Air vs liquid: The cooling challenge behind India's 1-GW data centres

After months of local protests, Google has reportedly agreed to use air-cooling for its planned 1-GW data centre in Visakhapatnam district. Tata Consultancy Services (TCS) has also announced a similar-scale facility, named HyperVault, in the same region. While air-cooling cuts water use, it brings its own engineering and economic challenges.

A 1-GW data centre, in principle, generates 1 GW of heat. That heat originates in billions of transistors inside millions of processors. Each transistor releases heat as electrical current flows through it, and the cumulative effect across a facility is enormous. The first step in cooling is moving heat from the silicon chip to a heat sink, but what happens next depends on the cooling technology chosen.

Air-cooling is the conventional approach. Fans push cool air through server racks, and the hot exhaust is collected and cooled. Variants include computer-room air conditioners (CRACs), chilled-water air handlers (CRAHs), hot-aisle/cold-aisle containment, and free-cooling, which uses naturally cooler outside air. Rear-door heat exchangers attach a water-cooled coil to the back of each rack, improving efficiency.

Liquid-cooling options are more effective per unit volume because water has a higher heat capacity than air. Cold-plate cooling places a metal plate with internal channels directly on the processor, circulating liquid to carry heat away. Immersion cooling submerges entire electronics in a non-conductive fluid, with single-phase or two-phase (boiling) variants. Evaporative cooling uses water evaporation in cooling towers, but demands large water volumes. Dry-cooling exposes heat exchangers to ambient air, avoiding water but requiring large surfaces, especially in warm climates. Other methods include chilled-water plants, geothermal rejection, natural water bodies, and heat reuse.

Air-cooling has lower upfront costs — liquid-cooled facilities may cost 7-10% more to build — and benefits from a mature maintenance ecosystem. For racks under 20 kW, air-cooling remains the preferred choice. In cool, dry climates, air-side economisers can cut mechanical cooling energy by 70%.

However, for a 1-GW facility, air-cooling alone is now considered physically and economically impractical. The key limitation is the 'thermal wall': air can remove at most about 40 kW per rack. Modern AI chips like Nvidia's Blackwell generate 700-1,000 W per GPU, pushing rack loads to 120-150 kW. Cooling such racks with air would require wind-tunnel-like setups, driving up power consumption. Noise is another issue — chillers and air handlers can produce up to 100 dB, and low-frequency hum can travel kilometres if not dampened.

Given these trade-offs, many large data centres today use hybrid cooling: air for low-density areas, rear-door heat exchangers to extend air-cooling to about 50 kW, and direct liquid cooling for high-density hotspots. The choice for Visakhapatnam will depend on balancing water scarcity, energy efficiency, and the performance demands of next-generation hardware.