Liquid Cooling for Flagship Racks in India: CDUs and Water Quality
Overview
Above roughly 50 kW per rack, air cooling stops being an engineering choice and becomes an impossibility, which is why every flagship AI rack ships liquid-cooled. In India the technology is arriving quickly — Asia Pacific is reported to hold over 38 percent of the coolant distribution unit market in 2026, and a Bengaluru manufacturing facility for liquid-cooling products opened in February 2026. What Indian operators consistently underestimate is not the equipment but the discipline: a direct-to-chip loop is a chemical and mechanical system that must be maintained to specification for years, and the consequences of not doing so are corrosion and blocked cold plates.
Key takeaways
- Two loops, not one — a clean technology cooling loop to the chips, and a facility water loop to the outside world.
- The CDU is the isolation point — it transfers heat between loops and keeps facility water away from the servers.
- Water chemistry is non-negotiable — DI or RO water with conductivity typically below 5-10 microsiemens per centimetre.
- Indian ambient conditions hurt heat rejection — dry-cooler capacity must be sized for peak summer, not annual average.
- Leak detection and service response belong in the contract, not in a runbook written after commissioning.
How the loop is structured
A direct-to-chip system has two hydraulically separate circuits. The technology cooling system carries treated coolant from the CDU through manifolds to cold plates mounted directly on GPUs, CPUs and other high-power components, and back. The facility water system carries heat from the CDU to the building’s heat rejection plant — dry coolers, cooling towers or a chiller.
The CDU sits between them with a heat exchanger, pumps, filtration and instrumentation. Its separation function is the important one: facility water is comparatively dirty and variable, and allowing it anywhere near a cold plate would foul the narrow channels that make direct-to-chip cooling work. Everything on the technology side is a controlled environment; everything on the facility side is ordinary building services.
Sizing a CDU
| Parameter | What it governs | Common error |
|---|---|---|
| Heat rejection capacity (kW) | Total IT load served | Sized to nameplate, not to peak plus growth |
| Approach temperature | How close TCS gets to FWS temperature | Ignored until summer performance disappoints |
| Flow rate and pressure drop | Number of racks and manifold design | Long runs add drop the pump cannot overcome |
| Redundancy configuration | Behaviour during maintenance or failure | N configuration on a critical load |
| Filtration specification | Particulate protection for cold plates | Filters not on a maintenance schedule |
The approach temperature row matters more in India than in temperate climates. A CDU cannot deliver coolant colder than its facility water plus the approach; if facility water rises in May because the dry coolers are working against a 45-degree ambient, the technology loop rises with it, and GPUs throttle. Sizing the heat rejection plant against peak summer conditions rather than an annual average is the difference between a system that works all year and one that works from October to March.
Water chemistry is the maintenance item
This is where liquid-cooled deployments fail slowly. The technology cooling loop requires deionised or reverse-osmosis water, typically with conductivity held below 5 to 10 microsiemens per centimetre, plus corrosion inhibitors and often a biocide. Ordinary treated building water is not acceptable, and neither is topping up a DI loop with tap water when a technician cannot find the DI supply.
Three failure modes follow from poor chemistry. Galvanic corrosion where dissimilar metals meet, which attacks fittings and eventually cold plates. Scale deposition inside the narrow cold-plate channels, which raises thermal resistance gradually until components throttle for no visible reason. And biological growth, which blocks filters and, in the worst case, channels. None of these announce themselves early; they present as slowly degrading performance months after commissioning.
What Indian sites specifically must plan for
Four items. Water availability and treatment: a DI or RO plant is part of the facility, not an accessory, and its capacity, consumables and maintenance need a budget line. Ambient design: peak dry-bulb and wet-bulb conditions at the site govern heat rejection sizing, and in much of India those are severe enough to make dry coolers alone marginal at high density.
Power quality for the pumps: a CDU pump failure at a megawatt-class load is a fast thermal event, so pumps belong on protected power with genuine redundancy. And skills: Indian data centre teams have deep experience with air cooling and comparatively little with hydronics, so the operations team either needs training or the service contract needs to cover routine chemistry and filter maintenance explicitly. The facility readiness context is in hosting 100 kW racks in India.
Leak management
The objection every facility manager raises is leaks, and it deserves a straight answer: leaks happen, and the design assumes it. Modern implementations use dripless quick-disconnect couplings so servers can be removed without opening the loop, leak-detection cabling along manifolds and under racks, and automatic isolation valves that shut a segment on detection.
What matters operationally is the response procedure and who executes it. A leak detected at 2 a.m. needs someone who knows which valve isolates which manifold and what to do about the racks downstream. Write that procedure during commissioning, rehearse it, and make service response time a contractual term rather than a best effort. The day-2 discipline generalises from day-2 operations for rack-scale AI.
What belongs in the contract
Five clauses. Coolant chemistry specification with a testing schedule and who performs it. Filter replacement intervals and consumable supply. Guaranteed CDU performance at the site’s design summer conditions, not at a laboratory ambient. Leak response time with a defined escalation path. And spares holding for pumps, couplings and control components, since import lead times for hydronic components are long and a failed pump on a critical loop cannot wait a month.
One design note worth deciding early: whether to standardise on a facility-level CDU serving many racks or in-rack CDUs serving one each. Facility CDUs are more efficient and easier to maintain centrally; in-rack units contain failure to a single rack and suit incremental deployment. For most Indian enterprises deploying two to six flagship racks, in-rack or small-cluster CDUs reduce the facility engineering burden considerably. The air-versus-liquid decision threshold is covered in air-cooled vs liquid-cooled GPU racks.
Frequently asked questions
At what density does liquid cooling become mandatory?
Broadly above 50 kW per rack, air cooling ceases to be practical, and flagship AI racks at 120 kW and beyond ship liquid-cooled by design. Below that, high-density air with containment or rear-door heat exchangers remains viable.
What does a CDU actually do?
It separates the clean technology cooling loop that reaches the cold plates from the facility water loop, transferring heat between them through a heat exchanger while providing pumping, filtration and instrumentation. That separation keeps variable facility water away from narrow cold-plate channels.
Why is water quality so important?
Because poor chemistry causes galvanic corrosion, scale deposition in cold-plate channels and biological growth, all of which develop slowly and present as unexplained thermal degradation months later. Specifications typically call for DI or RO water with conductivity below 5 to 10 microsiemens per centimetre.
What is different about liquid cooling in India?
Peak ambient conditions make heat rejection harder, so dry coolers must be sized for summer rather than average conditions or coolant temperature rises and GPUs throttle. Water treatment capacity, protected power for pumps, and hydronics skills in the operations team all need explicit planning.
How are leaks managed?
Through dripless quick-disconnect couplings, leak-detection cabling along manifolds and under racks, and automatic segment isolation. The operational half matters equally: a written, rehearsed response procedure and a contractual service response time rather than a best-effort arrangement.
Ready to deploy?
Talk to an RDP architect about power, cooling and lead time.