How to Estimate Cooling Requirements for an AI Data Center

Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.

Estimate an AI data center’s cooling from the heat released in the conditioned space, not from a generic watts-per-rack rule. Start with measured or nameplate GPU, CPU, memory, storage, network, fan and rack losses; add room-side lighting, people, pumps, UPS and distribution losses; then convert the result to refrigeration tons and loop flow. As a first-order check, the U.S. Department of Energy (DOE) gives 0.28 refrigeration tons per kilowatt of room heat: 1,000 kW of heat is about 280 tons before project-specific margins. Rack density then determines the architecture. General-compute rooms may remain air cooled, while AI racks around 50–100 kW or more commonly need rear-door heat exchangers, direct-to-chip liquid cooling or a dedicated technology-cooling system.

This guide shows how to build the load, size air and liquid systems, select heat rejection, account for redundancy and growth, and verify the design with operating data.

#	Preview	Product	Price
1		Thermal Guidelines for Data Processing Environments by TC9.9 MISSION CRITICAL FACILITIES...	$147.83	Buy on Amazon

As an Amazon Associate I earn from qualifying purchases.

What exactly is the cooling load in an AI data center?

The cooling load is the rate at which heat must be removed to keep IT equipment and people within their allowed temperature and humidity limits. Almost every watt consumed by equipment inside the conditioned white space becomes heat there. That includes the power used by processors as well as memory, storage, network switches, fans and conversion losses inside the rack.

Keep three quantities separate in the calculation:

• IT load: power delivered to servers, accelerators, storage and networking.
• Room load: IT heat plus lighting, occupants, pumps, fan heat and any other loads physically in the conditioned room.
• Facility load: room heat plus the energy used by chillers, dry coolers, cooling towers, computer-room air handlers (CRAHs), pumps and other plant equipment outside the white space.

The cooling equipment is sized from the room load. Facility energy is used later for electrical-capacity, PUE and operating-cost analysis. Mixing these categories can double-count overhead or leave a mechanical system undersized.

#1 Best Overall

Thermal Guidelines for Data Processing Environments by TC9.9 MISSION CRITICAL FACILITIES (2004-06-30)

$147.83

Buy on Amazon

How do I inventory the IT load?

Build the inventory rack by rack and use telemetry whenever it exists. Nameplate ratings are useful for an early envelope, but they can be substantially above normal draw and may not capture synchronized AI workload spikes.

Collect the components that produce heat

• GPU or accelerator boards, including the number of boards and their sustained and peak power modes.
• Host CPUs, memory, local storage and network interface cards.
• Top-of-rack switches and optical equipment.
• Server, storage and switch fans.
• Power-supply and voltage-conversion losses inside the rack.
• Rack-level power-distribution losses and any equipment-specific conversion hardware.

For each rack, record at least a design maximum, a typical operating value, a diversified peak and a future-expansion value. Diversification should reflect the fact that not every rack peaks at exactly the same moment, while AI training or inference orchestration can make many racks change load together. Do not apply a diversity factor without documenting the workload assumption and the time interval over which it was measured.

Use a load schedule rather than one headline number

A useful schedule has one row per rack or pod and columns for present IT power, expected peak, cooling technology, supply and return temperatures, and the planned phase date. Mark racks that will be liquid cooled separately from air-only racks; their residual air load is not zero because memory, drives, power supplies and room equipment still reject heat to air.

When measured data are available, reconcile the sum of rack telemetry with UPS output and branch-circuit meters. Explain any difference before selecting chillers or CDUs. A persistent gap can indicate unmetered network equipment, an incorrect power factor assumption, or heat that is being rejected outside the room by a liquid loop.

What non-IT heat must be added?

Add every source located on the room side of the cooling boundary. Typical additions are lighting, people, UPS and power-distribution equipment installed in the room, CRAH or in-room fan heat, pumps, humidification or dehumidification equipment, and process loads. A pump installed in a mechanical room may belong in the facility load instead; classify it according to where its motor heat is released.

Use the equipment manufacturer’s electrical input for motors and fans. If a motor is outside the conditioned zone but only the fluid heat enters the room, count the heat that actually crosses the room boundary. For a conservative early estimate, list uncertain items explicitly and carry a documented allowance rather than hiding them in an unexplained percentage.

Separate sensible and latent heat

Nearly all IT heat is sensible. People and humidification systems can add latent load, which affects both coil capacity and humidity control. Keep sensible and latent values separate when selecting CRAHs, coils or air-handling units; a room may have enough total tons but still fail its humidity or dew-point requirement.

How do I convert kilowatts of heat to refrigeration tons?

One refrigeration ton equals 12,000 Btu/h. Convert the room heat to Btu/h, then divide by 12,000. For a fast check, DOE’s 2013 data-center CHP guidance gives 0.28 tons of cooling per kilowatt delivered to the space. Thus, 100 kW of room heat is approximately 28 tons, and 1,000 kW is approximately 280 tons.

The 0.28 value is a first-order conversion, not a final equipment rating. Add the project’s design margin only after the underlying load is clear, and state whether the margin covers measurement uncertainty, future racks, fouling, extreme weather or a specified availability requirement.

These figures describe room heat before project-specific redundancy, part-load, distribution and future-growth decisions. A plant that serves 1,000 kW of IT may require more installed chiller capacity than 280 tons if the room contains substantial non-IT heat or if a chiller is unavailable during maintenance.

Worked example: a 100 kW AI rack

Suppose a rack’s measured IT draw is 100 kW, with 4 kW of local fans and power-conversion losses already included in that meter. Add 2 kW of nearby network equipment and 1 kW of room-side pump and lighting heat allocated to the rack’s zone. The room load is 103 kW, which is approximately 103 × 0.28 = 28.8 tons. If direct-to-chip cooling removes 80 kW through liquid, the air system still has to remove roughly 23 kW of residual rack and room heat, plus any neighboring air-only load.

Worked example: a 1 MW AI hall

If rack and room measurements total 1,000 kW at the design case, the first-order requirement is about 280 tons. If the design calls for 10% additional capacity for a defined expansion phase, the target becomes about 308 tons of useful cooling, subject to the selected redundancy arrangement. Do not silently apply that 10% to every subsystem: pumps, heat exchangers, electrical feeders and heat-rejection equipment each need their own capacity and failure analysis.

How do I calculate airflow or liquid flow?

Capacity in tons tells you how much heat must be removed; flow calculations tell you whether the selected pipes, coils, heat exchangers and fans can carry it at the chosen temperature difference.

Airflow calculation

For air, use the sensible-heat relationship Q = V × ρ × cp × ΔT, where Q is heat-transfer rate, V is volumetric airflow, ρ is air density, cp is air specific heat and ΔT is the supply-to-return temperature rise. Use the actual site altitude and design temperatures when precision matters. A larger allowable ΔT reduces required CFM, but only if rack inlet temperatures remain within the IT manufacturer’s envelope.

After calculating total CFM, check distribution, not just the fan nameplate. Containment, perforated-tile placement, bypass airflow, raised-floor leakage and cable openings can make a mathematically adequate fan system deliver hot air to a rack inlet. Validate rack inlet temperatures with computational-fluid-dynamics analysis or an equivalent airflow study, then confirm with sensors during commissioning.

Liquid-loop calculation

For any liquid loop, use Q = ṁ × cp × ΔT. The mass-flow term, specific heat and allowable temperature rise must match the actual fluid. For water without glycol, DOE’s engineering reference gives the field form q = 500 × GPM × ΔT in Btu/h. Glycol mixtures have different density, specific heat and viscosity, so the 500 constant no longer applies directly; use the supplier’s temperature-dependent properties.

Specify flow and temperature at the same boundary. A CDU secondary loop may have a different ΔT and flow from the facility primary loop, and a direct-to-chip cold plate has its own pressure-drop limit. State whether the quoted capacity is at entering-water temperature, leaving-water temperature, or an average condition.

When does an AI rack need liquid cooling instead of air?

Rack density is the practical decision point. ASHRAE’s current AI data-center framework discusses racks around 50–100 kW and higher and recommends liquid cooling, rear-door heat exchangers, direct-to-chip systems or a dedicated technology-cooling system for high-density zones. DOE guidance shows the transition away from all-air designs as processor power rises through roughly 200–600 W and above.

Do not treat the threshold as a universal cutoff. A 40 kW rack with very restrictive inlet temperatures may need liquid assistance, while a 60 kW rack with unusually efficient hardware and a permissive envelope may be served by a hybrid design. The decision should be based on heat flux, allowable inlet conditions, available airflow, service procedures and the next expansion phase.

Why liquid cooling does not eliminate room airflow

Cold plates usually remove heat from processors, not every component. Memory, storage, power supplies, voltage regulators, fans and network devices may remain air cooled. The room also needs airflow for personnel comfort, fire and smoke control, and removal of incidental heat. Keep containment and rack-level inlet sensors even in a liquid-cooled hall.

How do I size a cooling distribution unit (CDU)?

Size the CDU from the heat assigned to its secondary loop at the specified entering and leaving fluid temperatures, then check flow, pressure drop and control range. Start with the connected cold-plate or rear-door load, add the secondary-pump and heat-exchanger losses that enter the loop, and define a future connection allowance separately from today’s operating load.

1. Define the boundary. State whether the CDU serves one rack, a row, a pod or an entire hall, and list every device connected to the secondary loop.
2. Calculate design heat. Apply Q = ṁ × cp × ΔT with the actual water or water-glycol properties at operating temperature.
3. Select flow and pressure. Sum cold-plate, hose, manifold, filter and heat-exchanger pressure drops at design flow. Confirm the pump can meet the worst branch while maintaining minimum flow through every active device.
4. Check control range. The CDU must remain stable at low AI utilization as well as at synchronized training peaks. Verify valve authority, minimum pump speed and bypass behavior.
5. Provide filtration and chemistry control. Specify particle filtration, conductivity or inhibitor limits, materials compatibility, sampling points and a fill-and-drain procedure.
6. Provide detection and isolation. Include leak detection, drip management, automatic or manual isolation valves, and quick-disconnect procedures that limit water release during service.
7. Coordinate redundancy. Decide whether CDUs are N, N+1 or 2N, and confirm that a failed unit can be isolated without exceeding rack inlet or coolant-temperature limits.

A CDU’s nameplate tonnage is not enough. Ask for its rated capacity at the project’s actual primary and secondary temperatures, fluid chemistry and altitude, and check turndown at the minimum expected load.

Which heat-rejection system should I use?

Compare the heat-rejection options against climate, water availability, noise, plume, electrical capacity, maintenance access and expansion. There is no universally best plant.

ASHRAE describes facility-fluid operation around 45 °C as an enabler for dry-cooler heat rejection in some current AI designs. That is an architectural possibility, not a guarantee that every rack, CDU or climate can operate at that temperature. Confirm the allowable facility-fluid and IT-inlet envelopes with the equipment manufacturers and model the hottest design day.

Use economization only when the site supports it

Economizers can cut annual compressor energy, but their benefit depends on outdoor temperature, humidity, particulate levels and water policy. Establish the hours when an airside or waterside economizer can operate without violating IT conditions. Include changeover controls, freeze protection and the energy used by fans and pumps in the annual model.

How should supply temperatures, ΔT and fluid chemistry be chosen?

Set supply and return temperatures from the IT hardware envelope, heat-rejection capability and condensation risk. The design must cover normal operation, a failed component, sensor error and a controlled restart. Higher liquid temperatures can improve dry-cooler hours, but they reduce the available temperature difference to the environment on hot days and may narrow the control margin.

For water-glycol systems, document concentration, freeze point, specific heat, viscosity, corrosion inhibitors and materials compatibility. Glycol raises viscosity and pumping power while changing heat capacity. A calculation that uses the water-only 500 GPM constant for a glycol loop will overstate capacity.

Specify filtration class, side-stream treatment, sampling frequency and a response plan for conductivity, biological growth or corrosion products. Keep the technology loop isolated from untreated make-up water and provide a method to drain or replace fluid without exposing electronics.

How much redundancy and future capacity should I install?

Choose the resilience target before selecting equipment. N provides exactly the design capacity; N+1 adds one spare module; 2N duplicates the full capacity. The label alone is insufficient: state the largest component failure, whether maintenance is concurrent, and which valves, feeders, controls and heat-rejection paths are also redundant.

Model synchronized AI workload swings, not only the average. A plant can meet the daily mean and still trip when a training job starts across a pod. Use staged starts, thermal storage or control limits only when the workload and availability requirements permit them; do not count software throttling as mechanical capacity unless it is an approved operating mode.

Reserve physical space, electrical capacity, pipe headers, CDU connections and structural load for the next rack generation. ASHRAE warns that AI power and thermal loads can change the data-center design paradigm and recommends adaptive planning. A phased build with capped initial capacity is often safer than installing a large plant that cannot be serviced or expanded.

How do I validate the estimate during commissioning?

Commissioning turns a spreadsheet into an operating design. Compare the calculated load with independent measurements at several load levels and document the conditions for each test.

1. Record UPS output, rack branch power and major mechanical loads at steady state.
2. Measure rack inlet and outlet air temperatures, supply and return air temperatures, and airflow in representative hot, cold and mixed zones.
3. For liquid systems, record CDU primary and secondary flow, entering and leaving temperatures, pump speed, pressure differential and fluid quality.
4. Run representative AI utilization patterns, including a synchronized high-load case, rather than relying only on a resistive load bank.
5. Open and close containment panels or service doors as operators would during maintenance, then verify that nearby rack inlets remain within limits.
6. Test failure and transfer sequences: a CRAH or CDU outage, pump failure, control-network loss, sensor fault and loss of a heat-rejection module.
7. Reconcile every material mismatch. Correct the model, instrumentation or operating procedure before accepting the final capacity.

DOE guidance emphasizes measured data, load calculations, monitoring and commissioning. Keep the resulting trend data as the baseline for later expansion and preventive maintenance.

How do PUE and WUE affect the cooling decision?

Power usage effectiveness (PUE) is total facility annual energy use divided by IT-equipment annual energy use, according to DOE FEMP (2019). Water usage effectiveness (WUE) is annual site water usage divided by IT-equipment annual energy use. Use both metrics with the same boundary and reporting period.

Lower PUE does not automatically mean a better project if it requires unacceptable water consumption, weak redundancy or difficult maintenance. Conversely, a dry cooler may improve WUE while increasing fan electricity in a hot climate. Compare annual and peak efficiency, water use, failure containment, service procedures, capital and operating cost, rack compatibility, floor loading, heat reuse and expansion options together.

DOE identifies 0.6 kW per ton as a “better” cooling-system-efficiency benchmark and 0.8 kW per ton as good practice in its 2024 guidance. Treat these as reference points, not guaranteed results: actual efficiency varies with entering temperatures, part load, climate, pumps, fans and the measurement boundary.

What are the most common sizing mistakes?

• Using nameplate IT power as the whole design. Add room-side non-IT heat and distinguish typical, diversified-peak and future cases.
• Applying 0.28 tons/kW as a final plant rating. It is a first-order room-heat check; it does not select redundancy, pumps, coils or heat rejection.
• Assuming liquid-cooled racks have no air load. Residual components and room equipment still require controlled airflow.
• Using water properties for glycol. Recalculate density, specific heat and viscosity at operating temperature.
• Ignoring synchronized workload steps. Test the power and thermal response when many accelerators change state together.
• Counting standby equipment as available capacity without proving controls. A nominal N+1 arrangement fails its purpose if isolation valves, controls or feeders are single points of failure.
• Choosing a dry cooler from average weather. Size for the local design day and verify noise, plume, footprint and fan power.
• Skipping inlet measurements. Total tons can look correct while a few racks overheat because of bypass air or poor distribution.

What should be in the final cooling-design package?

A reviewable package should let another engineer reproduce the result and understand its limits. Include:

• A rack-by-rack IT load schedule with measurement dates, operating modes and growth phases.
• A room-load schedule showing lighting, people, UPS, distribution, pumps, fans and process loads by location.
• Air and liquid heat-balance calculations, including fluid properties and selected ΔT.
• Supply and return temperature setpoints, allowable IT inlet conditions and humidity limits.
• Airflow or liquid-flow diagrams with pressure drops, control ranges, filtration and leak-detection points.
• CDU, CRAH, chiller, DX, dry-cooler or economizer capacities at stated design conditions.
• N, N+1 or 2N failure analysis, concurrent-maintenance assumptions and transfer sequences.
• Electrical, water, noise, plume, floor-loading and structural constraints.
• Commissioning scripts, sensor locations, acceptance limits and a process for reconciling measured data.

With those items documented, the estimate becomes a living design baseline rather than a single tonnage number that quickly becomes obsolete as AI racks and workloads change.

Frequently Asked Questions

How much electrical capacity should I reserve for the cooling plant?

Reserve capacity from the plant’s modeled peak input, including compressors, dry-cooler or tower fans, pumps, controls and heat-rejection auxiliaries at the site design condition. Convert that mechanical demand into the facility electrical model and test the largest-starting-motor or staged-start sequence; the cooling tons alone do not reveal feeder or generator requirements.

Can I retrofit liquid cooling into an existing air-cooled hall?

Sometimes, but the retrofit must confirm floor loading, ceiling or row space for CDUs, pipe routing, water-treatment provisions, leak containment, service clearances and electrical capacity. A mixed hall should segregate liquid and air zones so a leak, maintenance event or airflow imbalance does not compromise neighboring racks.

Which sensors are most useful for operating an AI cooling system?

At minimum, trend rack inlet and outlet temperatures, room supply and return temperatures, rack or branch power, CDU entering and leaving temperatures, flow, differential pressure, pump speed, leak status and fluid conductivity. Place sensors at hot-spot racks and at representative low-load racks; a single room-average sensor can hide localized excursions.

How often should the cooling calculation be revisited?

Recalculate at each rack-generation change, major workload change, expansion phase, control-sequence change or measured deviation from the model. A quarterly trend review is a practical minimum for a rapidly changing AI hall, while acceptance tests should establish the baseline immediately after commissioning.

Does free cooling work in a hot or humid climate?

It may still provide seasonal hours, but the benefit must be calculated from local weather bins, allowable facility-fluid temperatures, humidity limits and fan or pump energy. Keep mechanical cooling available for the design-day condition and verify economizer changeover, freeze protection, filtration and water-treatment requirements.

Quick Recap

Bestseller No. 1

Thermal Guidelines for Data Processing Environments by TC9.9 MISSION CRITICAL FACILITIES (2004-06-30)

$147.83

Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.

AI data centerASHRAEcooling distribution unitdata-center coolingdirect-to-chip coolingenergy efficiencyLiquid Coolingthermal design