Direct-to-Chip
vs. Immersion Cooling
for High-Density GPU AI Clusters
How each liquid-cooling architecture works, what affects deployment fit, and how to think about the infrastructure decision — not just the equipment comparison.
Why This Comparison Matters Now
Thermal management has become a core infrastructure decision — not a support function.
AI infrastructure is changing data-center cooling requirements faster than many facilities were originally designed to handle. As GPU density rises, thermal management becomes more than a support function — it becomes a strategic infrastructure decision that shapes how servers are deployed, serviced, and scaled.
That is why direct-to-chip and immersion cooling are now being evaluated as architecture choices, not just equipment options. Both approaches move beyond traditional air-only cooling by removing a larger share of heat with liquid. But they do not solve the problem in the same way, and they do not fit every deployment equally well.
MRX approaches this comparison from an engineering and deployment perspective. The right answer depends on the thermal density, system architecture, service strategy, facility constraints, heat-rejection path, and long-term operating priorities of the specific project.
Key Takeaway
The question is not "which technology is better?" The better question is "which architecture fits the deployment?" — and answering that requires evaluating thermal density, service model, facility fit, and the full heat-rejection chain together.
Architecture 1
Direct-to-Chip Cooling
Direct-to-chip cooling removes heat directly from major heat-generating components through cold plates mounted on CPUs and GPUs. Coolant circulates through these cold plates, absorbs heat at the chip level, and moves through a broader cooling loop where that heat is transferred out through a CDU, heat exchanger, or related thermal infrastructure.
Common architectural elements:
Often attractive because it targets the hottest components directly while preserving a more familiar server and rack architecture than full immersion.
Architecture 2
Immersion Cooling
Immersion cooling removes heat by placing server components into a dielectric fluid bath instead of cooling them only with air or cold plates. The fluid absorbs heat directly from immersed components, and the thermal system transfers that heat out of the bath and into the broader rejection loop.
Common architectural elements:
Often considered when operators are pursuing very high thermal density, strong liquid heat capture, or a different service model built around immersion-ready hardware and tank-based infrastructure.
Architecture-Level Comparison
The Core Architectural Difference
The most important difference is not just where the liquid goes — it is how the entire cooling architecture is organized. This comparison should be made at the system level, not the component level.
These differences can affect facility retrofit complexity, heat capture percentage, maintenance workflow, rack and hardware packaging, integration with existing white space, and long-term scaling strategy — which is why the decision should be made at the architecture level, not only at the component level.
Deployment Fit
When Each Architecture Often Makes Sense
Where Direct-to-Chip Often Makes Sense
Direct-to-chip is often a strong fit when the operator wants to increase cooling performance while retaining a deployment model that is closer to conventional rack-based infrastructure. This does not make it simpler in every case, but it often makes it more compatible with existing operational assumptions.
Where Immersion Often Makes Sense
Immersion is often considered when the deployment is comfortable with a more purpose-built liquid architecture and wants to maximize liquid heat capture through a tank-based approach. Immersion is not automatically better — it is a better fit when the deployment model, maintenance model, and thermal strategy all support it.
Evaluation Framework
What Buyers Should Compare
The best comparison is not "which technology is better?" The better question is "which architecture fits the deployment?" — and that requires evaluating four dimensions together.
Thermal Capture & Density
- How much heat must be removed by liquid versus air?
- What rack densities are expected now — and after future GPU generations?
- How much hybrid air support is acceptable in the design?
- How does the cooling architecture perform under sustained peak AI loads?
- What future densification is likely and how quickly?
Service & Operations
- How will hardware be accessed, serviced, and replaced in the field?
- What maintenance workflow does the operations team prefer?
- How does fluid handling affect day-to-day operations?
- How quickly must failed components be accessed or replaced?
- Is the site prepared for immersion-specific or liquid-loop-specific maintenance?
Facility Fit & Retrofit
- Is the project a retrofit of existing space or a new deployment?
- How much white-space disruption is acceptable during phased rollout?
- How is the heat-rejection path currently designed or planned?
- What floor loading, layout, and utility conditions exist on-site?
- Does the project need a highly adaptable or purpose-built architecture?
Controls, Monitoring & Reliability
- What is the leak detection and fluid management strategy?
- How are cooling controls integrated with the broader facility or BMS?
- What thermal monitoring and alarming is required?
- What redundancy is required for the cooling infrastructure?
- What are the recovery and maintenance procedures for thermal events?
End-to-End Thermal Chain
Why the Heat-Rejection Path Still Matters
Neither direct-to-chip nor immersion eliminates the need for broader thermal infrastructure. Both still depend on what happens after the heat leaves the rack or enclosure. The success of the architecture is often tied to what surrounds it — not just to the rack-level technology.
This is why liquid cooling should be evaluated as an end-to-end thermal chain — not just as a rack-level technology choice. The performance of the cooling architecture at the chip or tank depends entirely on the infrastructure behind it.
MRX engineers custom heat exchangers and pump and distribution skids for the secondary-loop infrastructure that both architectures depend on. The rack-level technology is only as reliable as the thermal chain behind it.
Real-World Deployments
Why Hybrid Thinking Is Often Necessary
Many real-world AI deployments are not choosing between air and liquid in absolute terms. They are building hybrid environments — where direct-to-chip often works within hybrid cooling strategies with some air support, and where immersion still depends on broader facility-level infrastructure beyond the tank itself.
Operators often need to evaluate:
MRX Liquid Cooling Infrastructure
MRX approaches liquid cooling from the infrastructure side — defining and building the thermal architecture around the actual deployment. That includes secondary-loop infrastructure, heat exchange strategy, pump and distribution systems, controls, and integration into the broader facility design.
Custom Thermal Infrastructure
When to Consider a Broader Custom Thermal System
In many AI deployments, the real challenge is not only choosing direct-to-chip or immersion. It is building the surrounding thermal infrastructure to make the architecture succeed at scale, under real operating conditions, with the serviceability and redundancy the application requires.
Custom Heat Exchangers
Sized and configured for the specific thermal duty, coolant chemistry, pressure conditions, and secondary-loop architecture of the deployment.
Thermal Management Skids
Integrated pump, valve, instrumentation, and control packages that manage coolant flow, pressure, and temperature across the full cooling loop.
Pump & CDU Infrastructure
Engineered pump and distribution systems matched to the flow requirements, redundancy targets, and facility constraints of the liquid-cooling deployment.
Fluid Distribution Systems
Piping, manifolds, and distribution architecture designed around the facility layout, rack arrangement, and serviceability requirements of the project.
Controls & Monitoring
Instrumentation, sensors, alarm logic, and integration with broader facility BMS or controls systems for thermal visibility and fault response.
Packaged Thermal Subsystems
Complete engineered subsystems — from heat exchange through distribution — designed and fabricated as integrated packages for the deployment.
What Distorts the Comparison
Many cooling decisions get distorted by oversimplified comparisons. The more AI-dense the deployment, the more expensive these misunderstandings become.
Treating them as interchangeable
Direct-to-chip and immersion solve thermal density differently, serve different service models, and have very different facility and hardware implications. Treating them as equivalent options leads to misaligned deployments.
Comparing only heat removal
Heat removal capacity is one dimension. Service model, maintenance workflow, hardware compatibility, facility retrofit requirements, and integration with the heat-rejection path all affect which architecture actually fits the deployment.
Treating hardware impact as secondary
Both architectures affect how servers are packaged, handled, and serviced. Hardware compatibility and serviceability are part of the architecture decision — not afterthoughts to be addressed during commissioning.
Ignoring the heat-rejection path
Neither approach eliminates the need for CDUs, secondary loops, heat exchangers, and pumping infrastructure. The success of either architecture depends heavily on what happens after the heat leaves the rack or enclosure.
Underestimating retrofit complexity
Retrofitting an existing facility for either architecture requires honest assessment of floor loading, utility availability, white-space disruption, phased deployment sequencing, and compatibility with existing IT infrastructure.
Assuming one is more future-proof
Future density increases, hardware generations, and service model evolution can all shift the right answer. Future-proofing depends on the specific deployment trajectory — not on a generic technology claim.
Failing to define operational preferences
Maintenance workflow, fluid handling familiarity, and hardware service access are operational realities that determine whether the chosen architecture works in practice. These must be defined before deployment, not discovered after.
Treating it as a component purchase
This is an infrastructure architecture choice. It sets the design path for how servers are cooled, how heat is rejected, how the facility is organized, and how the deployment can scale. Making it at the component level produces component-level thinking applied to a system-level problem.
Frequently Asked Questions
Liquid Cooling Architecture FAQs
Need Help Evaluating Direct-to-Chip
vs. Immersion Cooling?
Talk with MRX about your rack density targets, heat-rejection path, service model, facility constraints, liquid-cooling goals, and deployment strategy. We can help define the thermal infrastructure that best supports your AI cluster.