How to Size a Water Storage Tank for a Data Center

A water tank is the easiest piece of cooling infrastructure to get wrong, because it looks like a commodity. It has no compressor curve and no control sequence to argue about — just a volume. So it gets sized by rule of thumb, priced by the gallon, and installed. Then the first real chiller outage comes, and the tank delivers 8 minutes of the 15 the design brief promised.

The reason is almost always the same: the tank was sized as a volume, not as a cooling capacity. Water stores almost nothing per gallon. A single gallon of water cooled 16°F below its return temperature holds about 133 BTU — roughly the energy in a burning match, times a few hundred. Capacity comes from how much water you have, how wide a temperature difference you can hold across it, and how much of the vessel is actually usable. Get any of those three wrong and the tank is decoration.

This article walks through the capacity math for the three water vessels a data center actually needs — the chilled-water thermal storage tank, the make-up water reserve, and the fire water tank — plus the expansion tank that keeps the liquid cooling loop stable. It ends with a complete worked example for a 5 MW AI hall, and the mistakes that cost the most.

Name the tank before you size it

“Water storage tank” describes four different vessels with four different formulas attached. Nothing wastes more engineering time than applying a TES formula to a fire tank, or sizing a make-up reserve as if it had to ride through a chiller outage.

Table 1 — Four vessels, four formulas. The time scale in the fourth column is what separates them; it drives the required volume far more than the flow rate does.

A TES tank is a battery. A make-up reserve is a fuel tank. A fire water tank is a fire extinguisher with a code-mandated discharge duration. Each one is sized against a different clock, and the clock is the whole argument.

Reduce all of it to one equation

Every sensible-storage calculation is the same heat equation, rearranged. Heat moved equals mass flow times specific heat times temperature difference:

That 24 is the only number worth memorizing. It is 1,441 divided by 60, and it exists because a ton-hour is a fixed amount of heat while a gallon-degree is a very small one. Once you have it, the physics stops being the hard part.

Size the TES tank from stored cooling

Start with the load the tank has to carry. Two conventions exist, and mixing them up is the first place a calculation goes wrong.

Use the chilled-water load the tank is actually connected to, not the IT load. For an air-cooled hall, a 5 MW IT load lands near 1,422 tons at conventional efficiency ratios, because the plant has to move the IT heat plus the cooling overhead. For direct-to-chip liquid cooling, the ratio between IT power and cooling load closes toward one-to-one, because the loop captures 75–85% of the heat at the chip and the remainder goes to room air [7]. Never use nameplate chiller capacity — chillers are bought with redundancy baked in, so nameplate overstates the load by the size of the N+1 plant [2].

Then fix the ride-through time. A data center buys a TES tank for resilience, not for load shifting, and resilience is measured in minutes: long enough for chillers to restart after a utility event or for generators to pick up the load. Most US designs land at 10 to 15 minutes to match the UPS battery autonomy they are paired with; mission-critical sites sometimes specify 30 or more. Chinese national practice for a Tier-A facility sets the same floor from the other direction — chilled-water storage must cover no less than the UPS battery's 15-minute rated autonomy [8].

Then read off the volume. The relationship between ΔT and tank size is brutal and linear:

Table 2 — Calculated from V = (Q × t × 24) / (ΔT × η) at η = 0.90. Going from a 16°F to a 10°F design ΔT increases the required vessel by 60% for identical stored cooling.

That last row is the entire argument for chasing ΔT in the plant. The tank does not care how much water flows through the loop; it cares how much temperature difference the loop can hold between supply and return. A plant designed for 16°F that runs at 10°F has given away more than a third of its storage capacity before anyone touches the tank.

Two published worked examples are worth carrying around as sanity checks. A 1,400-ton facility with 10 minutes of ride-through, a 14°F ΔT and 0.90 stratification efficiency works out to 26,667 gallons [2]. A 1,000-ton load for 15 minutes at a 16°F ΔT and 0.9 efficiency lands near 25,000 gallons [1]. Both round up to a 28,000–30,000 gallon shell once freeboard and the diffuser zones are added.

Or size it from flow rate

Some specification packages express the requirement as a flow rate and a thermal maintenance period instead of tons and ΔT — usually because the pump schedule is already fixed. The arithmetic is even simpler:

This method answers a slightly different question: it sizes the vessel to deliver the design flow at design temperature for the duration, and it assumes the tank starts fully charged and fully stratified [9]. It is a screen, not a design. Run it, then check the result against the ton-based formula, and take the larger number.

For the 5 MW hall in this article, the two methods agree closely at a 16°F ΔT: 17,064,000 BTU/hr of heat rejection corresponds to 2,133 gpm of chilled water, and 10 minutes of that flow is 21,330 gallons — which becomes 23,700 gallons once the 0.90 stratification efficiency is applied. When the methods disagree by more than about 15%, the ΔT in the flow number and the ΔT in the ton number are not the same ΔT. Go find out which one is real.

Subtract the volume the tank cannot deliver

A tank's nameplate volume is not its usable capacity, and the difference is not a rounding error. In a stratified tank, cold water sits below and warm water above, separated by a thermocline — a thin, stable layer where the temperature gradient does the work. Storage capacity is defined by how sharp that layer stays, and everything that blurs it costs volume [10].

Four deductions come off the top:

1. Thermocline and diffuser zones. The inlet diffusers at the top and bottom of the vessel occupy real height, and the transition layer itself is not at usable temperature. This is why tank designers talk in terms of equivalent lost height and lost capacity rather than gross volume [10][11].
2. Stratification efficiency. Well-designed tanks with properly sized radial or octagonal diffusers deliver 85–95% of nominal volume. Poorly designed tanks with oversized inlets, short diffusers or squat geometry drop to 60–70% [1][2]. The inlet Richardson number and the ratio of diffuser diameter to tank diameter are the parameters that decide which you get; inlet Reynolds number matters far less than intuition suggests [11].
3. Dead water below the outlet. Whatever sits under the discharge connection, or below the anti-vortex plate, never leaves the tank. It counts toward the invoice, not toward the ride-through [5].
4. Freeboard and standby capacity. A tank with no margin between design charge and overflow is a tank that cannot be charged on a warm day. Add 10–15% of shell volume on top of usable volume — and note that this margin is above the η deduction, not the same thing as it.

Shape the tank after the volume

Only after the volume is fixed does geometry become a decision. Aspect ratio is set by three competing constraints: stratification wants enough height for a stable thermocline, structure wants a sane shell and foundation, and the site wants a footprint that fits between the switchgear and the property line.

Table 3 — Volumes include a 12% shell allowance above usable capacity. Geometry calculated as a vertical cylinder; all figures rounded.

A squat tank is easy to build and hard to stratify, because a short water column gives the inlet jet less distance to dissipate in before it reaches the outlet, and the thermocline grows as a fraction of total height. A very tall tank stratifies beautifully and then creates a structural problem: the shell carries hydrostatic load proportional to height, the foundation sees it, and seismic restraint gets more expensive because the vessel's center of mass rises. Between H/D 0.7 and 1.1 is where most successful chilled-water TES vessels land.

Two more geometry decisions usually get made at this stage. Vertical versus horizontal orientation is driven by footprint, but horizontals stratify worse. And any outdoor tank in a climate that reaches 40°F needs freeze protection designed in from the start — circulation, immersion heating or an insulated enclosure — because a fire water tank that freezes is a fire water tank that does not exist [4].

Size the make-up water reserve

The second vessel answers a different question: not “how long can we cool without a chiller,” but “how long can we run without a city water connection.” That matters because evaporative heat rejection consumes water continuously, and because the failure that empties a water system is usually a municipal supply interruption, not a cooling failure.

Start with the plant water balance. Four streams matter: evaporation, drift, blowdown, and the make-up that replaces all three [12][13].

Run the numbers for the 1,422-ton plant, with a 10°F cooling range and 5 cycles of concentration:

Table 4 — Cooling tower water balance for 1,422 tons of heat rejection, 10°F range, COC 5.

Three useful things fall out of that table. Daily consumption is 239,700 liters, or about 240 m³. At full load that is 2.0 liters of make-up per kWh of IT energy — the same ratio the industry reports as water usage effectiveness. For context, the Uptime Institute's global survey puts the industry-average WUE near 1.9 L/kWh, and an evaporatively cooled plant running at four cycles of concentration typically lands in a 1.4–2.2 L/kWh band [14]. Seasonal WUE will read better than this figure, because evaporation falls with wet-bulb temperature and the plant carries less load on mild days; size the reserve on the design condition, not the average. Annual make-up at 8,000 operating hours is 21.1 million gallons. And raising cycles of concentration from 5 to 6 cuts blowdown and total make-up by roughly a sixth, which is the cheapest water saving available — as long as the make-up chemistry and the Langelier index say the tower will not scale at those cycles [13].

Now the reserve volume. The working industry convention is a 12-hour reserve at peak make-up rate, sized from the mechanical engineer's peak rate and then divided across multiple tanks manifolded in parallel so that one can be taken out of service for inspection without losing the reserve [3]:

For the 5 MW hall, 31,663 gallons is a 120 m³ reserve. That is not one tank: it is a parallel array, sized so that pulling one vessel for its five-year interior inspection still leaves the full 12 hours available. Three 12,000-gallon vessels is a common shape for that requirement.

Do not forget the closed loop

Liquid cooling adds a third water volume that is easy to leave out of the schedule: the technology cooling loop's own make-up and expansion capacity. The secondary loop is closed, so almost nothing evaporates, but it still needs somewhere for expanded coolant to go and enough reserve to refill after maintenance.

Size the expansion capacity from loop volume and temperature swing. A direct-to-chip secondary loop serving 4 MW of liquid-cooled load carries roughly 6,000 to 8,000 liters of coolant in cold plates, manifolds, CDUs and headers, assuming 1.5–2.0 L per kW [7]. Water expands about 0.06% per °C, so a 15°C operating swing on a 6,000-liter loop moves only 54 liters — but that is the thermal term, not the tank size. Industry practice sizes closed-loop expansion vessels at 6–8% of system volume, which is 360–480 liters for this loop, because the vessel has to hold the thermal volume plus the gas-side pre-charge margin [6].

The make-up reserve for that loop is a different calculation again, and it is driven by the two events that actually consume water: slow leakage, and maintenance that drains and refills a CDU circuit. Allow 0.5% of loop volume per week for leakage, add the volume of the largest single CDU loop that can be drained at once — about 500 liters for a 100 kW-class CDU — and carry 72 hours of combined demand. For this loop that is a reserve of roughly 800 liters, which is small enough that it is usually a buffer vessel inside the CDU plant rather than a tank outside it.

Size the fire water tank to NFPA 22

The fire water tank is the one vessel in this article where the calculation is trivial and the compliance work is not. Capacity is flow times duration, full stop [4][5]:

Three details change the answer more than the flow rate does:

• Net usable capacity is measured, not assumed. Capacity counts between the overflow inlet and the discharge outlet. Water below the anti-vortex plate is dead water and does not count toward the fire protection volume [5]. On a 40-foot-diameter tank, that deduction is thousands of gallons.
• Insurance can exceed code. FM Global and similar carriers frequently require more storage than the NFPA calculation alone, along with redundant suction connections and seismic criteria beyond the building code minimum [4].
• Freeze protection is mandatory below 40°F ambient, and the tank carries an inspection obligation — annual inspection with a five-year interior cycle under NFPA 25 [4][5].

Construction follows AWWA standards referenced by NFPA 22: AWWA D100 for field-welded steel, D103 for factory-coated bolted steel, plus separate chapters for concrete and fiberglass vessels [5]. Bolted steel in the 15,000–500,000 gallon range assembles in two to four weeks versus six to twelve for field-welded, which is usually the deciding factor on a fast build [4].

Worked example: a 5 MW AI hall end to end

Put the four vessels side by side for a single facility: 5 MW of IT load, direct-to-chip liquid cooling on 80% of it, air cooling on the balance, a 10-minute ride-through requirement matched to the UPS batteries, a chilled-water plant designed at a 16°F ΔT, and evaporative heat rejection at 5 cycles of concentration.

Table 5 — One 5 MW facility, four water vessels, four different time scales and four different formulas.

Two observations that only appear once the numbers sit in the same table. First, the fire water tank is more than six times the volume of the TES tank, and it is the one nobody thinks of as “cooling infrastructure.” Second, the TES tank — the tank with the formula engineers spend the most time on — is the second smallest vessel on the list. The capacity that takes the most math to get right is not the capacity that takes the most steel.

Sensitivity matters more than any single figure here. Hold the load at 1,422 tons and 10 minutes, and vary only the design ΔT: 10°F needs 37,920 gallons, 14°F needs 27,086, 16°F needs 23,700, 18°F needs 21,067. The spread between the best and worst case is 80% of the vessel. A TES tank decision is really a ΔT decision with a tank attached.

Six sizing mistakes that cost the most

1. Sizing from nameplate chiller capacity. Nameplate includes the redundant chiller. The tank ends up sized for a load the plant will never see, or worse, gets derated in the value-engineering pass because the number looked inflated [2].
2. Using gross tank volume as usable capacity. Jumping straight from the formula to a catalog tank skips the 0.85–0.95 stratification efficiency, the diffuser and thermocline deductions, and the freeboard. On a poorly designed vessel the error compounds to 30–40% [1][2].
3. Designing for a ΔT the plant cannot hold. A 16°F tank in a 10°F plant loses about 38% of its ride-through. Low-ΔT syndrome is the single largest cause of undersized TES performance in the field [1].
4. Treating the make-up reserve as a nicety. The reserve covers the interruption, not the cooling. Sizing it at 24 hours of evaporation while ignoring blowdown understates it by roughly 25% at COC 5, and ignoring drift understates it twice [12][13].
5. Combining process and fire water. A combined tank is permitted under NFPA 22, but the fire volume must remain inviolable — and a plant that draws make-up from the same vessel can strand the fire reserve [5].
6. Leaving the expansion vessel to the CDU vendor. Expansion capacity is set by total loop volume, and total loop volume is set by piping design decisions made by someone else. If nobody owns the sum, the vessel is sized on a number that exists nowhere.

What to put in the specification

Every number below should be an output of the calculations above and an input to the tank submittal. If any of them is blank, the tank is not specified yet.

• Ride-through time in minutes, with the reference event named (chiller restart, generator pick-up, utility transfer) and the load case it applies to.
• Design ΔT across the tank, plus the minimum acceptable ΔT in operation and how the control sequence protects it.
• Usable volume, not nominal volume, with the stratification efficiency the vendor is required to demonstrate.
• Diffuser type and dimensions, tied to the Richardson number and diffuser-to-tank diameter ratio that produced the claimed efficiency [11].
• Shell volume with freeboard, shell material and coating, and the structural loads the foundation must accept at full charge.
• Peak make-up rate, reserve hours, and tank count for the water reserve, with the manifold arrangement that preserves capacity during maintenance.
• Fire flow, duration and margin for the fire vessel, with net usable capacity measured between overflow and discharge, and the inspection access that keeps NFPA 25 compliance practical.
• Loop volume, expansion percentage and refill volume for every closed liquid cooling loop, including future racks.

If the tanks are being bought as part of a prefabricated or containerized plant, the sizing questions move earlier in the program rather than going away — the discussion in why modular container data centers are winning AI buildouts covers how factory integration changes the sequence. The loop side of the calculation — flow, diameter and the pressure drop that decides how much of your ΔT actually reaches the tank — is covered in sizing CDU pipework for a coolant loop, and the long-term water chemistry that decides how much blowdown you will really run sits in preventing and removing fouling in liquid cooling loops.

When the volumes are fixed, the tank specification is a procurement document, and it should be quoted as one. Send us your load case, ride-through requirement, design ΔT and site constraints, and we will come back with vessel options, geometry and the control interface between the tank and your chilled-water plant: request a quote.

FAQ

How many gallons do I need per ton of cooling?

None, until you pick a ride-through time. The useful ratio is gallons per ton-hour, which is 1,441 divided by the design ΔT in °F — 144 gal/ton-hr at 10°F, 90 at 16°F, 72 at 20°F. Multiply by the tons and by the ride-through in hours, then divide by stratification efficiency. A 1,000-ton load for 15 minutes at a 16°F ΔT and 0.90 efficiency needs 25,000 gallons of usable volume.

Should the tank be on the chilled-water side or the condenser-water side?

Chilled-water side. Stored cooling is only useful at the temperature the load can accept, and a condenser-water tank stores water that still needs a chiller to become useful cooling. The exception is a tank used purely as a condenser-water buffer to protect the tower and chiller from short cycling, which is a volume and control problem rather than a ride-through calculation.

Does a bigger ΔT really let me shrink the tank?

Yes, linearly. Capacity is volume times temperature difference, so a 16°F design needs 60% less volume than a 10°F design for the same stored cooling — 23,700 gallons versus 37,920 in the worked example above. The catch is that a wide ΔT has to be designed into the coils, the valves and the control sequence, not just written on the tank specification.

Can I use one tank for fire water and cooling make-up?

NFPA 22 permits combined fire protection and potable water service in the same vessel, but the fire volume has to remain fully available at all times, and the make-up draw has to be arranged so it cannot compromise it. In practice, most data centers keep the fire water tank separate and dedicated, because the compliance and inspection story is simpler and the fire reserve does not have to share chemistry with a cooling water treatment program [5].

1. Anvilfield, “Thermal energy storage chilled water tank field guide for data centers” — stored cooling formula, gallons per ton-hour (1,441/ΔT), stratification efficiency 0.85–0.95, ride-through versus load-shifting sizing, low-ΔT syndrome. anvilfield.com
2. Red River, “How Do You Size Thermal Storage for Data Centers?” — the four sizing inputs, the volume formula, the 1,400-ton worked example, and the warning against using nameplate chiller capacity. redriver.team
3. ProTank, “Water Storage Tanks for Data Centers” — the 12-hour reserve convention (peak make-up rate × 720 min), the 100 gpm and 1,000 gpm examples, parallel arrays for concurrent maintainability, and vessel construction types. protank.com
4. MMI Tank, “Fire Protection Water Storage Tanks” — NFPA 22 scope, flow and duration ranges by hazard, freeze protection below 40°F, FM Global requirements above code minimum, and bolted versus welded erection timelines. mmitank.com
5. Center Enamel, “NFPA 22: Standard for Water Tanks for Private Fire Protection — Complete Guide” — capacity calculation, net usable capacity measured between overflow inlet and discharge outlet, dead water below the anti-vortex plate, and the AWWA D100 / D103 / concrete / FRP chapter mapping. cectank.com
6. Smith Industries, “Expansion Tanks for Data Center Cooling Systems” — ASHRAE description of expansion tank function, the sizing inputs for closed loops, and the ASME pressure vessel considerations for mission-critical service. smithindustriestx.com
7. Rax, “Coolant Distribution Units (CDU) for Data Center Liquid Cooling” (2026) — the 70–85% share of rack heat captured by direct-to-chip loops, the flow-rate formula, the 1.44 LPM/kW figure at a 10°C ΔT, and CDU control requirements. rax.ae
8. Yi Zhu Mechanical & Electrical, “Principle and function of data center chilled water storage tanks” (Chinese) — the Chinese national practice that chilled-water storage must cover at least the 15-minute rated autonomy of a Tier-A UPS battery, plus the stratified tank volume expression with fill and volume efficiency terms. xayizhu.com
9. Masterflow, “How to Size a Storage Tank Correctly” — the flow × thermal maintenance period method with three worked examples (60 L/s, 78 L/s, 41 L/s), the 10-minute default when the period is unknown, and stratification requirements. masterflow.net.au
10. Bahnfleth, Song & Cimbala, “Measured and Modeled Charging of a Stratified Chilled Water Thermal Storage Tank with Slotted Pipe Diffusers,” ASHRAE Research Project 1185, HVAC&R Research 9(4), 2003 — stratification as the controlling performance mechanism, density maximum at 39°F, and the thermal performance metrics (thermocline thickness, lost capacity, equivalent lost height). iifiir.org
11. Musser, Bahnfleth, Wildin & Albuquerque, “Parametric Study of Charging Inlet Diffuser Performance in Stratified Chilled Water Storage Tanks with Radial Diffusers,” ASHRAE Transactions 107(2), 2001 (RP-1077) — inlet Richardson number, diffuser-to-tank diameter ratio and diffuser-to-inlet-height ratio as first-order parameters; inlet Reynolds number not significant. pure.psu.edu
12. Industrial Monitor Direct, “How to Calculate Cooling Tower Make-up Water Volume” — the energy-model method (ton-hours → latent heat → gallons), drift at 0.02% typical, blowdown as E/(COC−1), the 0.001 × flow × ΔT screen, and treatment cost per 1,000 gallons. industrialmonitordirect.com
13. Cooltech Engineers, “Evaporation Loss in Cooling Towers: Conservation & Calculation” — evaporation as 0.00085 × circulation × range, drift ranges by tower type, blowdown expression, and the total make-up identity M = E + B + D. cooltechcoolingtowers.com
14. Alliance Chemical, “Glycol Cooling & WUE in Arid-Region Data Centers” (2026) — WUE defined as liters per kWh of IT energy per The Green Grid and ISO/IEC 30134-4:2017; Uptime Institute 2023 global average of about 1.9 L/kWh; 1.4–2.2 L/kWh for a large cooling-tower campus at four cycles of concentration; best-in-class evaporative near 0.8 L/kWh; closed-loop dry-cooler plants below 0.05 L/kWh. alliancechemical.com