Surviving 45°C: How Closed-Loop Liquid Cooling Could Reshape India's Next 2GW of AI Data Centers
NVIDIA's 45°C closed-loop liquid cooling architecture could reshape India's next generation of AI data centers. Explore the engineering, system design, and cloud infrastructure implications.
By Ayush
Introduction
AI has introduced a new bottleneck that software alone cannot solve.
For years, improvements in cloud infrastructure came from faster processors, better virtualization, and more efficient distributed systems. But the latest generation of AI hardware has shifted the challenge from computation to thermodynamics.
Modern AI accelerators consume enormous amounts of power, and almost every watt they consume becomes heat. As GPU clusters grow denser, keeping them cool has become one of the biggest engineering challenges in modern data centers.
Last week, NVIDIA unveiled a reference design for its Rubin-generation AI infrastructure that takes an unconventional approach: instead of cooling servers with cold water, it allows coolant to circulate at temperatures as high as 45°C (113°F) inside a completely sealed liquid loop. The result is a system that can dramatically reduce cooling energy while nearly eliminating on-site water consumption in suitable climates.
At first glance, this sounds like a sustainability story.
For cloud architects and infrastructure engineers, however, it represents something much bigger.
As India prepares for more than 2GW of new AI-focused data center capacity, high-temperature liquid cooling could fundamentally change how future facilities are designed—from rack layouts and power delivery to thermal management and infrastructure economics.
The Real Problem Isn't Compute—It's Heat
Every generation of AI hardware delivers more performance.
It also generates more heat.
Traditional enterprise servers typically operated comfortably within air-cooled environments. Fans pushed chilled air through racks, cooling CPUs and memory before the warm air was extracted from the room.
That model is reaching its limits.
Modern AI racks concentrate dozens of high-performance accelerators into a single enclosure. These systems generate far more heat than conventional enterprise infrastructure was designed to dissipate.
Adding more fans is no longer an effective solution.
Moving enough air through increasingly dense racks becomes inefficient, noisy, and energy-intensive.
Eventually, physics wins.
The limiting factor is no longer how much compute can fit inside a rack.
It's how quickly heat can be removed from it.
Why India's Climate Makes Cooling Even Harder
Cooling has always been a regional engineering problem.
In countries with mild climates, outside air can assist cooling systems for much of the year.
India presents a very different operating environment.
Cities such as Mumbai and Chennai experience high temperatures combined with significant humidity for extended periods.
Traditional evaporative cooling systems become less efficient under these conditions while consuming large quantities of water.
As India accelerates AI infrastructure investment, these environmental realities become impossible to ignore.
A data center designed around yesterday's cooling assumptions may struggle to support tomorrow's AI hardware efficiently.
This is why cooling architecture is becoming a strategic infrastructure decision rather than simply a facilities concern.
NVIDIA's 45°C Approach: Cooling With Warm Liquid
The surprising part of NVIDIA's Rubin reference architecture is not that it uses liquid cooling.
It's that the coolant itself is intentionally warm.
The closed-loop system circulates a mixture of water and propylene glycol at temperatures reaching 45°C, allowing heat to be captured directly from processors through cold plates rather than relying on chilled air. Every major component—including GPUs and networking equipment—is liquid cooled, eliminating server fans and greatly reducing dependence on cooling towers.
Because the coolant can safely operate at higher temperatures, the heat it carries can often be rejected through outdoor dry coolers instead of water-intensive evaporative cooling systems.
The loop remains sealed for the life of the facility.
No continuous water replenishment is required for chip cooling.
In favorable climates, NVIDIA estimates this approach can reduce facility cooling water consumption from roughly 2.6 million gallons per megawatt annually to nearly zero, while also reducing cooling-related electricity usage.
The breakthrough isn't colder cooling.
It's hotter cooling.
How Closed-Loop Liquid Cooling Changes Data Center Architecture
Liquid cooling is not simply a replacement for air conditioning.
It changes the architecture of the entire facility.
1. Rack Design Changes
Traditional racks are largely independent.
Liquid-cooled racks become part of an integrated thermal system.
Cold plates, liquid manifolds, and coolant plumbing become as important as compute hardware itself.
Rack placement must now consider hydraulic routing alongside networking and power.
2. Coolant Distribution Units Become Critical
One of the most important components is the Coolant Distribution Unit (CDU).
The CDU manages coolant flow, temperature, pressure, and heat exchange between IT equipment and the facility cooling loop.
Instead of thinking only about server placement, architects now design around coolant pathways.
The CDU becomes infrastructure—not an accessory.
3. Power Infrastructure Must Evolve
Higher-density AI racks also require significantly greater electrical capacity.
That means:
- Larger Power Distribution Units (PDUs)
- Stronger electrical feeds
- Higher-capacity UPS systems
- More sophisticated power monitoring
Power and cooling can no longer be designed independently.
Each directly influences the other.
4. Mechanical Engineering Becomes Part of Cloud Architecture
Historically, cloud discussions focused on software.
Containers.
Kubernetes.
Microservices.
Distributed databases.
AI changes that.
Mechanical engineering now directly affects compute density, operating costs, and infrastructure scalability.
Infrastructure teams must increasingly collaborate across software, electrical, and thermal engineering disciplines.
Why This Matters for India's 2GW Data Center Pipeline
India is investing heavily in digital infrastructure.
New AI-ready facilities are being planned across multiple regions to support cloud services, enterprise AI, and hyperscale deployments.
Designing these facilities around traditional air cooling would risk locking them into infrastructure that may become increasingly inefficient as AI hardware evolves.
Building with liquid cooling from the beginning offers several advantages:
- Higher rack densities
- Better energy efficiency
- Reduced dependence on cooling towers
- Lower operational water consumption
- Greater readiness for next-generation AI accelerators
Perhaps most importantly, new facilities avoid the complexity and cost of retrofitting later.
The decisions being made today will shape India's AI infrastructure for the next decade.
Engineering Trade-offs
Despite its advantages, closed-loop liquid cooling is not a universal solution.
Several challenges remain.
Higher Capital Costs
Liquid cooling infrastructure requires additional plumbing, pumps, heat exchangers, and monitoring systems.
Initial deployment costs remain higher than conventional air-cooled designs.
New Operational Skills
Operating a liquid-cooled facility requires expertise that many data center teams are only beginning to develop.
Leak detection, coolant chemistry, maintenance procedures, and hydraulic balancing introduce new operational disciplines.
Not Every Climate Is Identical
Although 45°C coolant enables dry-cooler operation for much of the year, extremely hot conditions may still require supplemental cooling in some regions. NVIDIA notes that chillers may still be needed during a small fraction of the year depending on local weather.
Water Use Doesn't Completely Disappear
Closed-loop systems significantly reduce facility-level water consumption.
However, AI infrastructure still has indirect water usage through electricity generation and semiconductor manufacturing.
Reducing cooling water is an important step, but it does not eliminate AI's broader environmental footprint.
Looking Ahead
The transition from air cooling to liquid cooling is beginning to resemble the earlier transition from spinning disks to SSDs.
At first, it appeared specialized.
Eventually, it became the default.
As AI workloads continue increasing in scale, thermal management will become a defining factor in infrastructure design.
Cloud providers will compete not only on compute performance, but also on cooling efficiency, rack density, and power utilization.
For India, this presents an opportunity.
With large-scale AI infrastructure investments already underway, the country has a chance to build facilities optimized for next-generation hardware instead of adapting legacy designs.
The next generation of cloud infrastructure may not be defined by faster processors alone.
It may be defined by how effectively those processors stay cool.
Key Takeaways
- AI infrastructure is increasingly limited by thermal management rather than compute availability.
- NVIDIA's Rubin reference design introduces a fully closed-loop liquid cooling architecture capable of operating with 45°C coolant, reducing reliance on chillers and evaporative cooling.
- Closed-loop cooling can significantly reduce on-site water consumption and improve energy efficiency in suitable deployments.
- Liquid cooling changes rack architecture, power distribution, and facility design—not just cooling equipment.
- India's expanding AI data center pipeline provides an opportunity to adopt liquid-cooled infrastructure from the outset rather than retrofitting existing facilities.
- Future cloud architects will need to think beyond software and understand how power, cooling, and mechanical systems influence AI infrastructure at scale.
Published by Ayush
Software engineer focused on cloud computing, distributed systems, and scalable backend architecture. Writing about what I build and learn along the way.