The Data Center Heat Crisis: How AI Is Pushing Cooling Technology to Its Limits

When people talk about the cost of AI, they usually mean compute time, talent, or capital. What rarely makes the conversation is something far more physical: heat. Every query you send to ChatGPT, every image you generate, every model that gets trained — all of it produces massive amounts of thermal energy that has to go somewhere. And right now, the industry is sprinting to figure out where.

Cooling already accounts for roughly 40% of a data center's total energy use. That number alone should stop you in your tracks. Nearly half the electricity powering these facilities isn't running AI — it's fighting the heat that AI creates. And as chips get more powerful, the problem is getting exponentially worse.

Traditional data centers were built around a comfortable assumption: that air cooling could handle the job. For decades, it could. Standard server racks ran at power densities of 4 to 6 kilowatts — manageable with fans, raised floors, and cold aisle containment.

Then AI happened.

Average rack density is expected to grow from 36 kW in 2023 to 50 kW by 2027, with denser racks requiring more cooling for advanced AI chips. And that's the average. At the high end, it gets extreme fast.

Nvidia's Blackwell GPUs are projected to consume up to 1 kilowatt of power each — a substantial 40% increase from the 700W thermal design power of the Hopper architecture. Pack 72 of those into a single rack, and you have a very serious problem. Nvidia's GB200 NVL72 rack generates 120 to 140 kilowatts of heat — roughly equivalent to running 40 home ovens simultaneously, continuously, in a space the size of a refrigerator. Air cooling simply cannot dissipate that thermal output.

Nvidia CEO Jensen Huang announced at the 2025 GTC that upcoming racks like the Rubin Ultra NVL576, expected in 2027, could consume power up to 600 kW. The trajectory is clear and it only goes one direction.

To understand why the industry is scrambling, it helps to know what the existing toolkit looks like.

Air cooling has been the default for 40 years. Cold air is pumped into server rows, absorbs heat from the equipment, and gets exhausted as hot air. It works, it's cheap to build, and every data center operator knows how to run it. The problem is physics: air is a terrible conductor of heat. Water and specialized coolants have roughly 3,000 times the heat-carrying capacity of air — meaning you need enormous volumes of moving air to cool what a small stream of liquid can handle trivially. At rack densities above roughly 20-30 kW, air cooling starts struggling. Above 50 kW, it's effectively useless.

Chilled water systems are what most large data centers already use to supplement air cooling. Water is circulated through cooling towers on the roof, chilled, and then used to cool the air inside the facility. This works, but it consumes enormous amounts of water — more on that in a moment.

Evaporative cooling is the most water-intensive approach. Water is evaporated to absorb heat, the same principle as sweating. It's energy-efficient but AI data centers consume 10 to 50 times more cooling water than traditional server farms, with some Google facilities averaging 550,000 gallons daily.

This is where the story gets genuinely alarming, and where the intersection with broader infrastructure risk becomes impossible to ignore.

Larger data centers can each consume up to 5 million gallons of water per day — usage equivalent to a town of 10,000 to 50,000 people. Scale that across the industry and the numbers become staggering. Data centers powering AI systems consumed approximately 17 billion gallons of water in 2023, with projections showing water usage will surge to 68 billion gallons by 2028 — a 300% increase in just five years.

Accelerated AI adoption alone could result in an additional 4.2 to 6.6 billion cubic metres of water withdrawal by 2027 — equivalent to four to six times the annual water withdrawal of Denmark.

And here's the part that compounds the problem: roughly two-thirds of the data centers built since 2022 have been located in water-stressed regions, according to a Bloomberg analysis, including hot, dry climates like Arizona. The industry is building its most water-hungry infrastructure in places that can least afford to spare it.

A study by the Houston Advanced Research Center found that data centers in Texas will use 49 billion gallons of water in 2025, and as much as 399 billion gallons by 2030 — a figure that approaches the annual capacity of Lake Mead, the largest reservoir in the United States.

An assessment of 9,055 data center facilities indicates that by the 2050s, nearly 45% may face high exposure to water stress. The infrastructure being built today to power AI could itself become stranded by the climate conditions AI's energy demands are helping to accelerate. The irony is difficult to miss.

The good news is that the industry knows the problem, and capital is flooding in to solve it.

Direct-to-chip cooling (D2C) is rapidly becoming the preferred solution for hyperscalers. Cold plates — small heat exchangers — are positioned in direct contact with CPUs and GPUs, transferring heat away from the most energy-intensive components directly into a liquid coolant, which is then pumped away from the rack. Direct-to-chip cooling can remove about 70 to 75% of the heat generated by equipment in a rack, with the remaining heat handled by supplemental air systems. Microsoft's recent sustainability report outlined their lean towards closed-loop direct-to-chip cooling and its sustainability benefits, and it appears to be the emerging preferred technology among the hyperscalers.

Immersion cooling takes the concept further. Entire servers are submerged in a bath of dielectric fluid — a non-conductive liquid that doesn't damage electronics. Two types exist: single-phase immersion, where the coolant absorbs heat and is circulated through heat exchangers, and two-phase immersion, where the coolant actually vaporizes upon absorbing heat and re-condenses, transferring heat away with extremely high efficiency. Immersion systems have proven especially effective in high-density AI model training and HPC clusters where rack densities exceed 80 to 100 kW. The trade-off is cost — immersion-cooled infrastructure runs roughly twice the capital cost of traditional air-cooled setups.

Embedded cooling is the frontier. Multiple companies — including Nvidia, Microsoft, and TSMC — are developing forms of embedded cooling that bring microchannels directly into the silicon itself, dissipating heat at the transistor level rather than at the chip surface. TSMC's Direct-to-Silicon Liquid Cooling is considered the closest to commercialization, which matters enormously given that TSMC produces the majority of AI accelerators currently in production.

Zero-water closed-loop systems are the environmental answer. Microsoft's upcoming data centers will use closed-loop, zero-water evaporation cooling — once filled at construction, the system recirculates coolant indefinitely, eliminating evaporative water entirely and reducing usage by 125 million liters per facility annually compared to evaporative designs.

One of the more counterintuitive aspects of the cooling transition is that liquid cooling, despite sounding more resource-intensive, is actually dramatically more efficient.

Nvidia's GB200 NVL72 rack delivers 40x higher revenue potential, 30x higher throughput, 25x more energy efficiency, and 300x more water efficiency than traditional air-cooled architectures — because the precision of direct-to-chip heat removal means the facility-level cooling infrastructure can operate at warmer water temperatures, reducing or eliminating the need for energy-intensive mechanical chillers.

Because of the higher efficiency of liquid-to-liquid heat transfer, data centers can operate effectively with warmer water temperatures, reducing or eliminating the need for mechanical chillers in a wide range of climates. That's a significant operational cost reduction as well as an environmental one.

For Brezco readers thinking about where this all leads from a market perspective — the cooling crisis is a real investment theme, not just an infrastructure story.

Tech companies are projected to spend $375 billion on data centers this year, increasing to $500 billion by 2026. A meaningful and growing slice of that spend goes directly to cooling infrastructure — the pipes, coolant distribution units, heat exchangers, dielectric fluids, and embedded chip technology that keeps everything running.

The companies positioned in this supply chain — thermal management specialists, liquid cooling hardware manufacturers, and the materials suppliers behind dielectric fluids — are benefiting from a demand curve that doesn't depend on which AI model wins or which hyperscaler dominates. They win as long as the compute keeps growing. And by every available forecast, the compute keeps growing.

The data center cooling problem is not a bug in the AI buildout. It's a feature of the physics. And physics doesn't negotiate.

Educational content only. Not financial advice. Brezco Analytics is an independent research and media platform.