Updated: July 14, 2026
In Brief
A single modern GPU rack can throw off more heat than an entire row of the servers it replaced.
Air cannot carry that away. Not fast enough, not at that density, not on a warm afternoon in Chennai or Pune. That one physical fact, heat, is quietly rewriting how data centers in India have to be built, and it is the part of the AI story that the GPU announcements skip.
Liquid cooling used to be a niche for supercomputers and enthusiasts. It has become the price of running serious AI. This is why, and what it means for your facility.
Because the density jumped, suddenly and by a lot.
A conventional enterprise rack ran at perhaps 5 to 10 kW. A modern AI rack does not. NVIDIA's GB200 NVL72 systems draw in the region of 120 to 130 kW per rack, and the next generation is expected to push past 240 kW. Even a modest GPU deployment lands in the 25 to 40 kW range, several times what the facility was designed for.
Heat is the direct consequence of power. Every one of those kilowatts becomes warmth that has to be removed, continuously, or the hardware throttles and then fails. The GPUs got dramatically more powerful. The heat they produce scaled with them. And the cooling most data centers were built around, air, did not.
Less than the new hardware produces, which is the whole problem.
Air cooling is a good, cheap, well-understood technology, and it works well up to a point. That point is roughly 30 to 35 kW per rack. Beyond it, moving enough air to carry the heat away becomes impractical: the fans get louder and hungrier, the hot aisles get hotter, and the physics simply stops cooperating. You cannot blow your way out of a 120 kW rack.
So the moment your AI density crosses that threshold, and production AI racks now sit far above it, air is no longer an option you are choosing. It is a limit you have hit. Liquid cooling is what lies on the other side.
Liquid carries heat far better than air, so you bring it closer to the source. Two approaches dominate.
Direct-to-chip runs cooling liquid through cold plates mounted directly on the hottest components, the GPUs and CPUs, taking the heat away at its source through a sealed loop. It has become the practical standard for the dense AI racks most enterprises are deploying now.
Immersion goes further: the servers are submerged in a non-conductive fluid that absorbs heat from every component at once. It suits the most extreme densities, and it is more of a rebuild of how you run the room.
Both do the same essential job the air never could: get the heat out of a rack that produces too much of it.
Direct-to-Chip vs Immersion vs Air
The table sets out where each fits.
| Method | Handles Up To (approx.) | How It Works | Cooling Efficiency (PUE) | Best For |
|---|---|---|---|---|
| Air | ~30–35 kW/rack | Fans and airflow | ~1.5–1.8 | General workloads; low-density AI |
| Direct-to-Chip | ~150 kW/rack | Cold plates on the chips, sealed loop | ~1.1–1.25 | Most enterprise AI deployments |
| Immersion | 100 kW/rack and beyond | Servers submerged in cooling fluid | ~1.02–1.1 | Extreme-density AI factories |
Note the last column. PUE, the ratio of total facility power to the power the IT actually uses, is where liquid cooling pays for itself. Air-cooled facilities typically run around 1.5 to 1.8; liquid brings that close to 1.1 or better. For a power-hungry AI estate, that difference is a large and permanent saving.
Because the ambient starting point is warmer, and the power is dearer.
Cooling a hot rack in a hot climate is harder work than cooling it in a temperate one, and much of India is hot for much of the year. That raises both the difficulty and the energy cost of removing the heat. Layer on Indian commercial electricity prices, which are far from the cheapest, and the efficiency of your cooling stops being a facilities footnote and becomes a line item the CFO notices. A poor PUE on a large AI estate, in India, is expensive twice over: once for the compute, and again for the power spent fighting the heat.
Liquid cooling helps on both counts. It removes heat the air cannot, and it does so far more efficiently, which in a hot, high-tariff market is a stronger argument here than almost anywhere.
No, and it is worth being honest about that.
If your AI footprint is small and your racks sit below the air-cooling threshold, air, perhaps with some enhancement, may still be fine. Plenty of useful AI runs at densities air can handle. The trigger is density, not the word "AI". Many enterprises will run a hybrid facility for years: air for the general estate and the lighter AI, liquid for the dense GPU racks that genuinely need it.
The point is not that everyone must rip out air cooling tomorrow. It is that the moment you deploy the dense racks that production AI increasingly requires, cooling becomes a decision you have to make deliberately, not a default you can assume.
Here is the mistake that strands the most capital: buying the GPUs first, then discovering the room cannot cool them.
Cooling is a facility decision with long lead times. Retrofitting liquid cooling into a data center never designed for it is possible but disruptive, and sometimes the honest answer is that a particular building cannot host dense AI at all, which is one reason colocation in a purpose-built, liquid-ready facility has become a common path. Either way, the sequence matters: confirm how you will power and cool the density before you order the accelerators, not after they are sitting on a loading dock with nowhere to run.
Get that order right, and liquid cooling is simply part of a well-designed AI factory. Get it wrong, and it is an expensive emergency.
The cooling question is where a lot of AI ambition quietly meets physics. Sizing the density, choosing air, direct-to-chip or immersion, and confirming the facility can carry it, before the GPUs are bought, is exactly the kind of design work that separates a running AI factory from a stranded one.
Proactive Data Systems designs and builds AI-ready data center infrastructure for Indian enterprises, including the power, cooling and facility design that dense GPU racks demand, on-premises and through purpose-built colocation. We are a Cisco Preferred Cloud and AI Partner, Dell Platinum Partner and NetApp Preferred Partner, with 35 years in enterprise IT, more than 1,500 organisations served, and a 24/7 service desk in India. To find out whether your facility can carry the AI you are planning, ask Proactive for an AI-readiness assessment.
K Saravanan Product Sales Specialist, Data Center, Proactive Data Systems
We'll get back to you shortly.
