AI Cooling Technologies

The Numbers

Why liquid, why now.

Independent industry research on liquid immersion cooling in the AI era makes the case in plain arithmetic.

PUE 0

Single-phase immersion, vs ~1.58 for a traditional air-cooled data center

Compute per fully-populated 42U immersion rack — vs 5–7 kW air-cooled

−0

Cooling OPEX reduction reported with liquid immersion cooling

−0

Less water consumed than evaporative cooling setups

Immersion cooling can also reduce CAPEX build costs by more than 50%, cut footprint by as much as 60%, and repurpose excess heat for facility or district heating. Uniform temperatures eliminate hot spots; removing oxygen and dust extends hardware lifespan.

Immersion Cooling

Submerge the problem.

Servers and heat-producing components are submerged in a non-conductive dielectric fluid. As heat transfers to the fluid, a coolant pump cycles it through a heat exchanger and returns it to the tank at a lower temperature — continuously.

What an immersion system needs: a sealed tank or enclosure, dielectric fluid, a coolant distribution unit (CDU), and an external heat exchanger. Purpose-built immersion servers — with chassis engineered for bottom-to-top fluid flow — outperform retrofitted air-cooled hardware on thermal efficiency, maintenance, and reliability.

We work with a leading tank manufacturer specializing only in immersion cooling tanks. Standard sizes come in 12U, 25U, and 50U — and tanks can be custom-manufactured to your exact requirements.

Zero fan noiseMassive energy savingsExtreme densityHardware protectionSustainability

Immersion tank with dielectric fluid, CDU, and heat exchanger

Mythbusting Immersion

✕"Fluids wick up cables." Proper design and cable management prevent it.
✕"It needs high pressure." Immersion systems operate safely at standard pressure.
✕"Hard to service." Components are accessible without draining the tank.
✕"The fluid is flammable." Dielectric fluids used are non-flammable.
✕"It takes more space." Footprint can shrink by up to 60% versus air.
✕"Only for HPC labs." Adoption is growing across traditional data centers.

Direct Liquid Cooling

Cold plates on the hottest silicon.

Direct Liquid Cooling (DLC) — direct-to-chip (D2C) — runs a closed loop of coolant through cold plates mounted directly on CPUs and GPUs. The plates absorb heat, a CDU and heat exchanger dissipate it, and cooled liquid returns to close the loop. This can extract up to 80 kW per rack (60–70% of total heat) and cut energy costs by up to 45% — the remaining heat is still removed by air.

Direct-to-chip cold plate cooling loop illustration

⚡ Energy Savings

Eliminates power-hungry, high-speed fans and air conditioning units, drastically lowering energy consumption and operating cost.

▥ Density & Quiet

More processing power in a smaller footprint, free of airflow restrictions — and without the fan roar of a traditional data hall.

♻ Thermal Efficiency & Heat Reuse

Liquid transfers heat far faster than air, keeping components stable — and waste heat can be recovered and repurposed.

Single-Phase Cold Plate Technologies

▦ Microchannel Cold Plates

Sub-millimeter internal channels maximize surface area and drastically lower thermal resistance. Densely packed fins — typically copper for maximum conductivity, or aluminum where weight matters at scale — exponentially improve convective heat transfer.

◉ Microconvective (Microjet) Plates

Arrays of small fluid jets impinge coolant directly onto processor hotspots, breaking down thermal boundary layers. Outperforms standard microchannel designs by 20–40% in thermal resistance — for the hottest, most demanding silicon.

Cooling Racks

Right-sized rack cooling.

NordStar offers cooling rack solutions from Eaton and Rittal — both exceptional in this space. Eaton's in-rack units are a perfect fit for IT closets and smaller switching centers; Rittal's climate-control concepts scale from a single rack to entire data centers with safety and efficiency at the core.

Beyond that: portable cooling units, in-room precision air conditioners, in-row units between racks, rear-door heat exchangers, rack-top fans, and Eaton SmartRack modular data centers with 12 kW or 25 kW in-row precision cooling.

System Size	Typical Fit
Small · ≤ 5 kW	Computational systems or a small number of GPUs
Medium · 5–12 kW	1–2 racks (42U) up to a complete row of racks
Large · 12–25 kW	High-density rows; in-row precision and closed-loop enclosures
Beyond 25 kW	The liquid-cooling conversation — DLC or immersion

Feature	Air Cooling	Liquid (DLC)	Immersion
Methodology
Cooling medium
Max rack density
Heat transfer
Setup cost (CapEx)
Operating cost (OpEx)
Energy efficiency (PUE)
Space efficiency
Maintenance & risk
Hardware lifespan