Immersion Cooling for Data Centers: Does It Fit Your AI Hardware?

Immersion cooling places electronic equipment in an electrically insulating liquid to carry heat away from the components. For an AI data center, it is worth evaluating when the selected servers, fluid, cooling system and service model are supported together. A tank's advertised kilowatts cannot establish that a particular GPU platform will work inside it.

This guide is for operators choosing the cooling architecture before committing to hardware and facility design. For the wider set of options, start with our data center liquid-cooling guide.

TL;DR

• Choose immersion around an approved server configuration, including networking and service procedures.
• Compare it with direct-to-chip on useful workload, complete facility scope and operating costs.
• Single-phase liquid stays liquid; two-phase systems evaporate and condense the working fluid.
• The tank still needs a path to outdoor heat rejection. Dielectric fluid does not establish a water-free facility.
• Before buying, resolve compatibility, warranty ownership, floor loading and a witnessed maintenance trial.

What immersion cooling changes

In immersion cooling, the dielectric fluid touches the electronics. In direct-to-chip cooling, coolant stays inside cold plates and pipework attached to selected components. These are different equipment interfaces. The OCP Immersion Requirements, revision 2.1, distinguish single-phase circulation from two-phase evaporation and condensation. Both transfer heat onward to facility infrastructure.

A practical selection process therefore starts with two documents: the exact server bill of materials and the cooling supplier's supported equipment schedule. Match them line by line. A processor approval cannot settle the status of a storage device, connector, power supply or assembled server.

Immersion also changes the maintenance environment. Ask the operator to describe how an engineer will isolate a server, disconnect it, remove it, manage the fluid on it and return it to service. If that sequence cannot be demonstrated with the intended equipment, the design is still incomplete.

Does immersion cooling make sense for AI?

It can, but the answer depends on the hardware procurement route. A buyer developing a supported immersion server configuration has a different decision from a buyer who has already ordered complete OEM racks. In the second case, the cooling interface is part of the equipment contract.

For example, suppose a planned GPU cluster would exceed the existing hall’s air-cooling capability and hardware procurement is still open. Compare supported immersion and direct-to-chip proposals against that same workload, floor area and outdoor-plant limit. Immersion deserves investigation if a supplier can substantiate a workable configuration within those constraints; the tank alone does not prove an advantage.

NVIDIA's published GB200 NVL72 architecture, for example, describes compute trays with cold plates and liquid connections. That is evidence for a specific direct-to-chip arrangement, not authorization to immerse the same system. Ask the selected OEM for the supported configuration rather than inferring compatibility from the term “liquid cooled.”

For distributed AI, evaluate the complete cluster. A thermally acceptable server is insufficient if its network connections, cable routes, storage path or replacement process cannot support the workload. Test the intended application at its required concurrency and operating conditions; record temperature, power, throttling, errors and useful throughput together.

Use these three situations to decide what to investigate first:

• Hardware already fixed: establish the OEM cooling and warranty requirements before considering a conversion.
• Hardware still open: obtain supported quotations for immersion and direct-to-chip, using the same workload and availability brief.
• Existing facility constrained: identify the actual constraint—airflow, floor loading, power, outdoor plant or servicing space—before proposing a tank retrofit.

Our GPU data center planning guide follows that workload-to-facility sequence in more detail.

Immersion versus direct-to-chip: compare the project

The following table is a procurement framework. It tells you which evidence to request; it does not rank every product using either technology.

Compare a fixed application output and service requirement. If one proposal uses a different GPU generation, workload precision or utilization assumption, separate that effect from cooling. A cooling comparison should not quietly become a comparison of different computing systems.

Compatibility includes materials and high-speed connections

The OCP material-compatibility guide records different outcomes for different material/fluid pairings. “Dielectric” describes an electrical property; it does not establish chemical compatibility. The OCP IT equipment design guide also calls for signal-integrity validation and addresses optical connections.

Ask suppliers to identify what was tested, for how long, at which temperatures and against which acceptance criteria. Then check whether the evidence covers the production bill of materials rather than an earlier sample. Record substitutions: a new cable, seal or adhesive should trigger a review of the affected approval.

The buyer's deliverable is a controlled compatibility matrix: component, revision, fluid, operating range, evidence and responsible supplier. Keep unsupported entries visible. A list of brands without model numbers and conditions is too broad to freeze procurement.

Resolve warranty ownership before the order

Intel's immersion compatibility guidance says some Xeon SKUs can be used and recommends validation with the OEM. Its statement cannot be extended to every assembled server or GPU platform. The OCP warranty guidelines provide a framework for discussing responsibilities; they do not replace the actual supplier contract.

For procurement, request a written answer to a specific failure scenario. If a server fails after immersion, who diagnoses it, who transports it, which fluid records are required and what condition must it be in for return? Agree the handling of disputed causes and the replacement route while the system is being specified.

A useful support schedule names a first contact, escalation owner and evidence package for each interface. It should also say how a replacement server is checked before immersion and how changes in equipment or fluid affect support. Avoid relying on a verbal assurance that “the warranty is fine.”

Follow the heat beyond the tank

Write a thermal schedule covering the immersed IT, equipment outside the tanks and the auxiliaries entering each cooling path. Give each row a heat load, design condition and responsible supplier. A server tank does not remove the need to assess room loads, switches, electrical equipment or outdoor plant.

For a simple illustration, suppose the facility side must transfer 200 kW into a water loop with a 10 K rise. Using water density of approximately 1,000 kg/m³ and specific heat of approximately 4.18 kJ/(kg·K):

Mass flow = 200 ÷ (4.18 × 10) = 4.78 kg/s; volume flow ≈ 17.2 m³/h.

This is an illustrative heat balance, not a tank selection or hydraulic design. It applies to the water side under the stated assumptions, not the dielectric circuit. The actual design also needs fluid properties at operating temperature, pressure losses, heat-exchanger approach, pump duties, auxiliary heat and the required failure cases.

Check the outdoor design at the site's critical ambient condition. Request the available duty when a scheduled component is maintained or a specified failure occurs. Keep the tank supplier's operating limits aligned with the facility supplier's performance curves. See our CDU sizing guide for the interface method; confirm whether the chosen system requires a CDU and which interface arrangement applies.

Energy and water claims need a complete boundary

Dielectric liquid inventory and annual water consumption are different quantities. A facility can circulate its IT coolant in a closed system and still consume water outdoors. US Department of Energy guidance connects cooling-tower water consumption to the heat load and the heat-removal system.

Ask for the proposed rejection technology, operating temperatures, weather file, load schedule and control assumptions. Then compare annual electricity, annual water and peak make-up demand. Our liquid-cooling water guide explains the distinction between a closed loop and a water-consuming plant.

For energy, include the same loads and metering boundary in both options. If server fans are removed or IT power changes, disclose that change instead of crediting every difference to facility cooling. Add energy per useful application output to the comparison where hardware configurations differ. A universal PUE or percentage-saving claim cannot substitute for this model.

A witnessed service trial makes the operating model visible

Before a production commitment, agree a demonstration with representative equipment and the people expected to maintain it. Time the entire intervention, including preparation and return to operation. Use the following sequence as a proposed acceptance exercise:

1. Identify and isolate a selected server using the planned monitoring and isolation method.
2. Remove it with the intended tools and available headroom.
3. Handle fluid, cabling and the work area using the supplier's approved procedure.
4. Replace a representative component or exchange the server.
5. Restore the equipment and confirm application operation, alarms and fluid condition.
6. Record staffing, elapsed time, consumed materials and unresolved support questions.

Ask for the supplier's safety and operating instructions first. Demonstrate the hardest credible intervention as well as routine inspection.

What belongs in the immersion-cooling RFP?

A proposal should leave the buyer with a supported equipment schedule, a thermal interface sheet, a layout and a support plan. Request:

• Exact IT configuration, workload, operating limits and compatibility evidence.
• Capacity at stated fluid and facility temperatures, including the agreed failure cases.
• Tank dimensions, filled and installed weights, support reactions, lifting and access requirements.
• Fluid technical and safety data, sampling plan, replacement route and disposal responsibilities.
• Included electrical, hydraulic, monitoring and heat-rejection scope.
• Warranty and service responsibilities across every supplier interface.
• Factory and site acceptance criteria, test records and a witnessed service trial.

Compare cost over the same period using current quotations. Include servers, cooling plant, fluids, site work, commissioning, energy, maintenance and replacements. Identify exclusions explicitly. Without a site and supported hardware configuration, a generic price per immersed kilowatt is not a project budget.

Where ModulEdge fits

ModulEdge's documented architectures include rear-door cooling for air-cooled IT in MDC-500 and direct-to-chip interfaces with separate air-side cooling in MDC-2000. Those designs provide context for comparing facility interfaces. They do not establish a standard ModulEdge immersion product or universal support for immersing customer equipment.

If immersion is being considered, establish the supported IT and tank arrangement before mapping the modular building, electrical and thermal scope. The useful first discussion starts with the hardware bill of materials, workload, site conditions and supplier interface requirements.

Frequently asked questions

Can GPUs use immersion cooling?

A GPU-based server may use immersion when its complete configuration is supported for the chosen fluid and system. GPU identity alone does not establish server compatibility or warranty coverage.

Can I immerse a liquid-cooled OEM rack?

Obtain explicit OEM support for that configuration. A rack supplied for cold plates has a different cooling interface; its liquid-cooling description does not authorize immersion.

Does immersion cooling use ordinary water?

The electronics are immersed in an electrically insulating working fluid. A separate facility loop may use water to receive and reject the heat.

Is immersion cheaper than direct-to-chip?

There is no universal project price comparison. Use matched server/workload requirements, current quotations, site scope, energy assumptions and service costs.

Does immersion eliminate facility water consumption?

Not necessarily. Annual water use depends on the outdoor heat-rejection design and its operation. Read the water boundary separately from the tank-fluid inventory.

Which immersion technology should I evaluate?

Read our single-phase versus two-phase comparison for a focused framework covering phase change, fluid selection and operating evidence.