Direct to Chip vs Rear Door vs Immersion Cooling

Liquid cooling stopped being a research topic the day Nvidia shipped the GB200 NVL72, a rack that draws about 132 kW and exists only in a liquid cooled version. A few years earlier the average rack in a typical data center pulled around 7 kW, and the ceiling where containment air cooling runs out of physics, roughly 25 to 40 kW per rack, felt comfortably distant. That ceiling arrived much faster than the industry forecast, and teams planning new capacity are no longer deciding whether to use liquid. They are deciding which liquid approach fits their floor: direct to chip cold plates, rear door heat exchangers, or full immersion.

The three technologies solve the same problem, removing heat from a dense rack, but they put the water in different places. Direct to chip circulates coolant through metal plates bolted to the processors. Rear door exchangers mount a water cooled coil on the back of the cabinet and cool the exhaust air before it reaches the room. Immersion drops the entire server into a bath of dielectric fluid. Choosing between them is not about finding the best technology in the abstract; it is about your density target, your building, your hardware refresh cycle, and the operations model your team can actually run. This article compares how each one works, what density and efficiency each reaches, what they cost in practice, and when each is the right call, including the hybrid patterns most AI deployments end up running.

At a Glance: Liquid Cooling Options Compared

Three Ways to Put Water on the Problem

Behind all three approaches sits the same piece of hardware, and it is not the part people photograph at trade shows. It is a coolant distribution unit, or CDU, a rack mount or skid sized box of pumps, heat exchangers, filters, and valves. The CDU exists to keep two water loops apart. The facility loop is building scale plumbing carrying treated water to and from chillers, dry coolers, or cooling towers, and it is far too dirty and variable to run through a server. The technology loop is clean, tightly specified coolant at a controlled temperature and pressure that is good enough to touch electronics. A plate heat exchanger inside the CDU transfers heat from one loop to the other without ever mixing the two chemistries, and pumps, leak sensors, and flow control keep the technology side stable.

Once the CDU is in place, the three approaches differ only in where heat leaves the server and enters that loop.

The distinction matters for everything downstream. Direct to chip takes heat at the hottest spot, the most efficient point to intercept it, and leaves a residual air load that still has to leave the room. Rear door exchangers use the same water loop to cool air, which means they depend on the server fans and airflow that direct to chip deliberately reduces. Immersion surrounds the whole server with fluid, so there is no residual air load at all, but the server had to be built for the bath. One loop, three very different contracts with the hardware.

Direct to Chip: Coolant at the Source

Direct to chip is the 2026 default, around 45 percent of liquid cooling deployments and the technology Nvidia specifies for GB200 systems. A metal cold plate, usually copper with internal microchannels, mounts on each high power package: the GPUs, the CPUs, the switch ASICs, with thermal interface material conducting the die heat into the plate. Coolant, typically a treated water or propylene glycol and water mix, flows through the channels, warms by 10 to 20 degrees, and carries the heat out through quick disconnects into a rack manifold, then to the CDU, then to the facility loop.

Where direct to chip wins is capture rate and compatibility. Cold plates take roughly 70 to 90 percent of the heat out of the server before it ever becomes air, which is why a 100 kW rack only leaves 10 to 30 kW of residual load for memory, power supplies, and NICs to reject to the room. The servers stay standard form factor, standard racks, standard servicing, and hardware every major OEM ships today is cold plate ready, so procurement is a spec sheet conversation rather than a custom build. The loop runs warm, 30 to 45 degrees supply, which in most temperate climates lets a dry cooler reject the heat without a chiller for much of the year, and return temperatures of 55 to 65 degrees are high enough to make district heating or heat reuse realistic.

Direct to chip carries real operational weight. A coolant loop inside a live server needs leak detection, water quality monitoring, and disciplined service procedures, because a leak near electronics does damage in minutes and ionic or biological contamination in the loop is a maintenance problem that does not announce itself. The residual air still has to be removed from the room, which is exactly the gap rear door exchangers exist to close. And the loop is only as reliable as the CDU and its pumps, which is why N+1 pump redundancy and rack level flow monitoring are standard practice rather than optional.

Rear Door Heat Exchangers: Density Without Touching the Server

The rear door heat exchanger is the retrofit answer. The rack keeps its normal servers, fans, and airflow, but the rear door is replaced with a coil: exhaust air blows through the coil, heat moves into the water loop, and the air that reaches the aisle leaves noticeably cooler. There is no liquid inside the servers, which is the entire point. The risk profile, the hardware compatibility questions, and the service procedures all stay effectively air cooled.

Commercially available doors handle roughly 30 to 80 kW per rack depending on airflow and coil sizing, active units add fans for higher capacity, and the product is the lowest cost entry into water cooling, often around $5,000 per rack plus the water hookup. Because it does not require cooling the servers differently to cool the rack, a rear door is the natural first step for an existing hall with air containment that needs another 20 to 40 kW per rack without a rebuild. That simplicity of retrofit ranks as the most important adoption factor in industry surveys, and it is the reason rear doors are the most common first deployment in a brownfield.

The trade off is physics. The door still relies on the server fans to push air through the coil, so fan power, noise, and the air handling system stay in the design, and a liquid to air exchange is never as efficient as heat taken directly at the chip. Above roughly 80 kW the required airflow and coil sizes stop making sense, which is why rear doors occupy the middle band between air and direct to chip rather than replacing them. The water also has to reach the rack, so a brownfield team still needs space for piping and either a CDU or a chilled water connection near the row before the door swap is useful.

Immersion: The Server in a Bath

Immersion is the most total answer and the most operationally different one. Instead of bringing water to the server, the server goes into the water: a tank of dielectric fluid that does not conduct electricity, so the entire board is cooled by direct contact. There are two flavors. Single phase immersion keeps the fluid liquid, and pumps circulate it through a heat exchanger that transfers the heat to the facility loop. Two phase immersion uses a low boiling point fluid that boils at the chip, rises as vapor, condenses on a cooled surface, and drips back down, moving far more heat per unit of volume. Single phase is the deployed mainstream; two phase reaches the highest densities but brings engineered fluids, tighter regulation, and supply chain questions with it.

Immersion captures close to 100 percent of the heat, has no server fans, no residual air load, and no hot aisle, which is why it reaches the best efficiency of any approach or PUE around 1.02 to 1.08, and supports racks well past 150 kW with two phase systems pushing beyond 250 kW. It is also uniquely friendly to heat reuse, because the fluid leaves the tank warm enough to be genuinely useful to a district heating loop. For a fleet of identical high power nodes in a greenfield hall on a stable hardware roadmap, immersion is the most efficient way to get the heat out and often the lowest operating cost per kilowatt in the long run.

The cost is operational change. Servers need to be fanless, fluid rated, and on a compatibility list. Tanks take floor space and floor loading that a normal raised floor may not support. Every physical intervention means pulling hardware out of fluid into a different service model, and the fluid itself is a consumable that requires handling, filtration, sampling, and eventual replacement. Immersion therefore rewards predictability, a uniform fleet that changes rarely, and punishes the mixed, frequently churned environment most enterprise floors actually are.

Cost, Efficiency, and the Operating Load

The efficiency gap is recurring money. A facility with a fan and chiller plant at PUE 1.5 burns half again as much power as its IT load, while a warm water direct to chip hall at 1.1 in a temperate climate can let the chiller sit idle through most of the year. On a megawatt scale that difference is a seven figure annual electricity line. Immersion trims further by removing the server fans, and it delivers the most consistent thermal environment under full load, since every component is cooled by the same fluid rather than by whatever airflow a fan happens to deliver.

The capital picture is the mirror image. Rear door exchangers are the cheapest entry at roughly $5,000 a rack. Direct to chip sits mid range, priced per kilowatt of loop capacity and plumbing. Immersion carries the highest upfront bill, because tanks, engineered fluid, and compatible hardware add up before a single rack is powered. The honest comparison is not the price of the door or the tank but the total cost across the life of the hardware: the density you can actually deploy, the cooling energy you avoid, the throughput you gain from cooler silicon, the heat you can sell, and the operating headcount the technology demands. A direct to chip loop that lets a 100 kW rack run without throttling and without a chiller can pay for itself against air in a couple of years, while immersion on the wrong workload buys efficiency the floor never needed.

Matching the Option to Your Floor

The density target is the first filter, and it is usually decided by the hardware rather than by preference. A rack of ordinary web, storage, and general compute servers below 15 to 20 kW needs nothing more than containment and airflow discipline. Above about 30 to 40 kW the air path starts running out of value, and rear door exchangers become the cheapest way to buy headroom in an existing hall. Between 40 and 60 kW, which is exactly where today’s AI and GPU racks land, direct to chip stops being optional and becomes the spec sheet baseline. Immersion reserves itself for the top of the scale: uniformly dense fleets above 100 kW, plus specialized constraints like silence or harsh environments where a sealed tank is also a protection measure.

The second filter is your building. Brownfield work favors rear door and direct to chip in that order, because both reuse your racks, your aisle layout, and most of your room cooling, and both only need the water to reach the row. Immersion is close to a greenfield decision, since it reworks the floor, the service model, and the hardware it can hold. The third filter is fleet uniformity and service rhythm. If every rack is the same, runs the same load, and changes rarely, immersion rewards you. If the floor is mixed, churned by refresh cycles, and touched by hands every week, direct to chip with a rear door on the residual air is the pattern the market has standardized on: cold plates on the silicon, a door or small air plant for the leftover 10 to 30 percent, and one CDU loop serving the aisle. That hybrid is what most AI deployments ship with today.

The honest framing is that direct to chip is the 2026 default, rear door is the bridge for existing air cooled halls, and immersion is the specialist that wins at the extreme or on very uniform fleets. Most teams will not choose a single technology. They will choose a building where the approach can evolve from air to rear door to direct to chip as the density migrates through the life of the lease.

What a Liquid Hall Needs Besides Pipes

None of these options works on pipes alone, and the parts that are easy to under-budget decide whether a liquid deployment succeeds. Power comes first, because a rack that needs liquid cooling is a rack drawing tens of kilowatts, and 130 kW at the cabinet means high density power distribution feeding it. Power delivery and heat rejection are two halves of the same planning problem, not separate projects. Redundancy comes second, because the CDU and its pumps are now a single point of failure for the whole row. Operators plan N+1 pumps, extra fluid volume, and alerting that fires on flow and temperature before silicon protection clocks the GPUs down. Water quality and leak detection complete the safety loop, since a contaminated or leaking technology loop is a hardware reliability problem in a way air cooling never was.

The third under-budgeted item is testing and people. A liquid loop has failure modes nobody has practiced, so the commissioning window matters: flow balances, leak sweeps, pump failover drills, and a full cold start and hot pull sequence all deserve dry runs before the production load arrives. The same discipline applies to the team, because adding, moving, or servicing a liquid cooled server is a different skill from racking an air cooled one, quick disconnects have to seat correctly, bleed valves have to be closed, and the whole loop has to be watched through the transient. Budgets usually cover the extra racks and the pumps; the deployments that succeed also cover the drill time that proves the loop before it has to.

Whichever approach you land on, the decision looks the same from the operations side: pick the density target, verify the building and the loop can serve it, and lock in the flow and temperature budgets so the next refresh can move you up the ladder from air to rear door to direct to chip without starting the design from zero. The technology that fits today is the one your floor can actually run, and the plan that survives is the one built on verified assumptions.