Where the Coolant Actually Works in a Data Centre: Three Loops, Three Different Jobs

Ask where coolant works in a data centre and most answers stop at "in the liquid cooling". That is one loop out of three.

A modern hall runs heat-transfer fluid in three separate circuits. The technology loop behind the cold plates. The facility water loop that feeds dry coolers and plate exchangers. And the coil loops behind rear-door heat exchangers, where a hall is easing itself from air toward liquid. They share a building, a plant room and a procurement budget. They do not share a job.

At Glacier Coolant we read a hall as three loops. The fluid gets a different job description in each one, and grade, sampling rhythm and alarm thresholds all follow from that job.

Which loops actually carry coolant? The cold plate loop, the facility water loop with its outdoor sections, and any rear-door or in-row coil loop. A fully air-cooled hall still runs fluid in the chilled water circuit. An outdoor dry cooler puts a freeze-protection demand on it, whether or not a single rack is liquid-cooled.

Is cold plate coolant the same as facility water glycol? No. The cold plate loop asks for cleanliness and a closely watched chemistry in sub-millimetre channels. The facility loop asks for freeze margin across an outdoor winter. A fluid optimised for one job is merely tolerable in the other.

Does a rear-door exchanger use cold plate coolant? Usually it sits closer to the facility loop than to the cold plate loop. The coil channels are wider, the volume is larger, and the cleanliness demand is milder. But it still deserves its own answer, not an inherited one.

One hall, three jobs

The three loops connect through heat exchangers, not through the fluid. The fluid in each loop stays in its own circuit, which is why each circuit deserves its own specification line.

Job one: the technology loop behind the cold plates

This is the loop that made liquid cooling famous. Coolant leaves the CDU, enters cold plates bolted to processors, and returns carrying the heat that used to leave as air.

The channels inside those plates are sub-millimetre. That single fact reorganises the fluid specification. What sails through a strainer will sit in a plate. The Open Compute Project's commissioning guidance calls for filtration down to roughly 5 μm before a loop goes into service. It also calls for flush velocities of 3–5 ft/s, and it says plainly that the cold plate maker's numbers take precedence. Those figures belong to OCP, not to us, and they are the right place to start.

Chemistry matters as much as cleanliness. The OCP reference for direct-to-chip service is PG25 — propylene glycol at roughly 24.5–29.5 vol% with pH held between 8.0 and 10.5. It is a reference point, not a decision. Where the wetted list includes aluminium, the inhibitor package must be built for aluminium, and that requirement lives inside a grade number, not inside the words "propylene glycol". The eight questions that stall a coolant decision start with this one, and the second-year behaviour of the loop depends on the answer.

Conductivity gets watched closely in this loop, and for the right reason. The coolant normally does not touch electronics, so a conductivity number is not about electricity. It is a proxy. Rising conductivity means contamination, chloride ingress or inhibitor breakdown has entered the conversation. Treat it as a smoke alarm, set the threshold against your own baseline, and it will earn its place on the dashboard.

When a rack in this loop runs warm, the diagnosis path runs through flow, filter loading, concentration and air — in that order. Reading a warm rack before blaming the CDU covers the sequence.

Job two: the facility water loop and the dry coolers

The second loop is older than liquid cooling and will outlive every rack in the building. Chilled water leaves the plant room, picks up heat from CDUs, plate exchangers or rear doors, and dumps it — to a dry cooler on the roof, a cooling tower, or the outside air in winter.

Free cooling is where this loop earns its keep, and where the fluid earns its specification. In a free-cooling season the loop runs at outdoor temperature. The piping sits outside too, across the roof and up the riser. If that circuit holds plain water, the first hard freeze finds the thinnest pipe section and writes the maintenance ticket.

So the facility loop asks for freeze protection, and freeze protection is a concentration decision. The number comes from the local design temperature, not from the record low and not from a round figure. Glycol concentration sets the freeze point, and it taxes you twice. Heat capacity drops as concentration rises. Viscosity climbs too, exactly when the free-cooling season makes the loop coldest. A loop sized for water that is later filled with glycol moves less volume at the same pump head. That arithmetic belongs in the design stage, not in the first January.

Corrosion control does not take a holiday in this loop either. Glycol formulation includes inhibitors precisely because treated water alone does not protect mixed metals for a decade of service. ASHRAE's TC 9.9 water classes — W17, W27, W32, W40, W45 — describe facility water by maximum supply temperature. They are useful shorthand for which loop you are talking about. The fluid side still needs its own eyes: concentration, pH and inhibitor margin, checked on a schedule.

Water that enters this loop where it should not — through a vent, a make-up line or a weeping seal — behaves the same as it does in any cold circuit. How ice blockage starts and returns applies to a data hall's outdoor sections exactly as it does to a cold storage plant.

Job three: rear-door heat exchanger loops

The third loop is the transition loop. A rear-door heat exchanger replaces a server cabinet's back door with an air-to-liquid coil. The rack still breathes air; the door takes the heat out. For halls that cannot jump straight to cold plates, it is the pragmatic half-step.

The coil circuit sits between the other two jobs. Its channels are wider than a cold plate's, and its volume is larger. Its fluid demand reads closer to the facility loop: freeze margin where the circuit rides outdoor sections, corrosion control for copper or stainless coils, and a filter level a coil can live with. Conductivity still gets logged, but nobody mistakes it for the primary alarm it is in a cold plate loop.

The volume point deserves a note. A row of rear doors holds a lot of fluid compared to a cold plate circuit of the same heat load. Fill decisions, top-up discipline and baseline sampling all scale with that volume, and the cost of getting the grade right scales with it too.

The eight questions that stall a coolant decision cover the shipping, mixing and service-life queries that sit behind every door-coil installation, and they are worth answering before the volume is ordered.

The three jobs, side by side

The sampling rhythm follows published equipment guidance: high-density AI loads around every six months, general service around annually. Every top-up earns a test regardless of the calendar. Those intervals come from equipment makers' public maintenance guidance, not from our own statistics.

Why one campus ends up with two grades

Procurement likes one part number. The building does not care.

The technology loop and the facility loop ask different things of their fluid. A campus that serves both with a single grade either over-specifies one loop or under-protects the other, and the quiet failure mode is the second one. Two grades, matched to two jobs, is not procurement inefficiency. It is what the architecture asks for.

Before merging purchase lines, walk the loops. Which circuits touch outdoor temperatures, which wetted lists include aluminium, where the sub-millimetre channels are. The answers usually draw the grade boundary for you. That boundary usually runs exactly where a single-grade purchase order assumed it would not.

What we deliver

A grade per loop, with the reason. Matched to each circuit's temperature band, wetted materials and outdoor exposure. Written down, so the next engineer inherits the reasoning and not just the part number. The product range by temperature band shows the candidate families before the duty check narrows them.

A sampling plan per loop. Baseline at commissioning. A rhythm matched to load density. And a trend report that compares every sample against your own loop's starting point, because a number without your baseline is just a number.

A fill and top-up discipline. What may be mixed, what may not, and what to test after every top-up. The cheapest failures to fix are the ones caught at the sample bottle.

Answers that name their boundaries. Where a question belongs to the cold plate maker or the CDU maker, we say so. Where it belongs to us, you get a straight answer and the documents behind it.

Final thoughts

A data hall is not one cooling system. It is three loops wearing one building, and each loop hands the fluid a different job description. The cold plate loop wants cleanliness and a watched chemistry. The facility loop wants freeze margin that survives the design winter. The rear-door loop wants both, in a bigger volume, at a friendlier filtration level.

None of the three jobs is exotic. All three punish a specification that was written once, for a different loop, and copied forward.

Glacier Coolant builds heat-transfer fluids for circulating loops — the section between the machine and the point of use, which is exactly where these three circuits live. Tell us which loops your campus runs and what each one touches. We will answer per loop: a grade, a reason, a sampling rhythm and a baseline to compare against.