Retrofitting a hall for liquid cooling: what the existing building decides

1. The question has changed

Liquid cooling used to be a decision made at the design stage of a new facility. It is increasingly a decision made about a facility that already exists — a hall built for air-cooled racks at a density that no longer matches what the operator needs to install.

That changes which questions matter. A new build can choose its pipe diameters, its pump duty and its commissioning sequence. A retrofit inherits all three, and the constraints they impose are not performance questions that can be optimised — they are boundaries the design has to fit inside.

As of September 2026 the industry has moved this way quickly enough that the pattern is well described: retrofitting existing capacity is being treated as the faster route to supporting high-density compute than building new, because it reuses the building, the power and the connectivity that are already there. A brownfield data hall fitting warm-water liquid cooling to support new high-density tenants is a documented pattern rather than an experiment.

Three constraints follow from the building, and each limits the design in a different way: one caps it, one quietly erodes it, and one removes options before any performance question is asked.

2. What the existing pipe decides

Pipework is the least negotiable item in a retrofit, because it is the one thing that cannot be replaced inside a maintenance window. Once the diameter is fixed, so is the relationship between flow and velocity through it — and it is velocity, not flow, that the pipe limits, because pressure drop, noise and erosion all rise with it. Working that through, the heat the pipe can carry is proportional to the square of its internal diameter at a given velocity and temperature rise.

Two consequences follow from the squared term, and both matter in a retrofit.

Capacity scales with the square of the diameter. A pipe of twice the bore carries four times the heat at the same velocity and the same temperature rise. That squared term is why the size label is a poor guide: DN50 and DN80 carry bores of about 52.5 mm and 77.9 mm in schedule 40 pipe, a ratio of 1.48; the nominal sizes, 50 and 80, are in a ratio of 1.60. Squared, those become 2.20 and 2.56 — so sizing from the nominal label suggests about 16 % more capacity than the bore actually provides. The difference between a small and a large bore is not marginal either way; it is often the difference between serving the new load and not, and it is worth calculating from the actual bore.

When the pipe is the limit, the levers are elsewhere. If the existing diameter caps the heat, the ways to recover capacity are to widen the permitted temperature rise, to raise the velocity within what the pipe tolerates, or to split the load across a second loop. Each has a cost: a wider rise may not suit the hardware, a higher velocity costs pumping energy, and a second loop costs capital. Which is cheapest is a site question — but none of the three is a fluid choice. A different fluid changes how much heat each litre carries, which is worth a few per cent; it does not change how many litres the bore will pass.

Velocity bands for pumped water-based coolants in steel pipework sit in a familiar range, and the same figures apply here as in a new build — see what flow rate does my loop need for the bands and how a volume flow becomes a diameter. What changes in a retrofit is that the diameter is an input rather than an output.

3. What the existing pump decides

The pump and pipework referred to in this section are the ones the building already has. Where a distribution unit splits the circuit, that is the facility loop; the technology loop's pump and bore are chosen with the loop itself, and section 5 covers the split.

This is the constraint most often missed, because it does not appear as a limit at all — the pump keeps running and nothing fails.

A centrifugal pump does not deliver a fixed flow. It delivers whatever flow corresponds to the intersection of its own head-flow curve with the system's resistance curve. Raise the resistance and the operating point moves to a lower flow, at the same speed, with no alarm.

And raising glycol concentration raises resistance, because it raises viscosity. The fluid that reaches a colder freezing point is a thicker fluid, and the pressure drop through the same pipework rises with it. The honest description of what happens next is not primarily an energy penalty: it is less flow. Less flow means a wider temperature rise across the load, and a process temperature that drifts away from design. The pump power may even fall, because the pump is doing less work.

Three checks follow, and all three belong before the fluid is chosen:

1. Get the pump curve — and correct it for viscosity. Not the duty point: the curve describes what the pump will do at a different resistance, which the duty point does not. Manufacturer curves are drawn for water at about 1 cP, and a glycol charge at working concentration is many times that, so the curve has to be derated for the fluid actually being pumped before it is used. The correction reduces both head and flow, and it is the step most often skipped.
2. Get the system curve at the coldest condition, where viscosity is highest, for the concentration being considered. The coldest condition governs, not the design point.
3. Check the resulting flow against the heat the loop has to move. If the flow falls below what the duty requires, work through the options in order of cost. First, remove resistance that does not need to be there — a fouled strainer, a partly closed valve, an orifice left in from commissioning. That costs nothing and needs no outage. Then consider a variable-speed drive, which can recover flow only where the motor and drive have speed headroom above the duty point; a pump selected for a fixed duty often has none, and running above rated speed is constrained by the motor and the drive rather than by the pump alone. Then reduce the concentration, and only then replace the pump or the pipework.

The viscosity behaviour behind that is set out in over-concentration in glycol loops, including how much viscosity rises across the concentration range at a common temperature.

A variable-speed pump changes the shape of the answer, not the question. With speed control the pump can be driven faster to recover flow, which converts the problem from "lost cooling" into "more energy". That is a better failure mode, and it is worth knowing which one a given hall has before the concentration is raised.

4. What the shutdown window decides

The third constraint is not physical, and it can be the one that removes options.

A full fluid change needs the loop drained, cleaned and refilled, which needs it out of service. On a live hall that window may be a few hours, a weekend, or nonexistent. The window decides which of the following is available:

• A full changeover — the cleanest result, and the one that needs the longest outage.
• A partial drain and refill — possible on a live loop with isolation and temporary storage, but it leaves a mixture unless the two fluids are compatible, and it does not clean the system.
• No change at all — the existing charge is retained and treated, using an additive programme rather than a replacement.

The third option is not a failure. For an existing brine or alcohol-water charge that is performing adequately, an inhibitor or synergist programme can extend its life without a drain — which is what the system maintenance programmes exist for. What matters is establishing early which of the three the window allows, because it determines whether fluid selection is an open question or a foregone one.

A fourth option sits outside the fluid decision entirely: cool the high-density rows and leave the rest as it is. Hybrid halls with liquid-cooled and air-cooled racks side by side are a recognised pattern, and they limit both the new pipework and the shutdown exposure to the part of the hall that actually needs it.

5. Which loop the question lands on

The three constraints above apply to a specific loop, and a retrofit usually has more than one. The standard way to connect new high-density racks to an existing building is to add a coolant distribution unit that separates the two circuits, and the separation is what makes a retrofit tractable.

The facility loop is the existing one. It is the buried or cast-in pipework, the existing pump, and the connection to whatever heat rejection the building has. Its diameter and its pump are the constraints described above, and they cannot be changed cheaply.

The technology loop is new. It runs from the distribution unit to the racks, it is smaller in bore, and its material, diameter, pump and fluid are all chosen with the loop — subject to the cold-plate and distribution-unit manufacturers' approval, which governs the wetted-material list and the fluid quality limits.

That distinction changes where the concentration question belongs. On the facility loop, concentration is usually left as it is — the loop already works, and raising it costs flow through a pipe nobody can enlarge. The fluid specification that matters for the new load sits on the technology loop, which is a clean, closed circuit and can be specified on its own terms.

Two consequences are worth stating because they are easy to get wrong in the other direction.

The technology loop carries the electrical requirement. If the racks run at high voltage, the conductivity requirement applies to the fluid touching the cold plates — the technology loop — and it is a new-build decision rather than an inherited one. Section 3 of 800 V DC and the cooling loop covers how that requirement moves with voltage.

The two loops exchange heat, not fluid. They are hydraulically separate, which means the technology loop's water quality is protected from the facility loop's. Where the building's chilled water is not suitable to put near electronics, the separation is what allows it to stay in the facility loop.

Where the hall is small enough that a distribution unit is not warranted, a single loop is possible — but then every constraint above applies to the fluid that touches the racks as well, and the fluid choice carries the whole burden.

6. When the pump curve cannot be found

The check in section 3 assumes the pump curve is available. On a hall built a decade ago it often is not — the drawings are lost, the original supplier has changed hands, and the only thing on site is the pump itself.

That is a survey problem rather than a reason to proceed on assumption, and it has a practical answer.

Look for the documents before measuring anything. Drawings are often described as lost when they have simply moved: the O&M manual, the original commissioning report, the BMS trend history, or the spares purchase record will usually name the pump and often carry its curve. Failing those, the model number and impeller diameter are cast or stamped on the casing, and a manufacturer or distributor can usually supply the curve even for an obsolete model — a same-series sample curve scaled by the affinity laws gets close enough to size against.

Measure the operating point. A clamp-on ultrasonic flow meter and a pair of pressure gauges give the actual flow and head at the current condition. One point does not describe a curve, but it confirms what the pump is doing now — which is the number the design has to beat.

Take a short curve by throttling — with minimum-flow protection in place. With the flow meter and pressure gauges fitted, close the discharge valve progressively over several positions to trace the pump's curve from near shut-off toward open, without draining anything. It must be the discharge valve: throttling the suction side risks cavitation. Note what the test does and does not show — it traces the high-resistance end of the curve, which is the region a fouled or over-concentrated loop occupies, but it does not by itself demonstrate spare capacity at low resistance. The safety point is that throttling toward shut-off converts shaft power into heat in the liquid, and most pumps have a minimum continuous flow below which recirculation and seal damage begin; a bypass or an agreed minimum-flow limit is needed, along with a check on temperature rise in the casing. Done with those in place it is the standard field method, and it is worth commissioning before the design is fixed.

If none of that is possible, the conservative assumption is that the pump has no spare flow to give, and the design should not depend on it delivering more than the measured operating point.

7. The fluid decision under those constraints

By the time the three constraints are known, the fluid question has usually narrowed on its own. The rows below are about the loop the constraints apply to — the existing one, unless a distribution unit separates the two:

The common thread is that concentration is the variable, and on the loop the constraints apply to it is being pushed down by the retrofit rather than up. That is the opposite of the habit a new build can afford, where concentration is chosen for the coldest condition with margin. In a retrofit, every point of concentration has to earn its place against the flow it costs.

8. What to survey before deciding anything

Five items, and the first four are facts about the building rather than choices about the fluid:

1. Existing pipe route, diameter and material, recorded section by section — a retrofit often has more than one diameter in the run.
2. The pump curve and the current operating point, plus whether the pump is fixed- or variable-speed.
3. The available shutdown window, in hours, and whether a partial drain is acceptable.
4. The wetted materials in the existing circuit, including any part of it that predates the current fluid.
5. The coldest fluid condition the loop will see, including standby and any pipework exposed to ambient.

Those five determine what the loop can physically do, and they are the reason a retrofit specification is written around the building rather than around the load.

9. Limitations

The relationships above are qualitative where the building is unknown. The capacity scaling with diameter squared is arithmetic, and the pump behaviour is standard centrifugal characteristics, but neither predicts a specific hall's numbers. Both need the actual pipe schedule and the actual pump curve.

A partial drain leaves a mixture. The guidance on top-up and changeover — set out in can I mix two coolants — applies whether the change is planned or forced by a short window.

Hybrid halls carry an operating cost that is easy to underestimate. Two cooling regimes mean two monitoring routines, two sets of failure modes and two spare-parts holdings. That cost is real, and it belongs in the comparison against a single liquid-cooled hall.

No independent measurement was made. Product figures are reference values published by the manufacturer, and no independent testing or third-party endorsement is implied.

10. Where a retrofit succeeds or fails

A retrofit is usually decided by what was surveyed before the fluid was discussed. Where the pipe diameter, the pump curve and the shutdown window are established early, the design fits inside them and the outcome is predictable. Where they are discovered late — usually when a pump turns out to be unable to hold flow at the concentration the freezing point required — the correction is capital work rather than a specification change.

The practical sequence is therefore: survey the building, then size the loop, then choose the fluid. That is the reverse of how a new build is often approached, and it is the main thing that makes a retrofit a different problem rather than a smaller one.

Bring the pipe schedule, the pump curve and the shutdown window to your supplier before the fluid is discussed. Those three documents determine which options are open, and they are cheaper to read than to discover.