Author: Sumati Sahgal, VP, Data Centre & Secure Power
New AI data center builds in India face unique design constraints from day one. For hyperscalers, colocation providers, and neoclouds planning GPU-dense facilities, data centre liquid cooling has already become the default answer. The more consequential questions for a greenfield project are which liquid cooling architecture to design for and how closely that decision should be tied to India’s water constraints from the beginning.
Rack density has already been decided for you
NVIDIA’s GB300 NVL72 systems draw about 142 kW per rack, while Vera Rubin systems reach up to 246 kW per rack. Individual accelerator chips, including NVIDIA’s Blackwell GPUs, now have thermal design power above 1,000 watts each, more than triple what GPUs generated seven years ago.
For a facility planned today, this fuels out the option to cool with air first and add liquid cooling later if density demands it. A rack built for current-generation AI hardware, let alone Vera Rubin-class systems, needs liquid cooling from the first mechanical drawing. Retrofitting is still possible for existing facilities, but a new build that defers the decision is designed around outdated GPU generation rates.
Water makes India’s planning problem unique
Most liquid cooling arguments center on efficiency: lower PUE, reduced chiller load, and more headroom for compute. In India, that framing is incomplete. Water availability is the key constraint that distinguishes Indian data center planning from most other markets adopting liquid cooling at scale.
India’s data centers consumed approximately 150 billion liters of water in 2024–25. The Council on Energy, Environment and Water projects that this figure will more than double by 2030, driven directly by AI infrastructure growth. A single 100 MW hyperscale facility using conventional evaporative cooling can consume about 2 million liters of water per day, most of it lost to the atmosphere rather than recycled.
That consumption is disproportionately burdening the places that can least absorb it. Most of India’s data center capacity is concentrated in Mumbai, Chennai, Hyderabad, and Bengaluru, and S&P Global projects that 60–80% of India’s data centers will face high or extremely high water stress within the decade.
Closed-loop direct-to-chip liquid cooling is a game changer. Traditional evaporative cooling loses up to 85% of water to the atmosphere; a sealed, non-evaporative coolant loop circulating through cold plates virtually eliminates this loss. In a water-stressed city, this operating constraint determines whether the facility can secure water access at all as local governance tightens. India’s Ministry of Jal Shakti has already confirmed that the industry is adopting direct-to-chip and immersion cooling to reduce water demand, alongside rainwater harvesting and wastewater reuse.
Some operators are addressing water scarcity by sourcing water rather than relying solely on technology. Reliance Jio’s Jamnagar facility uses desalinated seawater for cooling rather than drawing from the municipal supply, an approach suited to coastal sites but not a substitute for closed-loop design inland where seawater access doesn’t exist.
Matching data center liquid cooling architecture to rack density
Three liquid cooling architectures cover the range of new AI builds that need to be planned for:
- Rear-door heat exchangers (RDHx)
These exchangers capture heat at the rack exhaust. They’re retrofit-friendly and suitable for moderate density increases, but they top out well below what current AI hardware requires. A new build targeting current- and next-generation AI rack densities will outgrow RDHx before it is finished. - Direct-to-chip liquid cooling
Supports much higher kW per rack by attaching cold plates directly to CPU and GPU surfaces. This is NVIDIA’s recommended approach for high-density AI deployments, and it should be the architecture most new AI-focused builds in India design around given current rack-density trajectories. - Immersion cooling
It handles even higher kW per rack by submerging full servers in dielectric fluid. It delivers uniform, largely silent cooling, but requires substantially more capital and specialized fluid handling.
Rack density targets should be set based on the GPU generation a facility plans for, rather than on current occupancy. Under-specifying the cooling architecture is a more expensive mistake to fix after construction than over-specifying it slightly at the outset.
CDU selection is a greenfield advantage, not an afterthought
Coolant Distribution Units (CDUs) are central to most liquid cooling architectures, and they come in three heat-rejection configurations with different infrastructure dependencies. Liquid-to-liquid CDUs are the most efficient but require chilled water infrastructure already in place. Liquid-to-air CDUs work without chilled water access and have a lower upfront cost. Liquid-to-refrigerant CDUs support modular, direct-expansion deployment where neither chilled water nor traditional air handling fits the site.
Greenfield builds have an advantage over retrofits: a new facility can incorporate chilled water infrastructure from the start and standardize on liquid-to-liquid CDUs, the highest-efficiency option. Retrofitted facilities without that infrastructure already in place are often locked into liquid-to-air or liquid-to-refrigerant configurations.
CDU physical topology is a related decision. In-rack CDUs suit phased or mixed deployments, in-row CDUs scale more efficiently for high-density pods, and gallery-scale (facility-level), floor-mounted CDUs fit large hyperscale campuses. Site layout and expected growth trajectory should drive this choice; a facility planning phased GPU rollouts over several years has different topology needs than one deploying a single hyperscale campus at once.
Climate introduces another variable specific to India. Chillerless operation (running liquid cooling without mechanical chiller backup) depends heavily on ambient conditions and design set points. In India’s hot, humid tropical and semi-arid zones, chillers typically remain necessary even with liquid cooling in place. CDUs and HDUs (Heat Dissipation Units, which reject heat to air where chilled water access doesn’t exist) need to be planned as a system that balances free cooling with mechanical backup.
Efficiency benchmarks are becoming a design target, not a compliance afterthought
India’s average data center PUE is currently around 1.5–1.6. The Ministry of Electronics and Information Technology (MeitY) has introduced a PUE requirement below 1.35 for its IndiaAI Mission GPU tender, covering across 10,000 GPUs of public-private cloud infrastructure. This benchmark has been resisted by parts of the industry on cost grounds.
There’s no binding national liquid cooling mandate yet, and the industry pushback against MeitY’s benchmark reflects real cost pressure, not manufactured resistance. The direction is consistent: the Telecom Regulatory Authority of India has recommended a certification framework for green data centers, and budget-linked incentives discussed for 2026 favor operators that build efficient, water-conscious infrastructure from the outset rather than retrofitting it later.
Direct-to-chip liquid cooling reduces cooling energy use by up to 60% compared with mechanical air cooling. It’s one of the more direct ways to improve PUE, as it reduces reliance on chillers and fans rather than optimizing around them. For a new build, designing to a sub-1.4 PUE target now costs less than retrofitting to a stricter benchmark once it becomes mandatory. Given the regulatory direction, treating today’s voluntary benchmarks as tomorrow’s floor is a reasonable planning assumption.
Schneider Electric opened a liquid cooling factory in Bengaluru in early 2026, built specifically to localize production of high-density cooling systems for domestic AI infrastructure. High-density cooling components no longer need to be imported and lead-timed against a global supply chain to reach a project on schedule, shortening the gap between design and deployment for new builds sited anywhere from Mumbai to Gujarat.
Schneider Electric’s Liquid Cooling Hub provides more detail on comprehensive direct-to-chip architectures, including CDU and HDU selection, as well as retrofit and greenfield planning guidance. For a closer look at how to select and design the right system for your facility, download the Liquid Cooling System Architecture Guide.
Add a comment