CoolIT Fanless Liquid Cooling Pushes AI Racks Beyond the Hybrid Cooling Limit

CoolIT says its fanless liquid cooling designs can capture effectively all heat from AI servers, addressing a 75 kW problem inside a 250 kW rack.

That figure exposes the limit of today’s common hybrid approach. A system capturing 70 percent of its heat with liquid still leaves 30 percent for air. At 250 kW, fans and facility equipment must remove the remaining 75 kW.

CoolIT argues that this residual load makes hybrid cooling increasingly difficult to justify. Its proposed alternative extends coldplates beyond processors to memory, networking, storage, and power components. The target is less than 1 percent of rack heat remaining in the air.

The company calls that near-total heat capture, while using 100 percent as a design objective. This distinction matters because cables, circuit boards, and other surfaces still release small amounts of heat.

The larger story is not whether liquid cools processors better than air. That contest has largely been settled for the densest AI hardware. The new contest is between hybrid servers that preserve an air-cooling layer and fanless systems designed around liquid from the start.

CoolIT’s argument arrives as Nvidia and server manufacturers push rack-scale computing well beyond traditional data-center densities. Nvidia’s GB200 NVL72 already requires about 120 kW of cooling capacity, according to its rack design. Future systems are moving toward several hundred kilowatts per rack.

At those levels, cooling the processor alone no longer solves the thermal problem.

CoolIT Moves Liquid Beyond the Processor

CoolIT fanless liquid cooling treats the entire server as a thermal system, not a collection of hot processors surrounded by air-cooled parts.

The company’s September 2026 argument starts with an important change in where heat appears. GPUs and CPUs remain the largest heat sources, but they no longer represent the complete cooling challenge.

Memory modules sit close to high-power processors and handle increasingly intensive data movement. Network switches and interface components support larger clusters with faster links. Storage, voltage regulators, and power-conversion hardware also carry heavier loads.

These components once operated comfortably within the airflow already moving through a server. That arrangement allowed manufacturers to place coldplates on processors while retaining fans for everything else.

Coldplates are metal heat exchangers mounted directly against electronic packages. Coolant passing through internal channels absorbs heat before carrying it toward a coolant distribution unit, or CDU.

A CDU isolates the equipment-side cooling loop from the facility-water loop while controlling temperature, pressure, and flow. It is a central part of most direct liquid cooling deployments.

The CoolIT heat-capture claim extends this loop to the components that hybrid designs leave behind. The company says it assembles these systems from modular coldplate blocks tested across six generations of fanless designs.

That modular approach matters because peripheral components are less uniform than processors. A CPU or GPU usually presents a flat package with defined mounting points. Memory, power components, switches, and storage devices arrive in different shapes and locations.

Each component also has its own temperature ceiling and mechanical constraints. A complete liquid loop must accommodate those differences without restricting service access or placing harmful force on a circuit board.

CoolIT is therefore presenting fanless cooling as a server-design problem, not simply a better processor coldplate. The coolant path, connectors, manifolds, controls, and component interfaces must operate as one system.

The company says its architecture can reduce the air-cooled share below 1 percent. It also acknowledges that literal 100 percent capture is nearly impossible because some heat escapes from surfaces outside the liquid path.

That qualification makes “fanless” more precise than “zero airborne heat.” A fanless server can tolerate a tiny passive heat release without using internal fans to maintain component temperatures.

This is the immediate change behind the announcement. Liquid cooling is expanding from the hottest packages into nearly every meaningful source of heat inside an AI server.

A 250 kW Rack Breaks the Hybrid Cooling Equation

The problem with a percentage-based cooling plan is that its residual air load grows with every increase in rack power.

A 70/30 liquid-air split can look effective when expressed as a ratio. It means liquid removes most of the heat, while familiar fans and room systems handle the remainder.

The absolute numbers tell a different story. Thirty percent of a 50 kW rack is 15 kW. Thirty percent of a 250 kW rack is 75 kW.

That residual load is comparable to the total power of several conventional racks. It sits inside one cabinet already crowded with processors, networking, power shelves, coolant manifolds, and high-speed cabling.

Removing that heat requires substantial airflow. Operators must supply cool air, move it through narrow server passages, collect the exhaust, and prevent it from recirculating into equipment inlets.

Fans also consume electricity and occupy space within servers. Higher airflow creates pressure, acoustic, maintenance, and filtration demands. Those requirements continue even after the facility installs liquid infrastructure for the processors.

The operator then maintains two cooling systems for the same rack. Pumps, heat exchangers, manifolds, and coolant controls serve the liquid loop. Fans, air handlers, containment, and room cooling serve the remaining air load.

CoolIT’s central claim is that this duplication becomes commercially unattractive beyond roughly 250 kW. That threshold comes from company modeling, not a universal engineering standard.

Rack feasibility depends on server geometry, coolant temperatures, climate, facility design, redundancy requirements, and allowable component temperatures. A carefully engineered hybrid system does not suddenly stop working at one exact number.

However, the underlying arithmetic is difficult to dispute. A fixed percentage of a rapidly growing total becomes a rapidly growing load.

ASHRAE’s AI cooling framework says rack densities have moved from about 120 kW toward several hundred kilowatts. It also anticipates megawatt-class racks in the near term.

The framework describes direct-to-chip liquid cooling as an industry standard for AI and high-performance computing. Direct-to-chip means coolant removes heat through coldplates attached to selected electronic components.

CoolIT takes that trend one step further. If nearly every high-load component receives a liquid interface, the room no longer needs to move large air volumes through the servers.

That can simplify white-space design, the data hall area where racks operate. It can also shift more heat rejection toward warm-water systems and outdoor heat exchangers.

Yet the cooling work does not disappear. Pumps must circulate coolant, CDUs must transfer heat, and the facility must ultimately release that heat outside.

Fanless AI servers change where the work happens. They replace high-volume internal airflow with a controlled liquid path that delivers heat to infrastructure built for concentrated loads.

This is why the 250 kW figure should be read as a design warning, not a universal cliff. It marks the point where preserving air cooling can create a second major infrastructure burden.

CoolIT Fanless Liquid Cooling Challenges the Hybrid Default

The main contest is now full liquid coverage versus hybrid cooling, not liquid versus air at the processor.

Hybrid cooling became a practical bridge because it preserved familiar server and data-center designs. Operators could add liquid to high-power chips without rebuilding every component or abandoning established airflow practices.

That compromise reduced deployment risk. It also allowed manufacturers to use proven air-cooled memory, storage, networking, and power designs around newly liquid-cooled accelerators.

Nvidia’s GB200 NVL72 illustrates the transition. The rack connects 72 Blackwell GPUs and 36 Grace CPUs through a high-bandwidth NVLink fabric. Its compute and switch trays include liquid-cooling connections, but deployment still involves extensive rack and facility integration.

The system’s 120 kW cooling requirement already exceeds what many existing data halls were designed to deliver. Newer platforms are expected to place even more power within a similar physical footprint.

CoolIT says flagship rack-scale systems will move toward full heat capture through 2028. That forecast aligns with the direction of processor power, although the timing remains the company’s assessment.

The industry is not waiting for one supplier to establish this category. Hewlett Packard Enterprise introduced its own 100 percent fanless direct liquid cooling architecture in October 2024.

HPE’s design covers GPUs, CPUs, memory, storage, networking, power, racks, pods, and CDUs. The company says its architecture reduces cooling power per server blade by 37 percent compared with hybrid direct liquid cooling.

HPE also claims a 90 percent reduction in cooling power compared with traditional air-cooled systems. Those figures come from vendor testing and should not be treated as universal operating results.

Still, HPE’s fanless architecture validates the broader direction. Cooling more components with liquid is becoming an architectural choice among major server suppliers.

Immersion cooling offers another route. It places servers, or selected assemblies, inside dielectric fluid that directly absorbs heat from exposed components.

Immersion can capture nearly all equipment heat without individual coldplates. However, it changes hardware packaging, maintenance procedures, fluid handling, warranties, and component qualification.

Coldplate systems preserve more familiar rack and server formats. Technicians can access dry components after isolating and disconnecting the appropriate cooling lines.

That familiarity gives direct liquid cooling an adoption advantage in facilities already preparing for Nvidia-style rack systems. It also leaves a demanding integration problem inside each server.

Adding a coldplate to every relevant component increases the number of interfaces. Designers must manage flow resistance, temperature differences, tubing routes, connector placement, and the risk of one component receiving inadequate cooling.

Modularity can help. CoolIT says its reusable coldplate building blocks shorten development while supporting components with different physical and thermal requirements.

The claim is plausible, but the decisive evidence will come from complete production systems. Operators need performance data across realistic workloads, coolant conditions, maintenance cycles, and failure scenarios.

This is where the opponent becomes clear. Hybrid cooling offers gradual adoption and component familiarity. Full liquid coverage promises higher density with less dependence on server airflow.

At lower rack densities, operators can still prefer the flexibility of a hybrid design. At several hundred kilowatts, the residual air load increasingly dictates the architecture.

Fanless Does Not Mean Infrastructure-Free

Near-total heat capture removes server fans from the critical path, but it transfers more responsibility to coolant delivery and facility engineering.

A fanless server cannot compensate for weak coolant flow by increasing fan speed. Its thermal stability depends on pumps, valves, manifolds, controls, coldplate contact, and heat exchangers operating correctly.

Redundancy therefore becomes essential. Operators must decide which pump, CDU, sensor, or control failures the system can tolerate without throttling processors or shutting down racks.

Leak management also moves closer to expensive hardware. Well-designed systems use tested connectors, pressure controls, dripless quick disconnects, monitoring, and isolation procedures.

The objective is not to claim that liquid never leaks. It is to make failures detectable, containable, and serviceable before coolant reaches vulnerable electronics.

Coolant chemistry creates another operational requirement. Water quality, additives, metals, elastomers, tubing, and temperature all affect corrosion and biological growth.

A system assembled from incompatible materials can degrade internally even when no external leak appears. Long-term reliability depends on validated material combinations and controlled maintenance.

The Open Compute Project’s coldplate requirements reflect this complexity. Its guidance covers coolant types, connection methods, CDU placement, temperatures, pressures, and cooling classifications.

Serviceability matters just as much. A technician replacing memory or a network component may need to disconnect liquid hardware that was absent from older servers.

More cooling interfaces can lengthen repair procedures unless manufacturers design access carefully. Facilities also need trained staff, spare assemblies, coolant-handling equipment, and documented isolation steps.

Fanless systems place new demands on commissioning. Engineers must confirm that each branch receives its required flow and that trapped air does not impair heat transfer.

They must also validate controls under partial loads. AI clusters rarely distribute work perfectly, so one tray can run hotter than its neighbors during a changing workload.

A second uncertainty concerns existing buildings. New AI facilities can be designed around liquid loops, heavier racks, larger pipes, and high-capacity heat rejection.

Retrofitting an older data hall is harder. Floor loading, pipe routes, electrical distribution, maintenance clearances, and CDU locations can restrict the density that operators can deploy.

Some facilities will use liquid-to-air CDUs, which reject coolant heat into room air. This approach can accelerate deployment where facility water is unavailable.

It also preserves the facility-level air burden that near-total liquid capture is intended to reduce. The server becomes liquid-cooled, but the building still moves the heat through air.

Water use needs similarly careful treatment. Direct liquid cooling uses a closed equipment loop, but the facility’s final heat-rejection method determines total water consumption.

A warm-water loop connected to dry coolers can reduce or avoid evaporative water use in suitable climates. A cooling-tower system can still consume significant water.

The appropriate comparison must include the entire path from silicon to the outdoor environment. Server heat-capture percentages alone do not establish total energy use or water performance.

This is the most important check on CoolIT’s narrative. Near-total capture solves the residual server-air problem, but it does not remove the need for system-level engineering.

The company has described a credible direction and a clear density threshold from its modeling. It has not published enough application data to make 250 kW a universal cutoff.

Operators should ask for validated pressure drops, component temperatures, pump energy, redundancy behavior, service times, and facility requirements. Those measurements will determine whether a fanless design simplifies operations in practice.

The Efficiency Case Depends on the Whole Cooling Chain

A stronger heat-capture percentage creates an opportunity for efficiency, but architecture and operating conditions decide whether that opportunity becomes a measured gain.

Server fans are not free. They consume power, occupy space, generate noise, and push air against increasingly restrictive component layouts.

ASHRAE has noted that fan power can represent 10 to 20 percent of server power in some air-cooled configurations. That overhead becomes more consequential when thousands of servers operate continuously.

Liquid has a much greater capacity than air to carry heat through a compact channel. Coldplates also collect heat near its source, avoiding the need to move large air volumes across an entire board.

These properties support denser packaging. Removing fans and large air passages can free internal volume for compute, networking, power, or more direct service access.

However, liquid systems consume pump power. A design with narrow channels, long tubing, or many parallel branches can require greater pressure to maintain sufficient flow.

The result depends on the complete hydraulic design. Lower server-fan consumption does not automatically guarantee lower total cooling energy.

Coolant temperature is another major variable. Warmer liquid can enable more hours of chiller-free heat rejection, depending on climate and workload requirements.

Higher coolant temperatures can also improve opportunities for heat reuse. Waste heat becomes more useful when its temperature is sufficient for nearby buildings or industrial processes.

Component limits constrain that opportunity. Processors, memory, optical devices, power electronics, and storage may require different inlet temperatures and operating margins.

A full liquid loop must satisfy the most restrictive components without overcooling everything else. Multiple loops or carefully controlled branches can address those differences, but they add complexity.

ASHRAE describes water classes ranging from lower-temperature loops to warm-water conditions above 45 degrees Celsius. Equipment compatibility determines which class an operator can use.

Capacity planning must account for transient behavior as well. AI processors can shift rapidly between idle periods and high utilization, changing heat output across racks.

Liquid has thermal mass, which can absorb short changes, but controls must still respond without creating unstable flow or temperature conditions.

Energy efficiency is also inseparable from utilization. A densely packed rack that runs consistently can use space and cooling infrastructure efficiently.

A rack built for peak power but regularly operating far below that level can leave pumps and facility capacity underused. Controls need sufficient range to match changing demand.

This is why vendor percentages require context. HPE’s 37 percent blade-cooling reduction and CoolIT’s below-1-percent air target measure different parts of the system.

One compares cooling power under specified designs. The other describes the share of equipment heat that reaches the air.

Neither number alone reports facility-wide efficiency. Operators still need power usage effectiveness, cooling-system energy, water use, available capacity, and delivered compute performance.

The strongest case for fanless AI servers is therefore not a single efficiency percentage. It is the ability to avoid scaling a separate air system alongside the liquid infrastructure.

At 250 kW and beyond, preventing tens of kilowatts from entering room air can reduce equipment, airflow, and containment requirements. That benefit grows as rack power rises.

The gain must still be verified for each site. Climate, water availability, redundancy, workload, and heat-rejection equipment can change the final result.

Three Signals Will Show Whether Fanless AI Servers Become the Default

The next test is whether near-total heat capture moves from vendor architecture to repeatable deployment across mainstream rack-scale systems.

The first signal is liquid coverage in the next generation of flagship AI racks. Buyers should examine whether vendors cool memory, networking, storage, and power components directly.

Processor-only liquid cooling would preserve the hybrid model. Broad component coverage would support CoolIT’s claim that heat has moved beyond the chip.

Nvidia’s future rack specifications will be particularly important because its reference architectures influence server manufacturers, cooling suppliers, and data-center plans.

The second signal is independently measured facility performance. Operators need results that include pumps, CDUs, heat rejection, water consumption, and workload utilization.

A design that captures more than 99 percent of server heat into liquid should materially reduce room-air demand. Published operating data must show whether total cooling energy falls as expected.

The most useful comparisons will evaluate similar workloads under hybrid and fanless designs. They should also disclose coolant temperatures, climate conditions, redundancy, and utilization.

The third signal is operational standardization. The industry needs interoperable connectors, coolant specifications, service procedures, monitoring interfaces, and safety requirements.

Open Compute Project guidance provides a foundation, but broad adoption depends on suppliers implementing those requirements consistently. Operators cannot treat every rack as a custom mechanical system.

Standardization would also make component replacement and multivendor procurement easier. Without it, facilities risk becoming dependent on proprietary manifolds, fluids, and service processes.

CoolIT’s sponsored fanless server analysis captures a real transition, even though its 250 kW threshold remains a vendor conclusion.

The important reversal is already visible. Liquid cooling was once the supplement added to processors when air reached its limits. Air is now becoming the supplement left behind after liquid handles nearly everything.

That change puts pressure on server manufacturers, facility designers, and operators at the same time. Each group must decide whether preserving hybrid infrastructure still offers useful flexibility.

For buyers planning several-hundred-kilowatt racks, the immediate action is to request complete thermal maps rather than processor specifications alone. Ask which components remain on air and calculate their absolute load.

Then trace every kilowatt through the liquid loop, CDU, facility system, and final heat-rejection equipment. That exercise will show whether CoolIT fanless liquid cooling simplifies the building or merely relocates complexity.

The future will not become fanless because the phrase sounds efficient. It will become fanless when production systems show that full liquid coverage is easier to operate than a 75 kW pocket of air.