How to Retrofit an Air-Cooled Data Center for Liquid Cooling: Options, Steps and Pitfalls

Last updated: November 2026

Most data centers running today were built for air-cooled racks drawing 5–15 kW. A fully loaded AI training rack can pull 60–100 kW or more. Building a new liquid-cooled facility means waiting on power, land and permits, often for years. A liquid cooling retrofit, which adds rear-door heat exchangers or direct-to-chip cold plates and CDUs to an existing air-cooled hall, is usually the faster way to add high-density capacity.

But a retrofit is not a matter of swapping air handlers for CDUs. Three things decide how far an existing building can go: how much weight the floor can carry, how much facility water is available, and how much electrical headroom is left. This guide walks through the work in the order it actually happens: auditing the site, choosing a retrofit path, phasing the work around live operations, and avoiding the mistakes that derail most projects.

How Do You Retrofit an Air-Cooled Data Center for Liquid Cooling?

To retrofit an air-cooled data center for liquid cooling, follow five steps:

1. Audit floor loading, power and facility water.
2. Choose RDHx or direct-to-chip cooling with a liquid-to-air or liquid-to-liquid CDU.
3. Build a liquid-ready zone and route the loop.
4. Pilot one rack or row, then convert in waves.
5. Rebalance air cooling for the residual heat.

Key Takeaways

• The site sets the limit, not the hardware. Floor loading, available facility water and electrical headroom decide how far a liquid cooling retrofit can go, so audit them before choosing CDUs.
• Your water decides the path. Rear doors need facility water, liquid-to-air CDUs need none but are capped by room air cooling, and liquid-to-liquid CDUs need spare chilled water flow.
• Phase it like a live-site project. Pilot one rack or row, convert in waves with a rollback plan, and plan to run air and liquid side by side long after the retrofit.

Why Air-Cooled Halls Run Out of Headroom

Air runs out of room long before AI racks reach their full power. ASHRAE TC 9.9 points out that a 40–50 kW rack can need around 5,000 cfm of airflow, while even best-in-class floor tiles deliver about 1,900 cfm. Meanwhile, more operators are reporting peak rack densities of 30 kW or more, and AI clusters push well past that.

Once racks outgrow what the air system can deliver, you either cap how many servers each rack holds, stranding space and power, or you move part of the heat into liquid. For guidance on which rack densities call for liquid cooling, see our data center liquid cooling overview.

Step 1: Audit the Site Before You Pick Hardware

The most common way a retrofit goes wrong: the team picks CDUs and cold plates first, then discovers the floor can't carry them, the pipes have no route, or the power isn't there. Start with a combined structural and MEP assessment, and finish it before you fix any CDU locations.

When the audit is done, you should be able to answer three questions: Which area can carry the load, take the piping and supply the power? Is there usable facility water there? And how many racks will you convert first?

Step 2: Choose a Retrofit Path Based on the Water You Have

The right path depends mostly on the water your building already has. The four options below run from least to most change to the facility.

In short: if you have chilled water and moderate densities, start with rear doors. If you have no facility water, use a liquid-to-air CDU for a pilot. If you can tap chilled water with spare flow, use liquid-to-liquid CDUs. If the chilled water plant is already maxed out, build a dedicated liquid zone.

Path A: Rear-Door Heat Exchangers for Moderate Densities with Chilled Water

A rear-door heat exchanger replaces the back of the rack with a water coil, so hot server exhaust is cooled as it passes through. Servers stay as they are, which makes this the gentlest retrofit option.

The catch is capacity. A rear door on its own typically handles on the order of 10–30 kW per rack, well short of a dense AI rack. That's why doors are increasingly paired with direct-to-chip cooling: an Open Compute Project white paper describes an 80 kW rack made workable by combining a 30 kW door heat exchanger with a 50 kW in-rack CDU. Use rear doors alone for moderate-density racks, and as a supplement to cold plates above that.

Path B: Liquid-to-Air CDUs When There's No Facility Water

If the building has no usable facility water, you can still fit servers with cold plates and feed them from a liquid-to-air CDU, which dumps the heat into room air through its own heat exchanger. This approach is often called liquid-assisted air cooling (LAAC). ASHRAE TC 9.9 describes it as a bridge: liquid-to-air units can later be swapped for liquid-to-liquid CDUs once facility water reaches the rack.

It comes with one hard limit: the heat never leaves the room. It just moves from the chip to the air more efficiently. Your spare air cooling capacity sets the ceiling, so this path suits pilots or a handful of high-density racks, not full rows of AI servers.

Path C: Liquid-to-Liquid CDUs When You Can Tap Chilled Water

This is the most common path for high-density AI racks. Servers get cold plates, and a CDU uses a plate heat exchanger to separate the primary side (existing chilled water or tower water) from the secondary side (the technology cooling loop that runs through the cold plates). The CDU controls secondary-side flow, temperature, pressure and filtration, so facility water quality never reaches the cold plates.

CDUs come in two common forms. An in-rack CDU sits inside a rack and suits a small number of racks; an in-row CDU sits in the row and can serve an entire row. See What Is a CDU? for how the two compare. On this path, the audit should confirm how much flow and pressure head the existing chilled water system can spare, and whether you can tap in without shutting it down.

Path D: A Dedicated Liquid Zone with a New Facility Loop

If the existing chilled water plant is already at capacity, or your target density is very high, you can carve out a dedicated liquid cooling zone with its own facility-side loop, such as dry coolers that reject heat straight to the outdoors. This path costs the most, but it doesn't compete with existing air cooling for capacity, and it leaves the most room for next-generation chips. For how dry coolers and other options affect water consumption, see our analysis of liquid-cooled data center water use.

What About Immersion Cooling?

Converting a whole hall to immersion means replacing racks with tanks and changing server form factors, which makes it the most disruptive option. A more practical move is to handle a small, dense workload with a self-contained edge immersion unit. For a full comparison, see direct-to-chip cooling vs. immersion cooling.

Step 3: Build a Liquid-Ready Zone and Route the Loop

Once you've chosen a path, set aside a liquid-ready zone in the hall, usually one hot aisle or one full row. Grouping liquid-cooled racks together, instead of scattering them, shortens pipe runs, simplifies leak detection and makes later expansion easier.

Overhead or underfloor piping? Underfloor runs keep the ceiling clear, but they eat into the supply-air plenum, and every service visit means lifting floor tiles. Overhead runs are easier to maintain but demand more clearance and stronger supports. In older buildings, new pipe has to compete for space with electrical conduit, fire suppression lines and structural columns, and off-the-shelf supports often won't fit. Low-profile custom supports are common, along with vibration isolation near pumps.

Manifolds and quick disconnects. The rack manifold distributes secondary coolant to each server, and quick disconnects let you plug and unplug servers without draining the loop. See our liquid cooling quick disconnect guide for sizing and specification tips.

Leak detection. Manifolds, quick disconnects and CDU connections are the likeliest failure points, so sensors need to cover all of them, with drip trays and alarms tied into the BMS. For the full detection and commissioning sequence, see our data center liquid cooling design guide.

From the manufacturing side, there's one simple way to cut on-site risk: have cold plate and manifold assemblies leave the factory with quick disconnects already installed and the whole assembly leak-tested. In a live hall, the field crew then just connects, fills and verifies, instead of making up joints on site and chasing leaks.

Step 4: Phase the Work Around Live Operations

A retrofit happens inside a working facility, where one leak, power interruption or dust event can hit customers. Phase the work:

1. Prepare the infrastructure first. Install CDUs, main piping and power upgrades without touching the racks that are already running.
2. Pilot one rack or one row. Use development or test workloads if you can, and verify flow, temperatures, alarms and failover.
3. Convert in waves. Move row by row, with a rollback plan for each wave so you can return to air cooling quickly if something goes wrong.
4. Control construction risk. Get hot-work permits, isolate dust and vibration, and flush, fill and pressure-test piping offline wherever possible before connecting it to racks.

Coolant selection, water quality and filtration all affect long-term cold plate reliability. Our direct-to-chip coolant guide covers them in detail.

Step 5: Run Air and Liquid Side by Side

Even with direct-to-chip cooling, memory, power supplies and drives still shed heat into the air, so the hall will run as a hybrid of air and liquid for a long time. For a rough idea of how that residual load splits, see the liquid-air split section of our data center liquid cooling overview.

After the retrofit, do two things. First, rebalance your CRAH/CRAC units: liquid-cooled racks put less heat into the air, so reduce airflow and revisit supply-air setpoints to match the lower air-side load, or you'll waste fan energy and overcool the room. Second, be clear about who sets the water temperature. The CDU's secondary side controls the coolant supply temperature to the cold plates. If the CDU's primary side shares chilled water with the air handlers, that chilled water still has to meet the air handlers' needs. Raising water temperature for efficiency only becomes an option once the liquid cooling loop is separated from the air cooling loop on the facility side, as in Path D.

7 Common Liquid Cooling Retrofit Pitfalls

1. Running the structural review after CDU locations are fixed. The result is slab reinforcement or a relocation, and a longer schedule.
2. Overestimating the existing chilled water. Checking cooling capacity alone, without verifying available flow, pressure head and supply temperature.
3. Finding the power shortfall too late. CDU pumps and higher rack power both need electrical support.
4. Leaving gaps in leak detection. Putting sensors only at the CDU and missing manifolds and quick disconnects.
5. Reading OEM warranty terms too late. Cold plates, coolant or supply temperatures that fall outside what the server vendor approves.
6. Mixing metals in the loop. Combining copper and aluminum in one loop raises the risk of galvanic corrosion, especially without a managed coolant and inhibitor program.
7. Designing only for today's chips. Leaving no flow or capacity margin for the next, hotter GPU generation.

Retrofit or Build New?

Retrofitting is usually cheaper. One market study estimates that a liquid cooling retrofit costs around USD 2 million per MW, compared with more than USD 11 million per MW for new liquid-cooled capacity. That gap doesn't include hidden costs, though: downtime risk, workload migration, redesign and contractor premiums.

As a rule of thumb, retrofit makes the most sense when power, land or permits are constrained and high-density demand is firm. If power is readily available, or the floor, ceiling height or electrical system simply can't support your target density, a new build or partial new build (Path D) often comes out ahead. Many projects end up in between, with a dedicated liquid cooling zone inside the existing building.

How APALTEK Supports Liquid Cooling Retrofits

APALTEK designs and manufactures liquid cooling products in its own factory, including the two components that matter most in a retrofit: CDUs and cold plates.

• CDUs: Single-phase in-rack CDUs (5 kW and 150 kW) for pilots and small rack counts, and in-row CDUs (1.5 MW and 2.5 MW) for full liquid-cooled rows.
• Custom cold plates: CPU and GPU cold plates built around your server and chip layout, using brazed copper, skived fin, pin fin, friction stir welded and tubed constructions. Browse sample designs on our cold plate page, or see our server GPU cold plate selection guide.
• Factory-installed quick disconnects, whole-assembly leak testing: Cold plate and manifold assemblies can ship with your specified quick disconnect brand pre-installed and the full assembly leak-tested, cutting down on joint work inside a live hall.
• AI data center projects: See how these components fit together on our AI data center solutions page.
• Testing and validation: Every product is helium leak-tested, and other tests are defined around your requirements. See our testing and validation capabilities. Paid prototype projects usually include thermal simulation.

FAQ

Can you add liquid cooling to a data center without facility water?

Yes. Fit servers with cold plates and use a liquid-to-air CDU to reject the heat into room air, and you don't need facility water at all. The room's air conditioning still has to remove that heat, though, so this approach suits a limited number of racks, and its scale depends on spare air cooling capacity.

Do you have to replace servers or racks to retrofit liquid cooling?

Not necessarily. Rear-door heat exchangers replace only the rack's back door and leave servers untouched. Direct-to-chip cooling needs cold plates on the CPUs and GPUs, either on servers that ship liquid-ready or on custom cold plates built for your existing servers. You can usually keep your racks but will need to add manifolds.

How long does a data center liquid cooling retrofit take?

It depends more on what the audit finds than on equipment lead times. Three factors drive the schedule: whether the floor needs reinforcement, whether you can tap chilled water without a shutdown, and whether the electrical system needs an upgrade. Start with a single-rack or single-row pilot, then roll out in waves once it's proven.

Can a raised floor support CDUs and liquid-cooled racks?

That depends on the rated loads of the raised floor and the slab beneath it. CDUs, water-filled piping and high-density racks add up quickly and can exceed what an older raised floor is rated for. Run a structural review before you fix CDU locations, and if needed, set CDUs directly on the slab or reinforce the floor locally.

Should you retrofit or build new for AI racks?

Retrofit is faster and cheaper when power and space are tight, demand is firm, and the existing building's floor loading, ceiling height and power distribution can support your target density. If those hard limits can't be met, a new build or a dedicated liquid cooling zone is the better fit.

Start with the Site, Then Choose the Hardware

A liquid cooling retrofit succeeds or fails on the early assessment, not on the equipment brand. Send APALTEK's engineering team your target rack power, existing chilled water conditions (supply temperature and available flow) and server models. We'll help you identify the right retrofit path and recommend a CDU and cold plate configuration, and we'll get back to you within one business day.