Talk to sales → TCO Analysis →
Product line · Immersion Cooling

Whole-board cooling
for PCIe fleets.

MDCX Immersion Cooling is a line of three single-phase immersion container blocks for PCIe inference fleets, from 200 kW to 400 kW IT. Choose by pad length and whether the UPS boundary sits inside the box.

Which GPUs

PCIe inference boards. Not SXM training trays.

Heat sits across the whole card — GPU, VRM, NIC, DIMM. A dielectric bath picks it up without a cold plate per SKU, up to 50 kW per tank. SXM / NVLink training density belongs on Liquid Cooling.

Why Immersion Cooling

A bath is the lowest-OpEx path to this density.

It cools everything, not just the chip

GPUs, VRMs, NICs, DIMMs sit in dielectric fluid. No per-SKU cold plate, no fans, no hot spots — up to 50 kW per tank.

It runs warm, so it runs almost free

Warm oil to a facility loop. Where peak dry-bulb stays ≤ 24 °C, dry coolers reject all heat — no compressors, PUE about 1.05. Hotter sites add a hybrid chiller for peaks only and are not 1.05.

Redundancy is built into every tank

Each A32 tank carries dual CDUs in 2N. A pump or CDU can be swapped without draining the tank; a single failure does not stop the servers.

The blocks

Three immersion containers.

Compare

Same tanks. Different pad.

I400C45I400C40I200C20
IT400 kW400 kW200 kW
GPUup to 512 Pro 6000 Server Editionup to 512 Pro 6000 Server Editionup to 256 Pro 6000 Server Edition
Box45 ft40 ft20 ft
Tanks884× A32
UPSInside the boxOutside the boxOutside the box
Built forComplete pack with powerYou already have distributionSmallest inference bay

If a 45 ft box will not fit, I400C40 keeps the same eight tanks in 40 ft. I200C20 is four A32 tanks in 20 ft — an edge bay, not a half-size hall.

Next

Price the pad, then talk to sales.

TCO Analysis

Start with I400C45 — switch to I400C40 in the wizard if the UPS stays outside.

Open TCO →