By Md Omar Faruk
Every ChatGPT query, every AI image generation, every model training run produces heat
— a lot of it. Modern AI racks that used to draw around 15 kW now regularly need 40 to 100 kW, with generative AI inference being especially power-hungry. Some next-generation racks are reportedly heading toward 370 kW. The industry's default answer has been water — but water is a finite, increasingly contested resource, and the AI boom is straining it in places that can least afford it.
The Water Problem
Roughly 70–80% of data centers rely on evaporative cooling because of its energy efficiency, even though it consumes large volumes of water compared to traditional air conditioning. The scale is staggering at a regional level: Texas data centers alone used an estimated 25 billion gallons of water in 2025, a figure projected to climb to roughly 161 billion gallons annually by 2030 as AI infrastructure expands. That trajectory has pushed operators, researchers, and even hyperscalers to look hard for alternatives.
1.Air Cooling — Reconsidered
Air cooling was never gone, but it's being redesigned for the AI era specifically to cut water use. Google has reportedly shifted toward air cooling for new data centers as a way to reduce water consumption while AI drives both computing power and heat higher. The catch: air cooling can practically handle around 40 kW per rack, while high-end AI racks can reach 120–150 kW, so air alone works only for lower-density workloads — it's a partial fix, not a universal one.
2.Direct-to-Chip (Cold Plate) Liquid Cooling
Instead of chilling an entire room, this approach attaches liquid-carrying cold plates directly onto the hottest components — the CPU and GPU. Direct-to-chip liquid cooling can be up to 3,000 times more effective than air, while cutting energy use by 30–60% and reducing water consumption. Crucially, the "liquid" here is rarely plain water. Most systems use engineered water-glycol blends or dielectric fluids in closed loops, which sidesteps a lot of the scarcity problem entirely since the fluid is recirculated rather than evaporated away.
3.Single-Phase Immersion Cooling
Here, entire servers are submerged in a thermally conductive but electrically insulating dielectric fluid. The fluid never changes state — it stays liquid, absorbing heat directly from components with no air gap at all. This eliminates fans, dramatically raises achievable rack density, and uses zero water. Vendors like Submer have built commercial single-phase systems that are increasingly used as drop-in retrofits for existing facilities.
4.Two-Phase Immersion Cooling
This is the more aggressive version of the same idea. Two-phase fluids enable 10 to 100 times greater heat transfer capacity than single-phase alternatives by letting the dielectric fluid boil at a low temperature, carry heat away as vapor, then condense back into liquid in a closed loop. In 2026 deployment terms, two-phase immersion supports roughly 100–300+ kW per rack with potential power usage effectiveness (PUE) as low as 1.01–1.05, though it's essentially greenfield-only — it's very hard to retrofit into an existing building. Companies like LiquidStack have built modular tanks specifically for this, some rated for hundreds of kilowatts per rack.
5.Microfiluidic / In-Chip Cooling
The newest frontier skips the server entirely and cools inside the chip package itself. Microfluidic cooling integrates microchannels directly into the chip or package substrate, minimizing the thermal resistance between the chip and the coolant for markedly more efficient heat dissipation. It's still largely experimental, but it points toward a future where cooling is designed into silicon rather than bolted onto a rack afterward.
6.Rear-Door Heat Exchangers (RDHx)
A hybrid approach: a radiator-like liquid-cooled panel mounted on the back of a server rack catches hot exhaust air before it re-enters the room, letting data centers raise rack density without full immersion. It's commercially mature and often used as a stepping stone before operators commit to immersion.
7.AI Managing Its Own Cooling
There's a feedback-loop angle worth noting: operators are now using machine learning to optimize cooling in real time — adjusting airflow, fluid rates, and even workload placement based on predicted thermal load, squeezing efficiency out of whatever cooling architecture is already in place.
8.The Speculative Edge
Underwater and Ofifi-Grid Data Centers. A few experimental projects have gone further, submerging entire data center pods in the ocean (using the sea itself as a heat sink) or exploring off-planet computing. These remain niche, but they illustrate how far operators are willing to go to decouple compute growth from freshwater consumption.
The Bigger Picture
None of these technologies is a silver bullet on its own — liquid cooling is not replacing AI compute, it's becoming a prerequisite for it. What's emerging instead is a layered strategy: air and RDHx for lower-density racks, direct-to-chip for mainstream AI training clusters, and immersion cooling reserved for the densest, most demanding workloads. The common thread across nearly every alternative is the same: get away from evaporating fresh water, and do it without sacrificing performance.
cooling tiers by rack density · air, RDHx, direct-to-chip, immersion
For engineers and policymakers alike, cooling architecture is no longer a back-of-house utility decision — it has become part of AI procurement itself.
Top comments (0)