Direct-to-Chip Cooling vs Immersion Cooling: Which Is Right for Your Application?

AI servers keep drawing more power, and in many facilities air cooling has simply run out of room. For most teams, the question is no longer whether to adopt liquid cooling. It’s which kind.

That decision carries more weight than it first appears. Once you commit to direct-to-chip (D2C) or immersion cooling, you’ve locked in your floor layout, your structural load requirements, your maintenance workflows, and often the direction of your next few server purchases. Changing course later is expensive.

So rather than declaring a winner, this guide breaks down how the two approaches actually differ, then lays out a practical framework for deciding which one fits your situation.

What’s the Difference Between Direct-to-Chip and Immersion Cooling?

Direct-to-chip (D2C) cooling pumps coolant through cold plates mounted on CPUs and GPUs, while the server stays in a standard rack and other components are usually still air-cooled. Immersion cooling submerges the entire server in dielectric fluid, so every component sheds heat into the liquid. D2C is easier to retrofit; immersion captures more of the heat load.

How Direct-to-Chip and Immersion Cooling Remove Heat

Direct-to-chip cooling. A cold plate sits on the chip package, and coolant flowing through the plate carries the heat out of the server. Hoses and quick disconnects tie multiple cold plates into a rack-level liquid cooling manifold, and a coolant distribution unit (CDU) regulates flow and temperature while exchanging heat with facility water. In single-phase D2C, the coolant stays liquid; in two-phase D2C, a dielectric working fluid boils inside the cold plate and absorbs heat as it vaporizes. For a closer look at the cold plate itself, see our guide to direct-to-chip cold plates.

Immersion cooling. The entire server board sits in a non-conductive dielectric fluid that directly contacts every component. In single-phase immersion, the fluid never boils. Pumps typically move it to a heat exchanger and back into the tank. In two-phase immersion, the fluid boils on hot component surfaces, and the vapor condenses on a coil above the bath before dripping back down.

Six Factors That Decide Between D2C and Immersion Cooling

1. Chip Heat Flux vs. Rack Density

“Density” means two different things here.

The first is chip heat flux: how much heat a given area of silicon has to shed. D2C has the edge on this one. Because the cold plate delivers liquid right on top of the heat source, D2C can often hold concentrated hotspots like GPUs at lower temperatures. Two-phase D2C goes further. Because boiling absorbs latent heat, it can move more heat at the same flow rate, which makes it a strong fit for chips with extreme heat flux.

The second is rack or floor density: how much compute you can fit into a given footprint. Here immersion leads. It doesn’t rely on air at all, so every component’s heat goes into the liquid and no residual air load holds the system back.

As for actual numbers, one industry estimate puts D2C at roughly 40–120 kW per rack and single-phase immersion at roughly 100–250 kW per rack. Other sources give noticeably different ranges, so treat these as ballpark figures rather than hard selection thresholds. And high density potential doesn’t mean immersion only makes sense for extreme densities. Immersion products today range from edge units drawing a few kilowatts to data center systems in the 100 kW-plus range.

2. Facility Constraints and Retrofit Feasibility

For most operating data centers, this factor matters more than raw performance. Uptime Institute’s 2025 cooling systems survey found that most operators consider ease of retrofit the most important factor in whether a direct liquid cooling system is viable.

This is where D2C shines. Servers stay in standard racks, and you only need to run coolant lines to the racks that need them. You can pilot a few racks, then scale up as confidence grows. Immersion works differently. Most tanks are horizontal, floor-mounted units with a very different footprint from upright racks, so the floor plan has to be redesigned. Once filled with dielectric fluid, they’re also heavy, and floor loading has to be calculated on the full, filled weight. In practice, immersion usually calls for a complete redesign of the data center, which makes it a better fit for new builds or planned large-scale renovations.

3. Coolant Selection and Supply Chain

• Single-phase D2C typically runs deionized water or a water-glycol mix. Water-based cold plates are the most common form of direct liquid cooling, with a mature supply chain and low fluid cost.
• Single-phase immersion uses dielectric fluids, most commonly hydrocarbons (such as mineral or synthetic oils) or esters.
• Two-phase systems, whether immersion or cold plate, use low-boiling dielectric working fluids.

Volume differs sharply too. The same estimate cited above puts the coolant volume of single-phase D2C at least an order of magnitude below that of an equivalent immersion deployment, and more fluid means higher fill, top-up, and replacement costs.

Two-phase systems call for one more check: fluid supply and regulatory status. 3M committed to exiting PFAS (per- and polyfluoroalkyl substances) manufacturing by 2025, and its Novec fluids were used in both two-phase immersion and two-phase cold plate cooling. If you’re evaluating a two-phase system, confirm the long-term supply source for its working fluid and the regulations that apply in your region.

4. Serviceability and Day-to-Day Operations

D2C largely preserves existing maintenance practices. Servers still slide out of the rack, and swapping a drive or a DIMM works as it always has; the main addition is inspecting coolant lines and quick disconnects.

Immersion requires a new way of working. Servers come out of the tank vertically, often with the help of lifting equipment. The dielectric fluid needs regular filtration and monitoring of its properties, and technicians need training on the new procedures. None of this is a deal-breaker, but it should be built into staffing and cost plans from the start.

5. Hardware Compatibility and Warranty

D2C cold plates have to match the chip socket precisely. Whenever a new CPU or GPU generation changes the socket or package, the cold plate base and retention hardware usually need a redesign. That’s why cold plate design has to account for the socket from the start.

Immersion is far less sensitive to chip package changes, since a new chip generation doesn’t require a new cold plate. It has a different hurdle, though: every material on the server has to survive long-term submersion in dielectric fluid, including cable jackets, seals, and label adhesives. To address this, the Open Compute Project (OCP) published immersion cooling requirements that specify suitable dielectric fluids and provide wetted-material compatibility guidelines. You’ll also want to confirm that your server vendor supports immersion use. Otherwise, your warranty may be at risk.

6. Energy Efficiency and Noise

Both approaches can run at relatively high coolant supply temperatures, which lets you lean more on free cooling and less on compressor-based chillers. The real differences come down to residual air load and pumping power.

• If D2C covers only the main chips, the remaining components still depend on fans and room air conditioning, which pushes overall energy use up. Two-phase D2C transfers heat through latent heat, so it runs at a lower flow rate and needs less pump power than single-phase D2C.
• Immersion eliminates server fans and leaves almost no residual air load. As a result, it can reach a lower power usage effectiveness (PUE) and runs the quietest. For typical PUE ranges across liquid cooling methods, see our overview of liquid cooling types.

Keep in mind that PUE depends heavily on local climate, heat rejection method, and system design, so published figures from different projects aren’t directly comparable.

Direct-to-Chip vs. Immersion Cooling: Which to Choose by Scenario

You Don’t Have to Choose Just One: Hybrid Deployments

D2C and immersion aren’t mutually exclusive. D2C systems that cover only the main chips are often paired with rear-door heat exchangers or liquid-assisted air cooling to handle the remaining air-side load. A single campus can also mix approaches by zone, using D2C for general-purpose compute and immersion for high-density dedicated zones or edge nodes.

Often the real decision isn’t which technology to use, but which workloads go where.

From Plan to Deployment: What Your Cooling Supplier Needs to Deliver

Once you’ve chosen a direction, a smooth deployment depends largely on whether every component is engineered to work as one system. APALTEK manufactures both D2C and immersion products. Here’s what each approach requires.

Single-phase D2C loop. You’ll need cold plates customized to your chip platform, rack-level manifolds, and CDUs, with the cold plates’ pressure drop and flow requirements matched to what the manifolds and CDUs can deliver. APALTEK offers a 2.5 MW-class In-Row CDU for row-based rack deployments. Pressure-hold and leak test items are defined with each customer to match their platform and qualification requirements.

Two-phase D2C system. A key challenge in two-phase systems is uneven flow distribution across multiple branches, which can leave some cold plates starved of fluid. APALTEK’s pump-driven two-phase cooling system uses wedge-shaped distributed jet cold plates to address this, and draws on our in-house two-phase flow research database for thermal calculations and custom rack-level designs.

Single-phase immersion system. The tank and CDU need to be sized to the same heat load. APALTEK’s IDC Immersion Cooling System, for example, pairs a floor-mounted tank with 46U of capacity and a separate CDU, supporting 15–60 kW of IT power per tank with 60 kW of CDU heat exchange capacity. For edge sites, our Edge Immersion Cooling System integrates edge servers, free cooling, immersion liquid cooling, and an optional UPS into one compact unit with 4U of server capacity. It supports 2.5 kW or more of IT power, carries an IP55 rating, starts working as soon as it’s connected to power and network, and can be stacked or deployed in rows.

Because APALTEK develops and manufactures all three, we don’t have to force your project into a single product line. We start by working out which approach fits, then move into design and prototyping.

Direct-to-Chip vs. Immersion Cooling FAQ

Is immersion cooling always more energy-efficient than direct-to-chip? Not always. Immersion eliminates server fans and leaves almost no residual air load, so in principle it can reach a lower PUE. But real-world efficiency also depends on heat rejection conditions, coolant supply temperature, and system design. When a D2C system has broad cold plate coverage and runs at a high supply temperature, the efficiency gap narrows considerably.

Can an existing data center be retrofitted for immersion cooling? Yes, but it’s difficult. Immersion tanks are typically horizontal, floor-mounted units with a different footprint from racks, and they’re heavy once filled. That means redesigning the floor layout and verifying floor loading. For this reason, immersion is more common in new builds, dedicated zones, or planned large-scale renovations.

Does direct-to-chip cooling still need air cooling? It depends on coverage. If cold plates cover only the CPUs and GPUs, memory, storage, and power supplies still need fans and room air conditioning. As cold plate coverage extends to more components, the need for air cooling drops sharply.

Is immersion cooling only for ultra-high-density racks? No. Immersion has high density potential, but products span a wide range, from high-density systems in the 100 kW-plus range to all-in-one edge units drawing a few kilowatts. At moderate densities, immersion’s value comes mainly from fan-free operation, low noise, and complete heat capture.

What’s the difference between two-phase direct-to-chip and two-phase immersion cooling? Both absorb heat by boiling a working fluid. The difference is where the boiling happens. In two-phase D2C, the fluid boils inside sealed cold plates, the servers stay in standard racks, and only the cold plate loop contains working fluid. In two-phase immersion, servers sit in a bath of low-boiling fluid that boils directly on component surfaces, which requires a sealed tank and a condenser. Two-phase D2C uses far less fluid, while two-phase immersion covers more of the server.

The Bottom Line: Start with Your Constraints

Neither D2C nor immersion is universally better. D2C integrates more easily into existing facilities and excels with high heat flux chips. Immersion offers complete heat capture, fan-free operation, and low noise, which makes it well suited to facilities designed from the ground up and to edge sites without a controlled data center environment. The right choice comes down to your rack power targets, facility constraints, operations capability, and heat rejection options.

If you’re planning a liquid cooling project, send your target rack power, server platform, whether it’s a new build or a retrofit, facility water conditions, and operations team setup to APALTEK’s engineering team. We’ll assess whether single-phase D2C, two-phase D2C, immersion, or a hybrid approach fits best, then recommend a specific configuration.