Immersion Cooling for Data Centers: What Developers Need to Know Before They Commit
A developer's guide to immersion cooling -- when it makes sense, what it changes, and where the hidden costs live.
High-density AI workloads are pushing power densities past what conventional air cooling can handle. Immersion cooling is the most direct solution. But for institutional data center developers, immersion is not a cooling equipment upgrade. It is a facility architecture decision that changes the building layout, maintenance model, power distribution design, and capital cost profile of the entire project.
Here is what you need to evaluate before it lands in your design brief.
What Immersion Cooling Is
Servers or server components are submerged in a non-conductive dielectric fluid. The fluid absorbs heat directly from the chips, removing the need for fans, raised-floor air management, and most of the mechanical air-handling equipment that dominates conventional white-space design.
Heat captured by the fluid is transferred via a heat exchanger to a facility-side loop, which rejects it to a chiller, dry cooler, or heat-recovery system outside the white space. The servers still generate heat. The question is how quickly and efficiently that heat moves from chip to atmosphere.
Two Types: Single-Phase vs Two-Phase
Single-phase immersion keeps the dielectric fluid in liquid form throughout the cycle. Servers sit in tanks, the fluid circulates through a heat exchanger connected to a coolant distribution unit (CDU), and heat is rejected via the facility cooling loop. The fluid does not change state, which makes the system operationally simpler.
Two-phase immersion uses a fluid engineered to boil at a low temperature. As heat is generated at the chip, the fluid vaporizes, rises to a condenser at the top of the tank, and returns to liquid. The phase-change process is thermodynamically efficient but introduces more fluid management complexity, tighter tank tolerances, and a more specialized maintenance regime.
For most institutional developers evaluating new builds in 2026, single-phase is the more practical starting point. Two-phase is technically impressive but has a smaller vendor base and higher operational learning curve.
Power Density Capabilities
Air cooling in a conventional raised-floor data center can typically handle 8-12 kW per rack without supplemental cooling. Direct liquid cooling and rear-door heat exchangers push that to 30-50 kW. Immersion extends the practical ceiling to 100 kW per tank and beyond.
For AI training clusters and GPU-dense inference racks, that headroom is not theoretical. Modern NVIDIA H100 and GB200 servers dissipate 10-14 kW per server unit. A 10-server rack can exceed 100 kW without overclocking. Air cooling cannot keep pace. Immersion is built for this load profile.
What It Changes in the Building
Immersion tanks do not fit the standard hot-aisle/cold-aisle rack paradigm. They require:
A different floor layout with tank footprints, service clearances, and CDU placement replacing traditional aisle architecture
Updated structural loading calculations -- immersion tanks filled with fluid are substantially heavier than rack rows
Revised power distribution to feed tanks and CDUs, including A/B feed redundancy planning
CDU connections to facility cooling loops, with appropriate isolation and leak detection
Dedicated pipework and fluid containment systems
For new builds designed around immersion from the start, this is manageable. For retrofit projects, it is often cost-prohibitive. The floor plan needs to be reconfigured, the CRAC/CRAH infrastructure removed, and new CDU routes installed through an existing building. Developers evaluating brownfield or shell-and-fit-out conversions should model the retrofit cost before assuming immersion is viable.
Water Use Comparison
Conventional evaporative cooling systems use significant amounts of water for cooling tower makeup. A 100 MW data center on evaporative cooling can consume 1-2 million gallons per day in peak summer conditions.
Immersion cooling removes most of the air-side evaporative load from the white space because heat is captured in the dielectric loop and rejected via the facility-side system. If the facility-side heat rejection is a dry cooler or air-cooled chiller, water use drops substantially. If it uses a cooling tower, some water demand remains.
The practical water savings depend on how the building-side heat rejection is configured -- not just on the choice of immersion at the server level. Developers in water-stressed markets should model the entire heat rejection pathway before treating immersion as a zero-water solution.
Deployable Vendors in 2026
The immersion market has matured from proof-of-concept to commercial deployments at scale. Vendors with production deployments include:
Submer (Barcelona-based, single-phase, strong European and hyperscale traction)
LiquidStack (two-phase, hyperscale deployments including major cloud tenants)
Green Revolution Cooling (GRC) (single-phase, US-based, colo and enterprise focus)
Vertiv (CDU and heat rejection systems supporting multiple immersion architectures)
Iceotope (chassis-level single-phase, modular approach)
Hyperscale operators including Microsoft, Google, and Meta have active immersion deployments or pilots. Colo providers including CyrusOne and Aligned have announced or are running immersion capacity. The technology is past the pilot phase for AI-focused builds.
The vendor question for developers is less "who has a product" and more "who can deliver a bankable, serviceable system with proven controls, CDU integration, and field support at your required scale and timeline." That narrows the list considerably for projects starting construction in 2026 or 2027.
Limitations Developers Need to Model
Upfront cost. Immersion tanks and CDU infrastructure add capital cost relative to air-cooled alternatives. The savings come through reduced CRAC/CRAH equipment, smaller mechanical plant, and potentially lower operating energy cost. The break-even depends heavily on power density, operating hours, and local energy costs. Model it explicitly.
Maintenance complexity. Accessing servers in an immersion tank requires removing, draining, or partially displacing fluid. Service procedures are more involved than pulling a server from a rack. Operators need trained staff, specialized tools, and a fluid handling protocol. Factor this into the operating cost comparison.
Fluid management. Dielectric fluids require periodic testing, top-up, and eventual replacement. Spills require containment and disposal procedures. The CDU primary loop needs monitoring for contamination and fluid degradation. These are manageable but not trivial.
Industry maturity. The immersion market is still developing compared to air-cooled and rear-door heat-exchanger deployments. Some vendors are still scaling manufacturing capacity, lead times are not yet as predictable as established mechanical equipment, and second and third tier service support in some markets is thin.
Not a universal fit. Immersion is the right choice when power density demands it or when energy and water constraints make it worth the additional upfront cost. For general-purpose colocation at moderate power densities, the operational simplicity of well-designed direct liquid cooling or rear-door heat exchangers may be preferable.
The Developer Decision Framework
Before committing to immersion in a new build, answer four questions:
What is the power density requirement? Immersion earns its complexity when racks exceed 50 kW and AI compute workloads are the primary product.
What is the heat rejection strategy? Immersion at the server level still needs a building-side solution. Model the full pathway from chip to atmosphere.
Who is the operator? Immersion requires trained operations staff and a maintenance program built around liquid systems. If the operator is inheriting the facility without immersion experience, budget for transition.
What is the construction timeline? Immersion requires early design commitment. Retrofitting later is expensive. If immersion is on the table, it needs to be in the schematic design phase, not as a late-stage change.
For AI data center developers building for high-density GPU tenants in 2026, immersion cooling is a legitimate design strategy rather than an emerging technology. The projects that succeed with it treat it as an architecture decision from day one, not a cooling system swap.