Key Takeaways
- AI racks are now routinely exceeding 100 kW of power draw, a density level that pushes traditional air cooling past its physical limits
- Immersion cooling submerges servers directly in a thermally conductive, electrically safe liquid, removing the need for conventional air conditioning hardware
- Switching to immersion cooling can shrink the physical footprint of cooling infrastructure, freeing up floor space for additional compute
- Single-phase and two-phase immersion systems each handle heat transfer differently, and picking the right one depends on workload intensity and facility goals
- Matching a cooling approach to specific workloads and facility constraints is the key to higher density without wasted spend
Data centers built for yesterday’s workloads are running into a wall today, and that wall is heat. AI training clusters draw more power per square foot than most facilities were ever designed to handle, forcing operators to rethink cooling from the ground up rather than patch around the problem.
AI Racks Can Now Exceed 100 kW
A rack that once pulled 5 to 10 kilowatts for standard enterprise computing now might house GPU-based AI hardware drawing well over 100 kilowatts, with some next-generation deployments pushing toward 200 kilowatts. This jump did not happen gradually. It emerged as large language models and generative AI systems proliferated, demanding sustained, near-maximum GPU utilization for days or weeks at a time, generating heat loads that traditional server rooms never had to manage.
The math gets uncomfortable fast. A facility that once comfortably housed 50 traditional racks might only have room, power, or cooling capacity for a fraction of that number once AI hardware enters the picture, even though total computing output goes up. Operators planning new buildouts or retrofits need a clear-eyed view of this shift. Evaluating each facility’s workload and constraints before recommending a cooling path, rather than defaulting to one product, gives a better starting point than guessing at future demand.
Experts like Triton Thermal say that rack density trends show no sign of slowing, and operators who plan around last year’s assumptions risk building infrastructure that is outdated before it’s even finished. Understanding why air cooling struggles at these densities is the first step toward choosing a system that can actually keep pace.
Why Air Cooling Can’t Keep Up
Rack Densities Outgrowing Old Designs
Air cooling works by moving large volumes of cool air across hot components, but air has limited thermal capacity compared to liquid. Removing 40 to 50 kilowatts of heat from a single rack requires airflow rates far beyond what standard raised-floor systems and CRAC units were built to deliver. Push past that threshold, and operators face a tough choice: over-engineer the airflow at significant cost, or accept hotspots and throttled performance.
Average rack loads have been climbing fast, and facilities sized for older assumptions are now being asked to support racks well beyond their original design limits. Retrofitting a legacy data hall for this kind of density often means ripping out and redesigning airflow paths entirely, which is expensive and disruptive compared to planning for liquid cooling from the start.
Cooling’s Heavy Share of Power Use
Cooling has always consumed a significant share of a data center’s total energy draw, and that share grows heavier as air systems strain to keep high-density racks within safe operating temperatures. When fans run at maximum speed and supply air temperatures get pushed artificially low just to stay ahead of thermal threats, the facility’s overall power usage effectiveness degrades, meaning more electricity goes toward cooling instead of computing.
More AI compute drives more heat, more heat drives harder-working cooling systems, and harder-working cooling systems drive higher electricity bills that eat into the value AI investments were supposed to deliver. Breaking that cycle calls for a cooling method that transfers heat more directly and efficiently than air can.
How Immersion Cooling Works
Immersion cooling takes a fundamentally different approach by submerging servers, GPUs, and other electronic components directly into a thermally conductive, electrically non-conductive liquid. Instead of relying on air to carry heat away from components, the liquid absorbs heat at the source, removing the need for the fans, ducts, and air conditioning equipment that traditional data halls depend on.
Single-Phase vs. Two-Phase Systems
Two main approaches exist, each suited to different operational goals:
- Single-phase immersion keeps the dielectric liquid in its liquid state throughout the cooling cycle. Heat absorbed from the components is carried by the liquid to a heat exchanger, cooled, and recirculated, delivering steady performance with a relatively simple system design.
- Two-phase immersion allows the liquid to evaporate when it absorbs heat, then condense back into liquid form in a separate chamber once the heat is released. This phase-change process transfers heat with even greater efficiency, which matters for the densest AI and HPC deployments.
Why It Suits AI and HPC Workloads
High-performance computing and AI workloads generate concentrated, sustained heat loads that outpace what air or even partial liquid solutions like rear-door heat exchangers can manage gracefully. Direct contact between liquid and component surface gives immersion cooling a heat transfer advantage that scales well with the kind of nonstop, high-utilization processing AI training demands. Two-phase systems tend to suit the most intense HPC and AI applications, since their heat transfer efficiency keeps pace with GPUs running at full load for extended periods.
Compute Gains, Smaller Footprint
Cutting Cooling Infrastructure Space
One of the practical advantages of immersion cooling shows up in square footage. Compact immersion tanks and their supporting equipment take up meaningfully less floor space than the CRAC units, ductwork, and air handlers that traditional cooling requires, freeing up that reclaimed space for additional server capacity or other operational needs. For operators working within a fixed building footprint, that space savings can translate directly into more compute per square foot, which is often the real constraint once power and cooling are solved.
Steadier Temperatures, Longer Hardware Life
Because the liquid absorbs heat directly from every submerged component, immersion cooling tends to prevent the uneven temperatures and localized hotspots that air cooling struggles to avoid, especially in densely packed racks. Consistent thermal conditions reduce the stress that heat cycling places on sensitive electronics, which can extend the usable life of GPUs and other expensive hardware. For data centers running AI training clusters around the clock, that reliability advantage matters as much as raw cooling capacity.
What Switching Actually Requires
New Infrastructure and Maintenance Needs
Moving to immersion cooling is a meaningful infrastructure change rather than a drop-in swap. It calls for new liquid distribution infrastructure, dedicated heat rejection systems, leak detection and containment measures, and maintenance procedures that differ from what most facilities teams are used to with air-cooled equipment. Staff need training on handling dielectric fluids safely, servicing submerged hardware, and monitoring systems built around liquid rather than airflow. These are real operational shifts, and planning for them early prevents costly surprises during deployment.
Matching the Solution to the Workload
Not every workload or facility benefits equally from immersion cooling, and that is where a technology-agnostic evaluation earns its value. A moderate AI deployment might do well with a hybrid air and liquid-assist approach, while a dense, space-constrained AI training cluster may justify a full immersion build-out. Matching the cooling solution to the specific workload, facility constraints, and business objectives at hand works better than assuming one technology fits every situation. Operators considering a shift can look at Triton Thermal’s modular AI infrastructure approach built around Atlas One for a sense of how a scalable, purpose-built deployment can come together.
Liquid Cooling’s Rapid Market Rise
The broader liquid cooling market reflects just how quickly this shift is happening across the industry. The global data center liquid cooling market grew substantially between 2023 and 2024, driven almost entirely by AI workload requirements, as operators move away from air cooling not out of preference for novelty but out of necessity.
Coolant distribution units supporting modular immersion and direct-to-chip setups can now provide capacity ranging from roughly 200 kilowatts up to 1 megawatt, letting operators scale cooling in step with GPU deployment rather than overbuilding capacity upfront. That kind of modular scaling is becoming a defining feature of how new AI data centers get designed.
Immersion Cooling’s Rise as the Preferred Choice
Across the industry, immersion cooling is increasingly viewed as the preferred thermal management path for the densest AI and HPC deployments, largely because of its cooling efficiency and its ability to support ultra-high-density GPU racks that air cooling simply cannot reach. As AI hardware continues its trajectory toward higher power draw per chip, facilities that have already built immersion or hybrid liquid cooling capability will be positioned to absorb new GPU generations without another round of disruptive retrofits.
Operators who move early on this shift build the kind of flexible, high-density infrastructure that lets future compute investments pay off instead of hitting a thermal ceiling. For a practical next step, looking into modular liquid cooling infrastructure built specifically for AI workloads is a reasonable place to start planning the transition.
Triton Thermal
3350 Yale St.
Houston
Texas
77018
United States
