Data Center / Hardware
Liquid Cooling TCO: Immersion vs. Air
Fans are dead. Why Immersion Cooling and Direct-Leads-to-Chip (DLC) are the only way to cool Nvidia Blackwell B200 clusters.
Liquid Cooling TCO: Immersion vs. Air

Hitting the Thermal Wall. For decades, we cooled computers by blowing cold air over hot metal heatsinks. But Moore's Law has a hot temper. Nvidia's Blackwell B200 GPUs draw over 1,000 Watts per chip. A standard rack of 72 of these GPUs consumes over 100kW of power in a few square feet.

You literally cannot cool this with air. To push enough air through the chassis to cool 100kW, the fans would require so much velocity that they would sound like jet engines (damaging hard drives with acoustic vibration) and consume 20% of the total rack power just to spin.

The Fluid Solution

The industry is rapidly shifting to Liquid Cooling. Water (or dielectric fluid) is 24 times more thermally conductive than air and holds 3,200 times more heat by volume.

There are two main flavors competing for the future of the AI data center:

1. Direct-to-Chip (DLC) / Cold Plate

This is the evolutionary step. Water blocks (cold plates) are mounted directly on top of the hottest components (GPU, CPU, Memory). Flexible tubes run to a rack manifold. This removes ~70-80% of the heat to fluid, with the rest distributed to the room air and handled by slow-moving fans.

  • Pro: Uses standard rack form factors. Easy for technicians to service.

  • Con: Still requires some air cooling. Risk of leaks near electronics (though glycol/water mix is standard).

2. Immersion Cooling

This is the revolutionary step. The "Full Dunk." The entire server board is vertically submerged in a tub (tank) of non-conductive dielectric fluid (like engineered mineral oil).

  • 1-Phase Immersion: The fluid circulates via pumps to a heat exchanger.

  • 2-Phase Immersion: The fluid boils off the chip surface (phase change to gas), rises as vapor, condenses on a cooling coil at the top of the tank, and rains back down. It is passive and extremely efficient.

The TCO Math (Total Cost of Ownership)

Immersion cooling requires high CapEx (buying tanks, expensive fluid at $100/gallon, reinforcing floors for weight). However, the OpEx savings are massive:

Metric

Air Cooling

Immersion Cooling

PUE (Efficiency)

1.4 - 1.6

1.02 - 1.05

Server Fan Power

15-20% of total load

0% (Removed)

Rack Density

15-30 kW/rack

100-250 kW/tank

Hardware Failure

Standard

Low (No dust, no vibration, stable temp)

Heat Reuse: The Circular Economy

The killer feature of liquid cooling is High-Grade Heat. Air capture produces lukewarm air (35°C) which is useless. Liquid capture produces hot water (60°C+).

This hot water can be piped directly into District Heating Systems. In Europe, data centers are becoming the furnaces for nearby residential blocks or swimming pools. Instead of venting waste heat into the atmosphere, you sell it as a commodity, turning a liability into a revenue stream.

The Verdict

If you are building an on-premise AI cluster in 2026, building raised floors for air cooling is building a relic. The TCO crossover point has passed: Liquid is cheaper than Air for high-density training. The data center of the future doesn't sound like a wind tunnel; it sounds like a bubbling aquarium.

See, Understand, Optimize -
All in One Place

Atler Pilot decodes your cloud spend story by bringing monitoring, automation, and intelligent insights together for faster and better cloud operations.