Why India's AI Ambitions Cannot Be Achieved Without Liquid-Cooled Data Center Infrastructure - Nasscom

Liquid cooling uses a fluid to carry heat away from servers. In direct-to-chip, or D2C, systems, coolant flows through cold plates attached directly to GPU dies. This puts the heat-removal system beside the hottest components instead of relying mainly on room air. The approach matters because AI GPUs now operate at much higher power levels than traditional processors.

A modern GPU rack can exceed 120 kilowatts. D2C cold plates absorb heat from the GPU and transfer it into the liquid loop. The warmed coolant then moves away for cooling before returning to the rack. This targeted process supports dense hardware while reducing the electricity needed for fans and air conditioners.

The article says D2C facilities can reach PUE levels as low as 1.09, even in challenging climates. New Indian campuses are being designed around liquid cooling because Hyderabad and Visakhapatnam can exceed 35°C. Cooling is becoming foundational to affordable AI computing.

Traditional data-center racks were built for roughly 5–15 kilowatts. A modern NVIDIA GB200 NVL72 rack reaches more than 120 kilowatts. That means heat density has risen by about eight to 40 times, depending on the comparison. Upcoming Vera Rubin and GB300 generations are projected to reach 150–600 kilowatts per rack.

The article gives a striking comparison. One rack of NVIDIA’s latest AI servers generates as much heat as 30 residential furnaces running simultaneously. It also draws the same power as 80 American homes. These figures reflect how nearly all electrical power used by computing equipment ultimately becomes heat that operators must remove.

This growth is reshaping facility design. Older sites were built around CPUs, air conditioning, and lower rack densities. AI campuses in Mumbai, Hyderabad, Pune, and Visakhapatnam are instead being designed for 100-kilowatt-plus racks, liquid cooling, and renewable-power-paired infrastructure.

Air cooling removes heat by moving room air across components and then conditioning that warmed air. Traditional facilities were designed for much lower densities, around 5–15 kilowatts per rack. AI racks now exceed 120 kilowatts, concentrating far more heat in the same physical space. The cooling system must therefore move much more heat continuously.

The article identifies practical thresholds. Once a GPU exceeds 700 watts, air cooling is no longer technically viable. Above 35 kilowatts per rack, direct-to-chip liquid cooling becomes mandatory. Above 100 kilowatts, immersion cooling is described as the only way to keep PUE under control. Liquid systems carry heat more directly and efficiently than air.

This is already affecting Indian infrastructure. Air-cooled sites can spend 30–40% of their electricity on cooling. Sustained AI inference loads make the problem worse because they run continuously. Retrofitting old air-cooled shells may therefore limit performance or produce uneconomic energy costs.

Efficient cooling determines how much computing equipment a data center can run at full performance. If a rack cannot remove heat quickly enough, its GPUs may need to be throttled. That reduces available AI capacity even when the servers and chips are installed. Cooling therefore becomes a direct limit on India’s ability to operate AI infrastructure.

The financial effect is equally important. Air-cooled facilities can spend 30–40% of their total electricity on cooling. A poor PUE means more power supports fans, pumps, and air conditioning rather than computation. For operators running thousands of GPUs, the difference between a PUE of 1.5 and 1.1 can determine whether the infrastructure remains profitable.

India’s planned expansion makes the stakes larger. IT load capacity is projected to rise from about 1.4 gigawatts to 6.5–17 gigawatts by 2030. If cooling remains inefficient, energy bills rise, compute is constrained, and AI campuses may fail to deliver their planned capacity economically.

India’s climate raises the cooling challenge. Cities such as Hyderabad and Visakhapatnam regularly exceed 35°C. Free-air cooling works well in cooler climates, but it is not practical across most Indian geographies for much of the year. Higher outdoor temperatures make it harder and more expensive to reject heat from dense AI equipment.

Electricity supply adds another constraint. Cooling can consume 30–40% of an air-cooled facility’s electricity. India’s IT load capacity is about 1.4 gigawatts today and is projected to reach 6.5–17 gigawatts by 2030. Every percentage point saved on cooling can leave more electricity for actual AI workloads.

Expansion makes the decision urgent. India’s colocation market is forecast to add $7.31 billion in 2026, while operators are targeting up to 15 gigawatts within five years. New campuses are being built around liquid cooling, dense racks, and renewable power instead of simply retrofitting older air-cooled facilities.

Air cooling remains suitable for lower-density equipment that stays within the older 5–15 kilowatt rack range. It uses fans and air-conditioning systems, but its energy burden grows sharply as heat density rises. Free-air cooling can work in cooler climates, yet the article says it is not an option for most of the year across many Indian locations.

Direct-to-chip liquid cooling is the main higher-density alternative. Coolant runs through cold plates attached to GPU dies, removing heat directly from the hottest components. The article identifies D2C as mandatory above 35 kilowatts per rack. Immersion cooling places equipment in a cooling fluid and is described as the only way to keep PUE under control above 100 kilowatts.

These choices are shaping new campuses. Operators are designing facilities for 100-kilowatt-plus racks and liquid cooling from the start. Existing sites may continue air cooling at lower densities, but retrofitting them becomes harder as GPU power and sustained inference workloads increase.

Power Usage Effectiveness, or PUE, shows how much total electricity a data center consumes for each unit used by its IT equipment. A lower number means less power is spent on supporting systems such as cooling. The article uses PUE to explain why thermal management is an economic and capacity decision, not merely an engineering detail.

For example, a PUE of 1.5 means the facility uses 1.5 units of total power for every unit reaching computing equipment. At 1.1, it uses only 1.1 total units for that same computing load. The difference represents electricity that no longer goes to overhead. Direct-to-chip facilities can achieve PUEs as low as 1.09, according to the article.

That saved power can support more GPUs or reduce operating costs. This matters in India because the grid is already power-constrained and AI demand is expanding quickly. Better cooling lets operators run dense racks without letting cooling consume the margin or limit computation.