AI Data Centre Cooling in India: Managing Power, Water, and Policy

Overview

Cooling next-generation artificial intelligence data centres presents a severe infrastructure challenge in India. High-density computing clusters generate immense thermal loads. Removing this heat requires up to 3,000 times more efficient heat transfer than traditional setups.

Modern artificial intelligence clusters draw between 40 kW and 120+ kW per server rack. This intense heat flux breaches the thermodynamic limits of legacy air cooling.

Meeting this demand requires a fast technological transition. Operators must pivot from water-intensive evaporative chillers to advanced closed-loop liquid cooling, captive renewable power, and enforceable resource-efficiency mandates.

India's computing infrastructure is expanding at an unprecedented pace, driven by structural policy support and intense enterprise demand:

• Surging Power Demand: National data centre electrical power demand is projected to scale from an operational base of 1.1 to 1.7 GW up to 26.3 GW by FY 2031–32.
• Strategic Policy Incentives: The Department of Economic Affairs granted Infrastructure Status to data centres exceeding 5 MW capacity in 2022. This classification unlocks institutional credit, external commercial borrowings, and lower-cost debt financing.
• Rapid Compute Expansion: As of July 2026, accelerated deployments under the national IndiaAI Mission and generative computing adoption have triggered multi-megawatt and gigawatt-scale construction across major metropolitan centres.

Hyperscale facilities require uninterrupted power and continuous thermal management. Without stable cooling, silicon processors throttle or suffer permanent thermal degradation.

Legacy data centres and modern AI clusters differ fundamentally in processor architecture and thermal concentration:

• Traditional Enterprise Facilities: Standard central processing units (CPUs) handle sequential web hosting, database management, and basic cloud compute tasks. Racks draw modest power densities of 5 to 15 kW per rack. Standard computer room air handlers circulate and exhaust this heat.
• High-Density AI Infrastructure: Massively parallel graphics processing units (GPUs) and accelerators, such as Nvidia H100 and GB200 configurations, process dense mathematical models. These clusters draw between 40 kW and 120+ kW per rack, packing unprecedented thermal loads into compact footprints.

Legacy forced-air cooling methods fail above 30 to 40 kW per rack. Air simply cannot circulate fast enough through tight chassis spaces to remove concentrated heat.

Silicon semiconductor dies concentrate intense heat flux over miniature surface areas. During continuous model training workloads, internal junction temperatures rapidly approach safe operating thresholds of 85 to 100 degrees Celsius.

Liquid cooling architectures bridge this thermodynamic gap:

• Superior Thermal Conductivity: Direct-to-chip liquid cooling and dielectric immersion dissipate heat with fluid thermal conductivity up to 3,000 times greater than air.
• Improved Energy Efficiency: Non-conductive dielectric fluids and chilled water blocks contact hot processor surfaces directly. This design pushes Power Usage Effectiveness (PUE) ratios down from legacy averages of 1.5–1.8 toward optimal ranges of 1.2–1.3.

Power Usage Effectiveness measures total facility energy divided by actual IT equipment energy. A score closer to 1.0 indicates minimal energy wasted on auxiliary cooling, fans, and power conversion.

The Central Electricity Authority faces serious grid-balancing challenges as round-the-clock hyperscale loads strain regional transmission systems. Hyperscale data centres run with baseload capacity factors exceeding 90 percent, unlike seasonal agricultural or commercial power demand.

Grid integration and clean energy supply involve several key mechanisms:

• Green Energy Open Access Rules (2022): The Ministry of Power lowered the open-access eligibility threshold from 1 MW to 100 kW for renewable consumers.
• Grid Stability and Storage: Intermittent solar and wind generation cannot supply 24/7 hyperscale uptime alone. Operators must install dedicated battery energy storage systems, off-site pumped hydro reserves, or dual-grid utility feeds.

Data centre geographic concentration creates severe local resource bottlenecks across India:

• Geographic Clustering: Over 65 percent of India's data centre capacity sits concentrated in three Tier-1 urban and coastal hubs: Mumbai-Navi Mumbai, Chennai, and Noida. These hubs face ongoing seasonal groundwater depletion and municipal water distribution stress.
• Evaporative Water Consumption: Evaporative cooling towers achieve low PUE figures by consuming between 1.8 and 9.5 litres of water per kWh of IT load.
• Closed-Loop Trade-offs: Dry-cooling and direct liquid systems drive Water Usage Effectiveness (WUE) near zero. However, they demand higher electrical power to run mechanical chillers and condensing fans during peak summer heat.

According to an industry study by CEEW and Financial Express (as reported in June 2026), annual water consumption by Indian data centres is projected to reach approximately 358 billion litres by 2030 unless operators adopt mandatory closed-loop architectures, wastewater recycling, or coastal seawater cooling.

Data centre operators evaluate thermal architectures across capital cost, spatial footprint, electrical consumption, and water utilisation rates.

The Ministry of Electronics and Information Technology (MeitY) circulated the Draft National Data Centre Policy to encourage domestic computing infrastructure through single-window clearances, Data Centre Economic Zones (DCEZs), and standardised building codes. However, the policy remained non-statutory through mid-2026, leaving environmental compliance governed by fragmented state guidelines.

The regulatory landscape shows significant enforcement gaps alongside strict international benchmarks:

• Domestic Standards Gap: The Bureau of Energy Efficiency (BEE) enforces the Energy Conservation Building Code (ECBC) under the Energy Conservation Act, 2001 (as amended in 2022). Yet data centres lack mandatory cooling caps and water ceilings under the Perform, Achieve, and Trade (PAT) scheme.
• State Policy Inconsistencies: State policies in Maharashtra, Tamil Nadu, and Uttar Pradesh provide land subsidies and electricity duty waivers. However, fewer than one-third of state policies mandate water recycling or captive zero-liquid-discharge (ZLD) plants.
• European Union Mandate: Directive 2023/1791 (Article 12) mandates that all data centres with an installed IT capacity of 500 kW or more publish annual audited metrics on PUE, WUE, and Energy Reuse Factor to a public database.
• Germany's EnEfG Standard: The Energy Efficiency Act legally requires new data centres commissioned from July 2026 onwards to achieve a target PUE of 1.2 or lower, alongside compulsory waste-heat reuse for municipal district heating.

The Supreme Court of India established in Subhash Kumar v. State of Bihar (1991) and affirmed in M.C. Mehta v. Kamal Nath (1997) that access to clean water is a fundamental right under Article 21 of the Constitution, governed by the Public Trust Doctrine. Under this doctrine, the State holds natural resources like rivers and aquifers in trust for public use.

Allocating potable municipal water to industrial cooling towers creates distributive friction in water-stressed regions:

• Domestic Water Inequity: Diverting millions of litres of daily freshwater to cool digital infrastructure during heatwaves shifts environmental burdens onto local farming communities and domestic consumers.
• Equitable Resource Governance: Public administration frameworks must balance digital sovereignty with constitutional commitments to equitable resource distribution and inter-generational environmental justice.

Sustainable expansion of India's computing infrastructure requires institutionalising statutory cooling standards, geographical decentralisation, and circular resource management across all hyperscale projects.

1. Mandatory Statutory Efficiency Standards: The Bureau of Energy Efficiency should incorporate hyperscale facilities as designated consumers under the PAT framework, setting mandatory PUE ceilings below 1.3 and capping potable water consumption.
2. Adoption of Closed-Loop and Immersion Systems: Industrial approvals should mandate direct-to-chip liquid cooling or closed-loop dielectric immersion for racks exceeding 30 kW, eliminating open evaporative cooling towers.
3. Mandatory Greywater and Tertiary Recycling: State pollution control boards should prohibit the extraction of potable groundwater for data centre cooling, requiring operators to utilise treated municipal wastewater or captive zero-liquid-discharge systems.
4. Spatial Decentralisation to Tier-2 Cities: State planning agencies should incentivise developers to build outside congested Tier-1 aquifers, establishing computing clusters in coastal zones that can utilise seawater heat exchangers or regions with cooler ambient climates.
5. District Heat Reuse Networks: Industrial zones should explore thermal co-location models where low-grade waste heat from liquid-cooled data centres supplies industrial processes, agricultural drying facilities, or commercial heating systems.

• Artificial intelligence clusters running advanced GPUs draw between 40 kW and 120+ kW per rack, breaching conventional air-cooling limits and necessitating direct liquid or immersion systems.
• India's data centre electrical demand is projected to surge from 1.1–1.7 GW to 26.3 GW by FY 2031–32 according to Ministry of Power parliamentary disclosures.
• Over 65 percent of operational and planned capacity sits concentrated in water-stressed metropolitan hubs including Mumbai, Chennai, and Noida.
• Evaporative cooling creates a direct resource trade-off by consuming 1.8 to 9.5 litres of water per kWh of IT load to achieve lower power overheads.
• Sustainable governance requires mandatory BEE efficiency targets, closed-loop cooling mandates, and strict protection of municipal aquifers under Article 21.