Liquid cooling for AI racks in India: a practical guide

For most of the last decade, an enterprise rack drew somewhere between 5 and 15 kW, and moving cold air through a contained aisle was enough to keep it healthy. AI hardware has broken that assumption. A single NVIDIA GB200 NVL72 rack draws roughly 120–132 kW, according to Supermicro's and Gigabyte's published specifications and SemiAnalysis's estimate. That is ten times the old norm, in the same floor space.

Where air runs out

There is no single hard limit, but industry sources consistently place the practical ceiling for air cooling at around 20 to 35 kW per rack, depending on containment and room design. Uptime Institute's 2026 Global Data Center Survey found a growing number of operators already reporting peak rack densities of 30 kW or more. Above that line, the volume of air required becomes impractical: fans get louder and hungrier, hot spots appear, and the hall's cooling plant works harder for diminishing returns.

The three options, in order of disruption

Rear-door heat exchangers replace a rack's back door with a water-cooled coil. Server fans still push air, but the heat is captured at the door instead of spilling into the hall. Published guidance puts their useful range at roughly 40 to 70 kW per rack. Strictly, this is still air cooling with a liquid assist — which is exactly why it is popular: it fits existing racks and existing servers, and it is the gentlest retrofit.

Direct-to-chip cooling bolts cold plates onto the hottest components — GPUs and CPUs — and pumps coolant through them. Depending on the design it removes roughly 70 to 90 percent of a rack's heat at the source; the remainder (memory, storage, power supplies) is still handled by air. This is the route the current generation of AI racks is built for, and the GB200 NVL72 ships designed around it.

Immersion cooling submerges entire servers in a dielectric fluid. It handles the highest densities and removes almost all fans, but it changes everything around it: tanks instead of racks, much heavier floor loads, different maintenance procedures, and fluid supply chains. Not every GPU vendor endorses immersion for its newest silicon, and some two-phase fluids face tightening regulation. For most Indian operators today, it is a specialist choice rather than the default.

The question is rarely 'air or liquid'. It is 'which racks need liquid, and how do we run both side by side'.

Warm water, and why it matters in India

ASHRAE classifies liquid-cooling systems by the warmest facility water they accept — W17, W27, W32, W40, W45 and W+ in its current naming. Hardware rated for warmer supply water lets the facility reject heat with dry coolers or cooling towers for more of the year and lean on chillers less. That matters in hot, humid coastal cities such as Chennai — which delivered half of India's new data-centre capacity in the first half of 2026, according to JLL — where cold-water designs would keep chillers running for much of the year. Choosing warm-water-capable hardware and designing the loop around it is one of the biggest efficiency decisions in an AI hall.

A checklist before the first rack is powered on

Power and cooling budget per rack: confirm the real draw at full training load, not the nameplate average, and size the coolant distribution units (CDUs) with redundancy so one failure does not throttle a cluster.

Water quality and the loop: secondary-loop chemistry, filtration and corrosion control are what keep cold plates from clogging over years. Flush and pressure-test the loop before any server is connected.

Leak detection and isolation: sensors under manifolds and at every quick-disconnect, with valves that isolate a single rack without draining a row.

The air that is left: direct-to-chip still leaves 10 to 30 percent of the heat to air. The hall's air system must be sized for that remainder — it does not disappear.

Commissioning under real load: burn the cluster in at full utilisation while logging supply and return temperatures, flow and pressure. Problems that never show at idle appear within hours at 100 percent.

Monitoring from day one: coolant supply and return temperature, flow rate, differential pressure and rack inlet temperature belong on the same dashboard as power. Drift in any of them is the earliest warning of a failing pump, a blocked filter or a rack that is about to throttle.

Where Zyvark fits

We do not manufacture cooling equipment. We work alongside the cooling OEM and the facility team on design-assist, installation, commissioning and day-two thermal operations — and NetPulse brings the coolant and power telemetry into one view. If you are planning racks above 25 kW, a short assessment before procurement is the cheapest point to get the design right.

Sources

Supermicro, SRS-GB200-NVL72 specification (≈132 kW per rack). Gigabyte, GB200 NVL72 GIGAPOD specification (up to 130 kW per rack). SemiAnalysis, GB200 hardware architecture (≈120 kW per rack). Uptime Institute, Global Data Center Survey 2026, press release (28 July 2026). ASHRAE, Thermal Guidelines for Data Processing Environments, 5th edition (liquid-cooling W classes), as summarised by Upsite Technologies and DatacenterDynamics. Data Center Knowledge and GRC (Green Revolution Cooling) guidance on rear-door, direct-to-chip and immersion ranges. JLL, India data centre 2026 mid-year report (31 August 2026). Ranges are as published; individual designs vary.