What Coolants Are Used in AI Data Center Liquid Cooling

Summary of Coolant Families

The architecture choice comes first and the coolant follows. This article focuses on the properties and selection criteria within each family.

PG25: the AI cooling default

What it is

PG25 is 25 % propylene glycol by volume, 75 % treated (typically deionized) water, plus a corrosion inhibitor package. Propylene glycol is used rather than ethylene glycol because of its substantially lower toxicity — a meaningful consideration where a leak inside a data hall is a realistic scenario and where personnel exposure during maintenance is routine.

Why it is the default

Three reasons, in order of importance:

1. Freeze protection. PG25 freezes at approximately −10 to −13 °C depending on the exact formulation and inhibitor package. This matters not only for cold climates but for any loop that can be isolated and left static — a rack under maintenance, a transport‑stage deployment, a plant shutdown over a winter weekend. Water that freezes in a microchannel cold plate destroys it, and the damage is not always visible until the loop is refilled and pressurized.

2. Corrosion inhibition. The inhibitor package — not the glycol — does this work. Modern data center coolants use OAT (organic acid technology) chemistry rather than the older silicate‑based IAT formulations. The distinction is important for AI hardware: silicates can form gel‑like deposits that foul microchannel cold plates, whereas OAT inhibitors form thin protective films on each metal surface. Specify silicate‑free, OAT inhibited coolant for any loop containing cold plates.

3. Microbiological stability. At 25 % concentration, PG is generally biostatic — microorganisms cannot readily proliferate. Below roughly 10 % glycol, the glycol itself becomes a carbon source and biological fouling becomes a real risk. This is why under‑dosing a loop "to save money" or "to improve heat transfer" backfires.

What it costs you

The thermal penalty is real and should be designed around, not ignored. At approximately 40 °C:

Three design consequences follow:

• Flow rates must increase for the same heat rejection, because ρcₚ is lower. Roughly 5‑10 % more volumetric flow than a water design.

• Pumping power roughly doubles at equal flow, because viscosity is roughly twice water's. At scale this is a real operating cost and a real contributor to PUE.

• Heat exchanger and cold plate effectiveness drops, because thermal conductivity is ~22 % lower. Approach temperatures at the CDU heat exchanger and junction‑to‑coolant resistances at the cold plate both degrade. Vendors' published performance

ASHRAE W‑class supply temperature classes

Single‑phase immersion fluids

Typical properties

The thermal penalty is severe. Specific heat is roughly half that of water and thermal conductivity roughly one‑third. Immersion compensates by moving much larger volumes of fluid and by wetting the entire component surface rather than just a cold plate footprint. It works, but it is not thermally efficient per unit of pumped volume.

What immersion buys you

• No cold plates, no manifolds, no quick disconnects — the whole server is cooled, including memory, VRMs and drives that cold plates leave to air

• Near‑total heat capture at the fluid, enabling high‑grade heat reuse

• No fans in the server, eliminating a significant noise and energy source

• Inherent leak tolerance — the fluid is already where the electronics are

What immersion costs you

• Fluid cost and volume — a tank of synthetic hydrocarbon for a rack is a substantial capital item

• Material compatibility — elastomers, cable jackets, adhesives and some plastics swell, soften or embrittle in hydrocarbon fluids. Every component must be verified.

• Service mess — a server withdrawn from a tank is coated in oil and requires cleaning before handling

• Weight and floor loading — a fluid‑filled tank is significantly heavier than an air cooled rack

Two‑phase dielectric fluids

Galden PFPE fluids, Standard Fluids' fluoroketone products, and supply from Chemours, AGC and Chinese manufacturers are all in the market. Some are positioned as drop‑in replacements; none should be assumed to be one without verifying material compatibility, thermal performance and — critically — long‑term supply commitments in writing.

The specification action: if you are evaluating two‑phase immersion, obtain a documented fluid supply strategy covering a ten‑year horizon, including the manufacturer's regulatory position and your contractual remedies if supply ceases. A cooling architecture whose working fluid may become unavailable is a stranded‑asset risk, not a technology choice.

Selection criteria, ranked

1. Cooling architecture. Cold plate or immersion? This determines the family before anything else. Cold plate → water‑based. Immersion → dielectric.

2. Temperature envelope. What ASHRAE W‑class does the IT equipment require? What are the minimum and maximum coolant temperatures across the year, including shutdown and transport conditions? This sets the freeze protection requirement and hence whether glycol is needed.

3. Heat flux and thermal performance. Junction‑to‑coolant thermal resistance and cold plate heat flux capability depend directly on k and cₚ. High‑heat‑flux platforms (500‑600 W/cm² at the die interface is cited for current GB200‑class hardware) constrain the fluid more than low‑density servers do.

4. Pumping power. Viscosity drives it. PG25 costs roughly 2× water; immersion fluids vary widely. Over a 10 MW facility this is a PUE line item, not a rounding error.

5. Material compatibility. Every wetted material — copper, stainless, aluminium if present, EPDM or FKM elastomers, PTFE, cable jackets, adhesives — must be confirmed against the specific fluid and inhibitor package. Elastomer selection is a common failure point: EPDM suits water‑glycol well but swells in hydrocarbons; FKM suits hydrocarbons but is attacked by some polar fluids.

6. Dielectric requirement. Mandatory for immersion. For cold plate systems, a non‑dielectric fluid means a leak is immediately an electrical event — which is why leak detection, drip trays and secondary containment are specified rather than relying on the fluid.

7. Water chemistry control capability. Can you actually monitor and maintain pH, conductivity, chloride, glycol concentration and microbial count over ten years? If not, choose the coolant that is most forgiving of neglect.

8. Regulatory and supply security. PFAS status, regional restrictions, manufacturer commitment, second‑source availability. Increasingly the deciding factor.

9. Lifecycle cost. Fluid price per litre × volume × refresh interval. Inhibited glycol is commonly refreshed on a 3‑7 year cycle depending on the formulation and monitoring results. Immersion fluids can last longer but cost far more per litre initially.

10. Environmental and safety profile. Toxicity, biodegradability, disposal route, flash point, and what happens in a spill inside a live data hall.

Commissioning and maintenance practice

Pre‑commission flushing. New loops contain manufacturing residues — drawing compounds, welding oxides, particulate, oils. Flush with deionized water (commonly ≥0.5 MΩ∙cm resistivity) until the effluent meets the target (often cited as below ~5 µS/cm) before introducing the final coolant. OCP has published pre‑commission preparation guidance for technology cooling system manifolds that is worth following even where not contractually required.

Verify glycol concentration at fill. Use a refractometer calibrated for propylene glycol. A 20 % fill where 25 % was specified loses freeze protection and can shift inhibitor performance.

Establish a baseline. Record pH, conductivity, chloride, glycol concentration, microbial count and dissolved oxygen (if measured) at commissioning. Every subsequent result is meaningless without it.

Trend, do not just test. A quarterly sample that is always "within limits" tells you nothing. A rising trend tells you everything.

Plan for coolant change‑out. Inhibited glycol degrades. Inhibitors deplete, glycol oxidizes, pH drifts, and the fluid's corrosion protection falls away gradually — long before it fails visibly. Budget for a change‑out in the 3‑7 year window, and design the system with drain points, sample ports and a fill connection that make the job practical on a live plant.

Frequently asked questions

Is PG25 mandatory for AI cold plate cooling?It is the overwhelmingly common specification and the formulation referenced across current NVIDIA‑class platforms, but it is not the only option. Treated water with an inhibitor package is used where freeze protection is not required. Always follow the IT equipment vendor’s requirements.