A cold plate on a modern AI GPU carries coolant through microchannels a fraction of a millimetre wide. Whatever fluid flows through those channels for the next five to seven years determines whether the cooling system holds its design performance or slowly loses it to corrosion, deposits and microbial growth. Most liquid cooling discussions focus on capacity: kilowatts per rack, CDU sizing, supply temperatures. Far less attention goes to the fluid itself and the condition it is in after the first year of operation.
The short answer to the fluid question is this. DI water gives the best thermal performance and the lowest pumping energy, but it has no built-in protection against corrosion or bacteria and needs continuous purity control. PG25 trades a small amount of thermal performance for an inhibitor package and resistance to microbial growth, which is why it has become the industry reference fluid. Immersion removes water from the IT hardware altogether, but it changes the building, the servers and the way the hall is serviced. Whichever fluid is chosen, the secondary loop is a closed chemical system that changes over time, and only monitoring tells you when it has moved outside its safe range.
How a CDU loop is arranged
ASHRAE TC9.9 and the Open Compute Project (OCP) both recommend separating the facility water system (FWS) from the technology cooling system (TCS) through a Coolant Distribution Unit, which keeps the two liquids apart and allows a different fluid in each loop. The facility side is fed from chillers, dry coolers or cooling towers. The technology side is the loop that runs through the rack manifolds, quick disconnects and cold plates.
This separation matters for the fluid decision. The FWS can tolerate ordinary treated water. The TCS cannot, because its passages are narrow, its metals are mixed (copper cold plates, brazed joints, stainless manifolds) and any failure lands directly on the IT hardware.
DI water: best heat transfer, highest maintenance burden
Deionised water is the thermally superior fluid. At 20 to 40°C, DI water has a specific heat of about 4.18 kJ/kg·K against roughly 3.85 for PG25, around 8% higher. PG25’s viscosity is about 2.5 times that of water at 20°C, and its thermal conductivity is roughly 19% lower. Lower viscosity means lower pressure drop across the cold plates and less pump energy. Water is also non-toxic and simple to dispose of.
The problem is that pure water is chemically hungry. With almost no dissolved ions, it pulls ions from whatever metal it touches, which means copper and aluminium surfaces corrode as the water moves toward equilibrium. It also offers no resistance to bacteria, so any microbes introduced during fill, top-up or maintenance can colonise the loop. A DI loop therefore needs continuous purification (typically mixed-bed polishing), tight control of conductivity, and a proper plan for microbial control.
Some OEMs prefer it regardless. Lenovo, for example, prefers deionised water for its Neptune systems where applicable, citing PG25’s lower heat transfer efficiency, higher viscosity and disposal considerations, while still accepting PG25. Notably, Lenovo ties its highest-performance DI water tier to continuous fluid pH and chemistry monitoring.
PG25: the industry reference fluid
PG25 is propylene glycol blended at about 25% by volume with water and a corrosion inhibitor package designed for the metals in the loop. It is the reference fluid in both ASHRAE TC 9.9 guidance and OCP standards for single-phase direct-to-chip cooling, and several major server OEMs specify it for the technology cooling loop.
The 25% concentration is not arbitrary. OCP guidance specifies a glycol content of 24.5 to 29.5% by volume for PG25, and notes that biocides are not necessary while glycol concentration stays above 25%, because the fluid is then considered bio-static. PG25 fluid pH is specified between 8.0 and 10.5, and the blend provides freeze protection to about -10°C.
In Malaysia, the case for PG25 looks different from the case in colder markets. Freeze protection is irrelevant for an indoor loop in the tropics (it only matters during shipping or storage), so the value of PG25 here rests almost entirely on its inhibitor package and its bio-static behaviour. Both depend on the glycol concentration staying in specification. A PG25 loop that has been topped up with plain water over several maintenance cycles can fall below 25%, at which point it loses its bio-static margin and its inhibitor strength together, with no visible sign at the rack.
The thermal penalty is real but manageable. PG blends need flow rates raised by roughly 15 to 20% compared with water because of their lower heat capacity and higher viscosity, which the CDU pump selection must account for from the start.
Immersion: removing water from the hardware entirely
Immersion cooling submerges server boards in a dielectric fluid. It can push power usage effectiveness close to 1.0 and eliminates server fans. Single-phase immersion keeps the fluid liquid throughout, using vegetable-derived, mineral oil or synthetic dielectric fluids. Two-phase systems use engineered fluids that boil at the chip surface, and these face growing regulatory pressure on fluorinated chemistries.
The limitations are practical rather than thermal. Tanks are heavy and demand floor loading most existing halls were not designed for. Servicing a server means lifting it out of the fluid. Not every server platform is warrantied for immersion, and materials such as cable jackets, labels and thermal interface compounds must be checked for compatibility with the fluid. Most current AI GPU platforms are designed around cold plates, not tanks.
Immersion also still uses a CDU. The dielectric fluid transfers its heat through a heat exchanger to the facility water loop, so the water side of an immersion installation still needs water chemistry control.
How the three options compare
| Criteria | DI water | PG25 | Single-phase immersion |
|---|---|---|---|
| Heat transfer | Highest | Good (about 8% lower heat capacity than water) | Lower per volume, but whole-board contact |
| Pumping energy | Lowest | Higher (15 to 20% more flow needed) | Depends on tank design |
| Built-in corrosion protection | None unless inhibited | Inhibitor package | Dielectric fluid is non-corrosive |
| Resistance to microbial growth | None | Bio-static at 25% and above | Not applicable to IT fluid |
| Maintenance burden | High (purity and microbial control) | Moderate (concentration and inhibitor control) | Fluid management and handling |
| Typical use | OEM platforms specifying water | OCP reference, most OEM specifications | Specialist and high-density deployments |
What is being used in Malaysia
Direct-to-chip cooling is the dominant liquid architecture in Malaysia’s new AI capacity, most of it concentrated in Johor. Market reports indicate that YTL Data Centre’s JDC2 facility allocated around 80% of its 100 MW IT load to direct-to-chip loops, while immersion remains a small share of the market. AI data centre operators generally favour direct-to-chip cooling, although a 65 MW immersion-cooled data centre has been announced for Cyberjaya, funded by state-backed developer Selangor Industrial Corporation.
On the fluid side, there is no public breakdown of PG25 versus DI water deployments in Malaysia. In practice, the TCS fluid follows the server OEM’s specification, which is why PG25 appears in most hyperscale direct-to-chip installations and DI water appears where the OEM platform calls for it.
Immersion has a local development story of its own. MPOB is developing Sawit EcoTherm, a palm oil-based single-phase immersion coolant that remains at the pre-commercial stage pending pilot projects and validation with operators.
The facility water side is where Malaysia’s conditions are changing fastest. Johor’s approval framework now weighs water usage effectiveness alongside power and connectivity. Operators and state water companies are developing recycled water schemes, and plants to treat municipal effluent for data centre cooling are under way in Johor. Reclaimed water carries more nutrients, more variable chemistry and more microbial load than potable supply, which raises the stakes for FWS water treatment and for any heat exchanger leak that could carry that water toward the technology loop.
Why coolant monitoring matters
A TCS loop is often treated as a fill-and-forget asset. It is not. Coolant chemistry changes throughout its service life, through several mechanisms that run at the same time.
Glycol oxidises under heat and dissolved oxygen, forming organic acids that pull pH down and consume the inhibitor package. As inhibitor depletes, copper corrosion accelerates, and corrosion products circulate as particulates that settle in the narrowest passages, which are the cold plate microchannels. Water top-ups after maintenance or quick-disconnect weeps dilute the glycol concentration. A pinhole leak in the CDU’s plate heat exchanger can let facility water into the technology loop, bringing hardness, chlorides and bacteria with it. In DI loops, and in PG25 loops that have fallen below 25%, bacteria that enter during fill or service can colonise surfaces and form biofilm.
Many systems monitor only a single parameter, such as conductivity or glycol concentration alone, and rely on periodic lab samples for the rest. Quarterly sampling catches a trend eventually. Continuous inline monitoring catches the event.
What each parameter tells you
Pyxis Lab has developed monitoring specifically for CDU loops. Its KRYOPTIX IK-1500 Series is a pre-engineered panel that measures pH, temperature, conductivity, turbidity and glycol concentration simultaneously, with each sensor providing 4-20 mA and RS-485 Modbus outputs for direct integration with CDU controls, PLCs, BMS or DCIM platforms. Each parameter answers a different question.
Glycol concentration. This confirms the fluid is still the fluid it was specified to be. The Pyxis PRISM RT-110 inline refractometer measures glycol concentration continuously with an accuracy of plus or minus 0.1%. A falling reading signals dilution from water top-ups and warns that the loop is approaching the 25% threshold below which it loses its bio-static protection and its inhibitor strength. Excess glycol, on the other hand, reduces cooling efficiency and increases pumping energy.
pH. This is the clearest indicator of inhibitor health and glycol degradation. A downward pH trend points to organic acid formation from glycol oxidation and to inhibitor consumption, both of which precede copper corrosion. The target band depends on the fluid supplier’s formulation, so monitoring setpoints should follow the fluid datasheet.
Conductivity. In a DI loop, conductivity is the primary purity measurement, and it needs an ultra-low range sensor because the values involved are very small; any rise means ions are entering the water from corrosion or contamination. In a PG25 loop, conductivity is naturally higher because of the inhibitors, so the value of monitoring lies in the trend. A sudden step change is a strong signal of cross-contamination from the facility water side.
Turbidity. This detects particulates: corrosion products, debris from installation, precipitated inhibitor or microbial growth. Because microchannel cold plates clog far more easily than conventional heat exchangers, rising turbidity is an early warning of flow restriction at the chip.
Temperature. Temperature underpins every other measurement through compensation, and it also governs how fast the fluid degrades. Sustained high return temperatures accelerate glycol oxidation.
Corrosion rate. Where the loop contains copper and mixed metals, a corrosion rate measurement shows directly whether the inhibitor programme is protecting the metal, rather than inferring it from pH and conductivity.
Biofilm: the parameter chemistry sensors cannot see
pH, conductivity and turbidity describe the bulk fluid. Biofilm grows on surfaces, and by the time it affects bulk fluid readings it is already established. In a cooling loop, biofilm restricts flow through narrow passages, insulates heat transfer surfaces and drives microbiologically influenced corrosion (MIC), which attacks copper and stainless steel at the metal surface underneath the film.
The Alvim biofilm sensor uses an inline electrochemical electrode that detects bacteria as soon as they settle on the probe surface, before a slime layer forms, with a signal that rises as the biofilm grows. It responds from as little as 1% surface coverage, and its data can be used to adjust biocide or cleaning treatments and to verify whether they worked. The same measurement helps prevent both MIC and the corrosion caused by overdosing chlorine and other oxidising chemicals.
Placement needs some thought. For biofilm sensing, the Alvim industrial sensor is specified for conductivity above 10 µS/cm and dissolved oxygen above 1 ppm at maximum sensitivity, and it outputs 4-20 mA with RS485 Modbus communication. That makes it a natural fit on the facility water side, especially where reclaimed water or cooling towers are involved, and on PG25 loops where the inhibitor package keeps conductivity well above that threshold. For a high-purity DI loop running at very low conductivity, the installation point should be assessed case by case.
The biofilm case is strongest in two situations. The first is DI water loops, which have no bio-static protection at all. The second is PG25 loops where glycol concentration has fallen below 25% through dilution, because the fluid has lost the property that made biocide unnecessary. In both cases, a biofilm signal is the earliest evidence that microbial control is failing, arriving well before turbidity rises or flow through the cold plates drops.
Matching the monitoring to the fluid
For a DI water loop, the priorities are ultra-low range conductivity, pH, turbidity and biofilm, because purity and microbial control are the two things that fail first.
For a PG25 loop, glycol concentration is the anchor parameter, since it confirms both the bio-static margin and the inhibitor strength. It is supported by pH, conductivity and turbidity, with biofilm monitoring as the backstop for the day concentration slips below specification.
For immersion, the dielectric fluid itself is managed through fluid supplier analysis, but the CDU’s facility water side still needs the same water chemistry and biofilm monitoring as any other cooling water system.
Across all three, the facility water loop deserves as much attention as the technology loop, particularly as Johor’s data centres move toward recycled and reclaimed water sources. The CDU heat exchanger is the only barrier between the two, and monitoring both sides is how an operator knows that barrier is holding.
Autoflo Technology is the Malaysian distributor for Pyxis Lab coolant monitoring solutions and Alvim biofilm sensors. For help specifying a monitoring package for a CDU technology loop or facility water system, contact us at info@autoflotechnology.com.