Direct-to-Chip vs Rear-Door vs Immersion: Three Things Sold as "Liquid Cooling"
When a data centre says it offers liquid cooling, it could mean three quite different things. One cools the air behind your rack. One cools the chips themselves. One dunks the whole server in a bath. All three are sold under the same two words, and only one of them will run a rack-scale AI system.
This guide separates them. It builds on our general comparison, Liquid Cooling vs Air Cooling for AI Servers, and goes one level deeper.
The three approaches in pictures
| Compare | 🚪 Rear-door heat exchanger | 💧 Direct-to-chip | 🛁 Immersion |
|---|---|---|---|
| How it works | A liquid-cooled coil replaces the rack's rear door and cools the hot exhaust air | Cold plates sit on the CPUs and GPUs; coolant carries the heat away | Servers are submerged in a tank of non-conductive fluid |
| What touches the liquid | Only the air | The hottest chips | Everything |
| Servers still need fans? | Yes Fully | Some For the remaining parts | No |
| Change to your servers | 🟢 None |
🟡 Must be built for it |
🔴 Different hardware and handling |
| Rack density it suits | Moderate |
Very high |
Very high |
| Runs an NVL72-class rack? | No Not on its own | Yes This is what they use | No Not how these racks are built |
| Covered by Singapore's SS 726 | No Out of scope | Yes Single-phase | Check Two-phase is out of scope |
Sources: Data Center Knowledge on liquid cooling options, Vertiv on direct-to-chip cooling, Singapore Standards eShop, SS 726:2026. Bars are illustrative, not measured values.
Rear-door heat exchanger: better air cooling, honestly labelled
A rear-door heat exchanger swaps the back door of the rack for a radiator. Hot air leaving the servers passes through it and is cooled before it reaches the room. The servers themselves are unchanged and still rely entirely on their fans.
That makes it the easiest step up for an existing air-cooled hall. It also explains the controversy: at least one immersion cooling vendor argues it should not be called liquid cooling at all, only a better form of air cooling. They have a product to sell, but the technical point is fair. The liquid never goes near a chip.
How far it stretches depends on who you ask. Vendor and operator figures range from a few tens of kilowatts per rack to claims of up to 100 kW. Treat any single number with caution and ask for a reference site at your density.
Sources: GRC comparison, Verne on liquid cooling technologies
Direct-to-chip: what AI racks actually use
Here, metal cold plates are fixed to the processors and GPUs, and coolant flows through them. The heat goes into the liquid at the source, then to a coolant distribution unit, then out of the building.
This is the method behind NVL72-class racks. It is not total: HPE describes its GB300 NVL72 as sending about 90% of its heat to liquid and about 10% to air, because parts such as drives and power electronics are still air-cooled. So a direct-to-chip hall still needs some air handling. Rear doors are sometimes added to these racks precisely to mop up that remainder, which is why HPE lists rear-door options alongside its NVL72.
Immersion: the most radical, and the least common in AI racks
Immersion puts the servers in a tank of dielectric fluid, which does not conduct electricity. In single-phase systems the fluid stays liquid and is pumped through a heat exchanger. In two-phase systems it boils on the hot components, and the vapour condenses and drips back.
It removes fans altogether and handles very high densities. The costs are practical ones. Servers have to be suitable for immersion, maintenance means lifting dripping hardware out of a tank, and warranties need checking. The flagship NVIDIA rack-scale systems are engineered around cold plates and manifolds, not tanks, so immersion is not the route to an NVL72.
Which one do you need?
| Your situation | Look at |
|---|---|
| 🏢 Existing air-cooled hall, racks getting hotter, same servers | 🚪 Rear-door heat exchangers |
| 🧠 Liquid-cooled 8-GPU servers or an NVL72-class rack | 💧 Direct-to-chip, with a CDU |
| 🧹 Direct-to-chip racks whose leftover air heat overloads the room | 💧 + 🚪 Both together |
| 🏗️ New build, custom hardware, extreme density or harsh environment | 🛁 Immersion, after checking hardware support |
| 💬 A few inference servers at ordinary density | Air You may not need any of these. |
The question to ask your data centre
Do not ask "do you have liquid cooling?" The answer will be yes. Ask "which kind, at what rack density, and is it at my rack position today?" A hall with rear doors has not prepared for a direct-to-chip rack. A hall with direct-to-chip in one room has not necessarily got it in yours.
For the box that makes direct-to-chip work, see What Is a CDU? For the local standard, see SS 726:2026 in Plain English.
Related on SourceIT
SourceIT Approved
Researched and verified by the SourceIT team for Singapore buyers.




