What Is Direct-to-Chip Liquid Cooling? How D2C Systems Work, Pros and Cons

AI silicon keeps getting hotter. Mainstream GPUs and accelerators already draw 700 to 1,000 W apiece, the next generation will draw more, and AI racks are now pushing past 100 kW. Bigger heatsinks and faster fans can't keep up. That's why direct-to-chip liquid cooling has become the default path for AI and HPC data centers.

Many explanations stop at one component: the cold plate. But what makes or breaks a deployment is the whole system. Where does the heat go after it leaves the chip? Which piece of equipment handles each step? And what do you gain, and give up, by switching to D2C? This guide answers those questions at the system level.

What Is Direct-to-Chip Liquid Cooling?

Direct-to-chip (D2C) liquid cooling is a closed-loop cooling method that pumps coolant through cold plates mounted directly on CPUs, GPUs, and other high-power chips. The coolant absorbs heat at the source and carries it through a rack manifold to a coolant distribution unit (CDU), which transfers the heat to facility water or, in liquid-to-air designs, to room air.

You'll also hear it called cold plate liquid cooling, and it's the most common form of direct liquid cooling (DLC). Unlike immersion cooling, D2C keeps the coolant sealed inside cold plates and tubing, so it never touches the electronics. Unlike air cooling, it pulls heat straight off the chip instead of dumping it into the room first.

Direct-to-Chip Liquid Cooling at a Glance

• What it cools: CPUs, GPUs, and other high-power chips, through cold plates. Lower-power components usually still rely on air.
• Where the heat goes: cold plate → rack manifold → CDU → facility water (or room air with a liquid-to-air CDU).
• Main types: single-phase (water or PG25, the mainstream choice) and two-phase (a boiling refrigerant or dielectric fluid).
• Biggest benefits: higher rack density, warm-water operation, and lower total energy use.
• Main trade-offs: high coolant flow, leak and corrosion management, and the added cost of CDUs and piping.

How Direct-to-Chip Liquid Cooling Works: The Heat Path Step by Step

Follow the heat, and a D2C system breaks down into six steps:

1. Chip to thermal interface material (TIM). Heat from the chip package crosses a thin layer of thermal interface material into the base of the cold plate.
2. TIM to cold plate. Microchannels or pin fins inside the cold plate transfer that heat into the coolant flowing through it.
3. Cold plate to hoses and quick disconnects. The warmed coolant leaves the server through flexible hoses and quick disconnects.
4. Quick disconnects to the rack manifold. Return lines from every server in the rack feed into a return manifold, while the supply manifold distributes cool fluid evenly back to each server.
5. Manifold to CDU. The return flow enters the CDU. Its pumps hold flow and differential pressure steady, its heat exchanger passes the heat to the facility side, and its controls set the supply temperature.
6. CDU to facility water to heat rejection. Facility water carries the heat to chillers, cooling towers, or dry coolers, which release it outdoors.

One design choice holds the whole thing together: the server side and the facility side are two separate loops that never share fluid. They exchange heat only inside the CDU's heat exchanger. The server-side loop is called the technology cooling system (TCS), and the facility-side loop is the facility water system (FWS). Keeping them apart means swings in facility water quality never reach the fine channels inside your cold plates. For how a CDU handles that isolation and control, see What Is a CDU? For the different ways facilities reject heat, see How Liquid Cooling Works.

Key Components of a Direct-to-Chip Liquid Cooling System

These parts have to be sized as a set. The cold plate's flow resistance sets how much differential pressure the manifold and CDU must deliver, and the quick disconnect's bore limits how much flow each server can get. Undersize any one link, and the whole loop underperforms.

Choosing a Direct-to-Chip Cooling Architecture for Your Data Center

The first architecture question isn't "Which cold plate?" It's "Does the facility have cooling water available?"

• No facility water: use a liquid-to-air (L2A) CDU. An L2A CDU pushes the heat from the cold plates through a coil and fans into the room air, where your existing air conditioning removes it. It needs no facility water piping, which makes it a good fit for piloting a few liquid-cooled racks in an air-cooled hall or retrofitting an older site in stages. The trade-offs are limited capacity and the fact that the heat still ends up in the room. For a related retrofit approach, see liquid-assisted air cooling.
• Facility water, rack-by-rack rollout: use an in-rack liquid-to-liquid CDU. The CDU sits inside the rack and serves only that rack. The loop is small, it needs little coolant, and you can add capacity one rack at a time.
• Facility water, deployment at scale: use an in-row liquid-to-liquid CDU. A single in-row CDU serves an entire row of racks, with capacities reaching the megawatt range. It's built for new high-density AI halls.

Single-Phase vs. Two-Phase D2C: The System-Level Difference

D2C comes in two flavors, depending on whether the coolant boils inside the cold plate. Our guides to direct-to-chip cold plates and two-phase cold plates cover how each plate works. Here, we focus on what each choice means for the rest of the system.

Single-phase systems are simpler to deploy and cost less. Single-phase cold plate cooling has an estimated 55% market share in 2026. Expect it to remain the default choice for the next several years. Two-phase systems move more heat per liter of flow, which makes them attractive as chip power keeps climbing. APALTEK's pump-driven two-phase cooling system is built on this approach.

Benefits of Direct-to-Chip Liquid Cooling

Higher rack power density. Liquid holds and conducts heat far better than air, so a cold plate can remove kilowatts of heat from a very small footprint. You can pack more GPUs into each rack and get by with fewer racks and less floor space.

Warm-water operation and more free-cooling hours. Air-cooled halls often depend on cold chilled water. Because a cold plate sits directly on the chip, D2C can run on much warmer supply water. At supply temperatures around 27 to 30°C, free cooling from outside air or cooling towers is viable for much of the year, with mechanical cooling needed mainly in the hottest months. How warm you can go depends on chip power and climate: the higher the chip power, the cooler the supply water it typically needs.

Heat reuse potential. Warmer return water carries more useful heat, which can be put to work for building heating or district heating networks.

Why PUE Understates the Gain

PUE is total facility power divided by IT power. D2C lets server fans slow down or disappear entirely, but that saving shows up in IT power. Since both the numerator and denominator shrink, PUE barely moves.

Here's a concrete example. In one modeling study, moving about 75% of the load to direct-to-chip cooling cut total data center power by 10.2%, yet PUE only dropped from 1.38 to 1.34. When you evaluate a D2C project, look at total energy use and your power bill, not PUE alone.

Drawbacks and Trade-offs of Direct-to-Chip Cooling

You'll still need some air cooling. If cold plates cover only the CPUs and GPUs, the heat from memory, drives, and power supplies still has to go out through fans and room air conditioning. For how cold plate coverage affects heat capture, see the coverage tiers in our direct-to-chip cold plate guide.

Flow demands add up. Single-phase D2C removes heat by warming the coolant, so higher chip power means more flow. Cooling a single 1,000 W chip takes roughly 1.5 L/min of coolant. Multiply that across a fully loaded GPU rack, and the requirements for pumps, pipe diameters, manifolds, and quick disconnect sizes climb quickly.

Leaks and corrosion are real risks. A water-based coolant leak can damage equipment, and mixing metals such as copper and aluminum in one loop invites galvanic corrosion. That's why cold plates and plumbing need reliable sealing, rigorous leak testing, and ongoing water quality management. Our D2C coolant guide covers this in detail.

Higher upfront cost and a steeper operations learning curve. D2C adds CDUs, manifolds, piping, and monitoring, and your operations team has to learn new skills such as coolant maintenance and servicing hardware on a live loop.

What Does Direct-to-Chip Cooling Cost?

D2C costs vary widely with rack count, chip power, existing facility conditions, and cold plate coverage, so there's no single cost-per-kW figure that holds up across projects. A more useful approach is to estimate by cost category:

• Server side: cold plates, internal tubing, and quick disconnects. The more chip types and the wider the coverage, the more this costs, and most of it is custom hardware.
• Rack and distribution: CDUs, rack manifolds, secondary piping, and leak detection and monitoring.
• Facility side: facility water piping, cooling towers or dry coolers, and any chiller modifications. New builds can plan this in one pass; in retrofits, it's usually the biggest unknown.
• Commissioning and operations: flushing, filling, pressure testing, and water quality checks, plus long-term coolant maintenance and spare parts.

Keep in mind that D2C raises rack density. The same compute needs fewer racks, less floor space, and less power distribution equipment, and those savings belong in your total cost of ownership.

When Does Your Deployment Need D2C?

A deployment typically needs direct-to-chip liquid cooling once chip power or rack density outgrows what air cooling can handle. Run through these questions for a quick read:

• Chip power: Is your CPU or GPU TDP close to, or beyond, what an air-cooled heatsink can handle?
• Rack power: Is your target rack power well above what your air-cooled hall was designed for?
• Facility water: Do you have chilled water or facility water available? If not, an L2A CDU is a practical starting point.
• New build or retrofit: A new build can be designed around D2C from day one. An existing site is usually better served by a phased rollout.
• Operations readiness: Is your team ready to manage coolant, leak checks, and live-loop servicing?

If your chip power and rack density have both outgrown air cooling and your facility water situation allows it, D2C is usually the most direct path forward. Still weighing D2C against immersion? See our direct-to-chip vs. immersion cooling comparison. Ready to pick cold plates? Start with our guides to choosing a server GPU cold plate and a server CPU cold plate.

Building a D2C Loop with APALTEK

APALTEK manufactures liquid cooling products and can supply the core hardware for both the server side and the rack side of a D2C loop:

• Custom cold plates. Our server and GPU cold plates use brazed copper construction (server cold plates are vacuum brazed), with options for nickel plating, integrated vapor chambers, and microchannel structures, all customized to your chip package and chassis. Browse cold plates.
• CDUs designed and built in-house. Our lineup runs from 5 kW and 150 kW in-rack CDUs to a 1.5 MW in-row CDU, covering everything from pilots to full-scale deployments. Browse AIDC thermal management products.
• Assembly-level delivery and leak testing. We can ship cold plates with quick disconnects pre-installed and leak-test the complete assembly. Every APALTEK product is helium leak-tested before it leaves the factory.
• Thermal testing and simulation. We run thermal tests on PG25 by default and can switch to water or deionized water to match your spec. Paid prototype projects usually include thermal simulation at no extra charge. To see how cold plate flow rate, temperature rise, and thermal resistance are calculated, read our liquid cold plate design calculations guide.

FAQ

How much heat can a direct-to-chip cooled rack handle?

That depends on cold plate coverage, flow rate, and CDU capacity, not on a hard limit in D2C itself. High-density AI racks already exceed 100 kW, and single CDUs now reach the megawatt range. In practice, the ceiling is set by how much flow you can route inside the rack and how much heat the facility can reject.

Does direct-to-chip liquid cooling lower PUE?

Yes, but usually by less than people expect. The fan power D2C saves counts as IT power, so it shrinks both the numerator and the denominator of PUE. Total energy use and your power bill are better measures of the real savings.

What is the difference between direct-to-chip cooling and a rear-door heat exchanger?

A rear-door heat exchanger (RDHx) mounts on the back of the rack and uses a coil to cool the hot exhaust air leaving the servers, so heat passes through the air before water carries it away. D2C absorbs heat with liquid right at the chip. Current rear-door products range from about 10 kW to 120 kW per door, and the two are often paired: cold plates take the chip heat, and the rear door captures what's left from memory, NICs, and power supplies.

Can direct-to-chip cooling be added to an existing data center?

Yes. Sites without facility water can start with an L2A CDU that releases cold plate heat into the room, where existing air conditioning removes it. Sites with chilled water can connect a liquid-to-liquid CDU. Before a retrofit, confirm rack space, floor loading, power capacity, and whether your servers come in liquid-cooled versions.

Next Step: Have APALTEK Evaluate Your D2C Loop

Send us three things:

1. Chip model and TDP
2. Target rack power
3. Facility supply water temperature (or let us know if you don't have facility water)

APALTEK's engineering team will put together a preliminary cold plate and CDU loop design and get back to you with a quote within one business day. Still pulling your requirements together? Start with our guide on how to prepare an RFQ for a custom cold plate.