How the rack cooling calculator works
Here is the one fact this whole page rests on: essentially every watt your servers draw becomes heat in the room. A computer is a space heater that does math on the way through. There is no used-up fraction, no portion of the electricity that leaves as anything else useful, so a 10 kW rack is a 10 kW heater running continuously, and the cooling has to remove all 10 kW, continuously. (The pedantic exception is the negligible trickle that leaves down the network cables as signal, and if that trickle ever matters to your cooling design, please write a paper about it.) Because of that fact, sizing rack cooling starts from a single number, the power draw, and everything after it is unit conversion: kilowatts to BTU per hour, BTU per hour to tons of cooling, and BTU per hour to the cubic feet per minute of air that must physically carry the heat away.
Two sibling pages own the neighboring questions. Our BTU calculator sizes an air conditioner for a living room, where sun, insulation, and people set the load and the answer is a heuristic; this page sizes a server closet, where the wattage on the meter IS the load and the answer is arithmetic. Same physics, very different confidence. And once you are done with the rack and want the whole building (meters, dollars, PUE, and CO2), that is the data center power calculator. This page owns the rack and the airflow; that one owns the facility.
The formula
tons of cooling = BTU/hr ÷ 12,000
CFM = BTU/hr ÷ (1.08 × delta T in F)
cooling kW = heat kW ÷ COP
kW is the rack's power draw, which is also its heat output. The 3,412.14 converts kilowatts to BTU per hour and is exact, derived from the definitions of the joule and the BTU. The 12,000 is exact by definition (more on the ice below). Delta T is how much warmer the air leaves the gear than it entered, and the 1.08 is standard air doing bookkeeping: density 0.075 lb per cubic foot, times specific heat 0.24 BTU per pound per degree F, times 60 minutes per hour. That factor assumes sea-level standard air; at altitude the air is thinner, the factor shrinks, and you need more CFM for the same heat (in Denver, roughly 15 to 20 percent more). COP is the cooling system's coefficient of performance: how many watts of heat it removes per watt of electricity it consumes.
Worked example
A 10 kW rack at a 20 F delta T, COP 3: the heat load is 10 kW, full stop, because every watt in is a watt of heat out. In BTU: 10 × 3,412.14 = 34,121 BTU/hr. In tons: 34,121 ÷ 12,000 = 2.84 tons of cooling. In airflow: 34,121 ÷ (1.08 × 20) = about 1,580 CFM. Running the cooling at COP 3 takes 10 ÷ 3 = 3.33 kW of electricity on top of the 10 kW of IT, a mini-PUE of 1.33.
Add-up-the-gear mode, a small homelab: one 300 W server, one 150 W NAS, one 50 W switch. Total 500 W, which is 1,706 BTU/hr, 0.14 tons, and about 79 CFM at a 20 F rise. Cooling it at COP 3 takes about 167 W. Small numbers, but they never stop; see the closet section below.
A BTU, a ton, and a block of ice
A BTU (British thermal unit) is the energy needed to raise one pound of water by one degree F, and BTU per hour is the rate at which heat flows, which is why air conditioners are rated in it. A kilowatt is also a rate of energy flow, so the two convert directly: 1 kW is 3,412.14 BTU/hr, always. The "ton" is better trivia: before mechanical refrigeration, cooling was literally sold as ice, and melting one ton (2,000 pounds) of it absorbs 288,000 BTU. Spread over 24 hours, that is 12,000 BTU/hr, and the industry kept the unit when the compressors arrived. So a 3-ton air conditioner moves heat as fast as three tons of ice melting per day, which is a genuinely useful mental image for what your rack is asking of the room.
Delta T is the design lever
Delta T sits in the denominator of the airflow formula, and that placement is the whole game: for the same heat, doubling the temperature rise exactly halves the air you need to move. Air is a courier, and each cubic foot can carry more heat if you let it warm up more before it leaves. The common design band for air crossing a server is 20 to 25 F, though modern gear tolerates more. This is also what hot-aisle containment actually does, and it is less magical than it sounds: containment does not add any cold, it just puts a physical wall between exhaust air and intake air so the hot output cannot leak back around the rack and pre-heat the input. Clean separation lets the room run a bigger, honest delta T, which means less air to move, smaller fans, and lower cooling bills. The trade is printed in the sensitivity table this calculator shows: a wider delta T means less air but a hotter exhaust aisle, and people occasionally have to work in that aisle.
The airflow reality check
Run the numbers on a single 10 kW rack at a 20 F rise and you get about 1,580 CFM, a serious volume of moving air aimed at one cabinet, continuously, forever. A traditional enterprise rack at 5 to 10 kW is comfortable air-cooling territory. Past about 30 to 40 kW per rack, air physically struggles: the CFM required will not fit through the front of the cabinet at sane velocities, which is exactly why the densest AI racks (the GB200 NVL72 class, at 120 kW and up) ship liquid-cooled, with rear-door heat exchangers or coolant piped directly to the chips. If your racks are the 120 kW kind, our GPU cluster power calculator speaks that dialect, chip by chip. The air-versus-liquid verdict this page prints uses those same thresholds, so this calculator and the facility pages can never disagree about where air gives up.
From COP to PUE
Removing heat is not free: a cooling system consumes electricity to pump heat out of the room, and its COP (coefficient of performance) says how efficiently. At a typical COP of 3, it takes about 1 watt of cooling electricity to remove 3 watts of heat, so a 10 kW rack quietly bills you an extra 3.33 kW at the meter. Add that to the IT load and you get 1 + 1/3, a mini-PUE of 1.33, and this is exactly where the PUE overhead on our data center power calculator physically comes from: cooling is the dominant slice of every facility's overhead. The best hyperscale buildings beat 1.33 by not running compressors at all when they can avoid it (outside air, evaporative cooling, warmer allowed temperatures), which is how fleet PUEs near 1.1 happen. Your server closet probably will not, and that is fine; now you know what the overhead costs and why.
The homelab in the closet
If you got here with one rack of homelab gear in a closet, the math is small but it does not stop mattering. A 500 W homelab is a 1,706 BTU/hr heater, roughly a hair dryer on low, running forever, in a space with no airflow and a door that seals. Closets heat up fast: the gear recirculates its own exhaust, intake temperatures climb, fans spin harder, and drives and power supplies age faster. In summer, a louvered door or a portable AC is not optional, it is the load rating of your hardware asking politely. The cheap fixes, in order: leave the door open, swap it for a louvered one, add a quiet exhaust fan high on the door (heat rises, pull from the top), and only then think about dedicated cooling. Your house's air conditioning is already removing that 1,706 BTU/hr today; the only question is whether the closet lets the house's air reach the gear.