Datacenter · Compare
Immersion cooling vs direct-to-chip liquid cooling: which to spec for AI racks
Direct-to-chip is the mainstream default for high-density AI on conventional racks; immersion wins on extreme density and efficiency if the floor and fluid handling allow it.
Short answer
For most high-density AI deployments, pick direct-to-chip liquid cooling: it bolts a cold plate onto a fairly conventional server, reuses much of the existing air handling, and scales to the densities GPUs need. Pick immersion when you need the highest density and best efficiency and the building can take the tank weight and the fluid handling. The single biggest deciding factor is the building itself: immersion is a floor-loading and fluid-handling question first, so the weight the slab can carry and how much retrofit disruption you can absorb usually settle the choice before the racks do.
Immersion cooling vs Direct-to-chip liquid cooling: side by side
| Factor | Immersion cooling | Direct-to-chip liquid cooling |
|---|---|---|
| What gets wet | Whole server submerged in a dielectric fluid bath | Sealed cold plate on the CPU/GPU only; rest of server stays in air |
| Heat to liquid | Effectively all of the rack heat | Roughly 70 to 80 percent; air still handles memory, drives, power (the residual 20 to 30 percent) |
| Efficiency and density | Most efficient of the liquid approaches; reaches the highest densities | Efficient for high-density AI; lowers PUE via warm-water free cooling |
| Floor / structural load | Filled tank is heavy, roughly a couple thousand pounds and up; often needs slab reinforcement | Conventional rack load; facility water routing is the bigger constraint |
| Retrofit into an air hall | Hardest of the three; weight and fluid storage/handling | Easier; hybrid rack reuses existing air handling. Facility water is the deciding constraint |
| Serviceability | Lift a hot, dripping node out of oil; two-phase tank must be opened and vapor managed | Dripless dry-break quick-disconnects; hot-swap a node without draining the loop |
| Leak / spill risk | Standing fluid volume; spill containment sized to the tank | Coolant inches from live electronics; layered leak detection, isolation valves, sometimes negative-pressure loop |
| Fire / code | Flash-point driven; NFPA 75 flash-point floor, NFPA 30 Class IIIB, AHJ sign-off before fill | No open-bath fire class; leak strategy and coolant spec govern |
| Server prep | Replace thermal grease, pull fans and baffles, seal/swap drives, immersion-rated optics | Cold plate and interface material on the chip; server otherwise conventional |
Which should you pick?
Choose Immersion cooling when
- You need the highest achievable rack density and the best cooling efficiency, and can design around it
- The floor or slab can carry the filled-tank weight, or you can reinforce it before the tank is set
- You can store, handle, spill-contain, and eventually dispose of a standing fluid inventory
- You are building new or can absorb the largest retrofit disruption of the liquid options
Choose Direct-to-chip liquid cooling when
- You are cooling high-density AI on fairly conventional racks and want the mainstream, well-understood option
- You are retrofitting an air hall and can get facility water (or bridge with a liquid-to-air CDU)
- You need to hot-swap servers routinely without draining the loop
- You want to reuse existing air handling for the residual load rather than rework the whole hall
Bottom line
It depends on the building and the density you actually need. Direct-to-chip is the form most of the high-density AI market took because it fits a conventional rack, keeps the existing air handling for the residual load, and hot-swaps cleanly through dripless couplings; its hard questions are facility water, the leak path near live electronics, and wetted-material compatibility. Immersion is the efficiency and density leader, but it is a floor-loading and fluid-handling decision first: a full tank is several times the weight of the rack it replaces, the fluid becomes a maintenance item and waste stream the owner inherits, service means pulling dripping hardware out of oil, and two-phase carries an open PFAS supply-and-regulatory question. Neither is a religion; match the approach to the density, the floor, and how much fluid handling and disruption the site can take.
FAQ
What is the difference between immersion and direct-to-chip cooling?
Direct-to-chip mounts a sealed cold plate on the hot chips and leaves the rest of the server in air, so liquid carries roughly 70 to 80 percent of the rack heat and air handles the remainder. Immersion drops the whole server into a bath of dielectric fluid, so the fluid touches everything and carries essentially all the heat. Immersion is more efficient and reaches higher densities, but the tanks are heavy and it is the hardest of the two to retrofit.
Which is easier to retrofit into an existing air-cooled data center?
Direct-to-chip is easier. It bolts onto a fairly conventional server and rack and reuses much of the existing air handling for the residual load, and where there is no facility water a liquid-to-air CDU can bridge by rejecting heat into room air. Immersion is the hardest to retrofit because the tanks are heavy, often requiring floor reinforcement, and the fluid has to be stored and handled. For a retrofit, facility water and floor loading are the real questions before the racks.
Is immersion or direct-to-chip better for the newest high-power GPUs?
Both handle the densities air cannot, and the choice is a building and density decision, not a performance verdict. Direct-to-chip is the mainstream choice for high-density AI on conventional racks and takes the chip load while air covers the residual. Immersion reaches the most extreme densities with the best efficiency if the floor can take the weight and the site can handle the fluid. Confirm the chip vendor's reference design and the platform split, since limits are vendor-governed.