blog.qct.io/wp-content/uploads/2025/04/QCT-Qoolrack-Stand-Alone_Advanced-Liquid-Cooling-for-NVIDIA-GB200-NVL72-Systems.pdf
• Alliance Chemical, “GPU Thermal Density & Coolant Flow Specs: B200, GB200, MI300 (2026),” https://alliancechemical.com/blogs/articles/gpu-thermal-density-b200-gb200-coolant-flow-specs
• GPUaaS.com, “GB200 NVL72: What Enterprise Buyers Need to Know (2026),” https://gpuaas.com/blog/gb200-nvl72-enterprise-buyers-guide-2026
• ToneCooling, “NVIDIA GB200 NVL72 Cooling Requirements,” 2026, https://tonecooling.com/nvidia-gb200-nvl72-cooling-requirements/