A liquid-cooled rack can look remarkably simple when its thermal path runs from silicon to coolant and from coolant to heat rejection, yet the physical system rarely behaves like that clean diagram. Heat crosses boundaries, fluid changes pressure, pumps establish circulation, filters resist movement, valves regulate flow, and heat exchangers impose additional thermal and hydraulic conditions before the heat reaches its final destination. A cooling distribution unit sits directly inside that chain in many architectures, separating the facility-side circuit from the equipment-side circuit rather than allowing the same fluid to move continuously through both. The separation can address legitimate requirements around chemistry, pressure, temperature, serviceability, and equipment protection, while an added boundary can also introduce additional heat-transfer and hydraulic requirements depending on the architecture.

The cooling bill starts before the heat reaches the plant

The useful quantity to follow is not simply the heat removed from the rack but the electrical work required to move that heat through every hydraulic and thermal boundary. For a pump, the core relationship remains straightforward: hydraulic power follows volumetric flow multiplied by pressure differential, while pump, motor, drive, and control efficiency determine the electrical input required to produce that hydraulic work. For a heat exchanger, the thermal approach and pressure drop remain linked design variables because increasing fluid movement can improve heat transfer while simultaneously increasing hydraulic resistance. When several heat exchangers sit in series across separate loops, each exchanger adds its own thermal approach and pressure drop to the overall temperature and hydraulic budgets, while exchanger arrangement and operating conditions determine how the system responds.

The central issue is not whether a CDU can perform its intended function, because established liquid-cooling systems demonstrate that CDUs can regulate temperature, pressure, flow, and fluid separation effectively. The more revealing question is what the complete cooling path would require if the system eliminated the intermediate transfer function and the facility loop could serve the equipment without another thermal conversion step, provided the resulting fluid, pressure, temperature, chemistry, and equipment requirements could be satisfied. That comparison changes the accounting from component efficiency to architecture efficiency, because a highly efficient pump inside a separate loop can still add electrical demand that a suitably designed continuous loop may not require. The same logic applies to heat exchangers: a well-designed plate exchanger can deliver strong thermal performance while still imposing a pressure drop that another pump must overcome.

The Heat That Pays Toll Before It Leaves

The heat leaving a processor does not travel as an abstract quantity from one temperature node to another without interacting with the equipment between those nodes. A liquid carries sensible heat according to its mass flow, specific heat, and temperature rise, but the system must establish the conditions that allow that liquid to reach the heat source and then release the heat downstream. Every liquid-to-liquid exchanger adds a thermal resistance between the two fluids, while the exchanger’s channels, headers, plates, seals, valves, and connections also create hydraulic resistance. The thermal resistance determines how closely the two fluids can approach one another in temperature, while the hydraulic resistance determines how much pressure the pumps must provide to sustain the required circulation. A CDU therefore performs two different physical jobs at once: it transfers heat across a boundary and it creates the hydraulic conditions needed for the secondary loop to operate.

Heat transfer is not free circulation

The cleanest way to model the architecture is to treat every interface as a node in a thermal-resistance and pressure-loss network rather than treating the CDU as a single black box. Heat exchanger effectiveness, approach temperature, fluid properties, channel geometry, and flow arrangement determine the thermal behavior of each node, while pressure drop, fittings, filters, control valves, manifolds, and piping determine the hydraulic behavior. The electrical demand then emerges from the combined requirement to maintain temperature and pressure conditions across the complete path. If a rack needs a particular coolant temperature, a facility loop that operates at a different temperature may require additional heat-transfer surface, increased flow, a different supply temperature, or some combination of those measures depending on the exchanger and control design. Colder supply conditions can shift more work toward the heat-rejection plant, while higher flow can shift more work toward pumps and their drives.

A useful architecture model therefore starts with an energy ledger rather than a component catalog. The ledger assigns electrical work to the facility pump, the CDU pump, the equipment-side pump, control valves, filtration, heat-rejection equipment, and any additional fan or refrigeration load that appears because coolant temperatures cannot remain high enough for direct heat rejection. The rack heat load remains the same in a like-for-like comparison, so the architectural variable becomes the amount of auxiliary energy required to transport and reject that heat. This approach also prevents the analysis from treating the CDU as a standalone electrical load while overlooking the effects that its temperature approach and pressure drop can have elsewhere in the cooling system. A direct-loop design can remove some of those consequences without changing the thermal requirement of the rack itself when the direct configuration can satisfy the rack’s required temperature, pressure, chemistry, and flow conditions.

The first penalty appears at the boundary

The temperature penalty also changes how the plant sees the rack’s return heat. A higher return temperature can improve the opportunity for heat rejection and heat reuse because the cooling system has more thermal heat available above ambient conditions. If an exchanger forces the equipment-side loop to operate at a lower supply temperature than the facility loop can naturally provide, the heat-rejection system must compensate for that lost temperature margin. The compensation can take the form of additional heat-exchanger area, greater fluid movement, colder source water, or mechanical cooling, depending on the site architecture and environmental conditions. None of those responses changes the heat generated by the compute equipment, yet each can increase the auxiliary electrical work required to move or reject it. The temperature approach therefore functions as an architectural tax whenever the system must create a larger thermal gradient than the rack itself fundamentally requires.

Approach temperature is often treated as an exchanger specification, but its real significance appears only when the entire cooling chain is considered. The approach represents the temperature difference that remains between fluids on opposite sides of a heat-transfer boundary under the chosen operating condition. A nonzero approach is unavoidable in practical heat exchangers, because finite surface area, material conductivity, fluid properties, flow arrangement, fouling, and economic design constraints prevent perfect temperature matching. The issue arises when several heat-transfer boundaries sit between the heat source and heat sink, because the cooling plant must maintain temperature conditions that satisfy the complete exchanger arrangement rather than only the rack. The system may then require a different facility supply temperature if the intermediate boundaries prevent the desired equipment-side temperature from being achieved at the available facility-water condition.

Approach temperature becomes an architectural constraint

A direct liquid path makes the temperature chain easier to see because the equipment coolant becomes more directly coupled to the facility heat-rejection condition. The rack still has internal thermal resistance between the chip and coolant, while a direct architecture can remove an intermediate liquid-to-liquid boundary from the path when the facility and equipment loops can operate as a compatible continuous circuit. That can allow the equipment loop to operate closer to the temperature supported by the heat-rejection system rather than requiring another controlled temperature conversion, provided the remaining system can satisfy the equipment’s operating requirements. A site that can use warmer coolant can potentially shift more operating hours toward heat rejection rather than refrigeration, although the actual outcome depends on climate, exchanger design, controls, fluid chemistry, and the thermal limits of the IT equipment.

Approach temperature also interacts with control stability because a cooling system must respond to changes in load without violating equipment temperature limits. When the architecture contains several thermal interfaces, each control loop may regulate a different temperature or pressure variable, creating more opportunities for one loop to constrain another. A secondary pump may increase speed to maintain equipment-side differential pressure while the facility system simultaneously adjusts flow to maintain its own supply condition, depending on the control strategy. The exchanger can then become a point where those control objectives interact, with its thermal behavior influencing the operating margin available between them. Direct architecture can reduce that coupling when it removes separate hydraulic and temperature-control stages from the path between the heat source and facility loop. Fewer variables do not automatically guarantee lower energy use, but a simpler hydraulic path can make the energy flows easier to measure and optimize.

Warm-water intent can die inside a cold boundary

Warm-water cooling works because the liquid can carry substantial heat while remaining at a temperature compatible with efficient heat rejection. The concept loses part of its advantage when an intermediate heat exchanger forces the equipment loop to operate substantially below the temperature that the facility loop could otherwise support. The rack may still receive technically suitable coolant, but the system has surrendered some of the temperature advantage that made direct liquid cooling attractive in the first place. A colder secondary loop also changes the required temperature lift at the heat-rejection side, potentially increasing the work required to reject the same thermal load. The penalty does not necessarily appear in the CDU nameplate because the exchanger and pump perform exactly as specified. It appears instead in the operating condition selected for the wider cooling plant.

The architecture becomes especially sensitive when the rack design has substantial tolerance for warmer coolant but the cooling plant continues to operate around a more restrictive interface. In that case, the exchanger is effectively setting the system temperature rather than the equipment. The plant must satisfy the most restrictive boundary in the chain, even if the other components could operate at warmer conditions. That constraint can reduce economizer opportunities, narrow the useful ambient-temperature window, and increase dependence on mechanical cooling. It can also influence heat-reuse opportunities because useful heat recovery depends strongly on the temperature of the recovered fluid. A direct loop does not automatically create a heat-recovery system, but it can preserve a higher return temperature that makes downstream recovery technically easier to evaluate.

One Thermal Job, Three Hydraulic Jobs

The hydraulic architecture of a liquid-cooled site can contain several pumps that appear to perform different duties while ultimately serving the same thermal objective. One pump moves facility water through the site distribution network, another sits inside the CDU and circulates the equipment-side fluid, and another may support a thermal-control subsystem, rack manifold, or secondary circulation path. Each pump operates against its own pressure requirement, flow requirement, efficiency characteristics, control logic, and operating range. From a component perspective, those are separate engineering functions, but from an energy-accounting perspective they can all consume electricity in support of the same overall thermal load. Pump power therefore has to be calculated from the complete hydraulic path rather than from the rated electrical draw of an isolated pump.

Pump power follows the heat through the loops

The governing relationship is simple enough to expose the architecture: hydraulic power is the product of volumetric flow and pressure differential, while electrical power reflects the efficiency of the pump, motor, drive, and associated controls. Pressure differential itself comes from static requirements and distributed and local losses across piping, fittings, valves, heat exchangers, filters, manifolds, quick-disconnects, and equipment channels. A duplicated loop therefore creates two opportunities for pressure loss where one continuous path might require only one. The additional pressure requirement can remain hidden when the CDU specification reports only its own pump consumption rather than the full cooling-path pressure budget. A site-level model must instead sum pressure losses by flow path and calculate the electrical work required to sustain the resulting operating point.

Variable-speed drives make the accounting more subtle because pump demand changes with operating conditions rather than remaining fixed. Reducing flow can substantially reduce pump power when the system curve and control strategy allow the pump to move toward a lower operating point. A poorly coordinated architecture, however, can force one pump to maintain pressure while another throttles flow, leaving electrical work in the system without delivering proportional thermal value. Control valves can also convert available pump head into local pressure loss rather than useful circulation. A direct loop can simplify the control problem by reducing the number of independently controlled circulation paths, although it must still provide adequate pressure and flow across every rack and branch. The energy advantage comes when the simplified hydraulic network lets the system satisfy thermal requirements with less total head and less duplicated circulation.

The hidden work sits between the pumps

A pump does not know whether pressure is being consumed by useful heat transfer or by an avoidable restriction somewhere in the network. The impeller or motor supplies the pressure required by the system curve, and the system curve includes every local and distributed resistance between the pump inlet and outlet. A heat exchanger with high thermal performance can therefore increase pumping demand if its channel geometry creates significant resistance, while a filter can gradually increase that resistance as contamination accumulates. A quick-disconnect can introduce another local loss, and a manifold can add distribution resistance even when its purpose is simply to divide flow among racks. The energy cost becomes particularly difficult to see when each item sits on a different procurement or maintenance boundary.

Seal drag, bearing losses, motor losses, drive losses, and control overhead add another layer between theoretical hydraulic power and actual electrical consumption. These effects matter because the system does not pay only for the pressure required to overcome the fluid network; it pays for the complete electromechanical chain that creates that pressure. A pump selected for a wider operating envelope may also run away from its best efficiency point under partial load, increasing the electrical penalty without increasing useful heat removal. Multiple pumps multiply the number of operating points that must remain stable as rack demand changes. A direct architecture can reduce the number of these machines, but the remaining pump must then cover a carefully designed distribution network rather than simply inherit a larger burden.

When ΔT Shrinks, Flow Has to Shout

The relationship between heat load, mass flow, specific heat, and temperature rise is one of the most important equations in liquid cooling because it shows why thermal interfaces can create hydraulic consequences. For a sensible liquid loop, the transferred heat follows the product of mass flow, specific heat, and coolant temperature rise, meaning that reducing the usable temperature rise requires greater flow for the same heat load. The equation does not care whether the reduction came from a heat exchanger, a control strategy, a rack constraint, or a conservative operating target. If the system cannot use as much temperature rise across the rack, the hydraulic system must compensate by moving more fluid. Greater flow then changes pressure drop across pipes, heat exchangers, filters, valves, manifolds, and cold plates. Pump energy therefore becomes coupled directly to the thermal decision made at every interface.

The temperature rise determines the required flow

This coupling explains why a seemingly modest approach-temperature penalty can become an architecture-wide energy issue. The exchanger itself may only require additional thermal driving force, but the system responds by adjusting supply temperature or flow. Lower supply temperature increases the thermal burden on the heat-rejection plant, while higher flow increases the hydraulic burden on the pumps. The chosen response depends on the equipment limits and control strategy, but neither response is free. A direct loop can preserve more usable temperature rise because it removes an intermediate fluid-to-fluid boundary and allows the facility loop to interact more directly with the rack cooling path.

Flow also affects the distribution network because pressure losses generally rise as velocity and flow increase. That means a thermal decision at the rack can propagate backward through the manifold and facility piping. A system that increases flow to preserve rack temperature may therefore consume additional pump head throughout the entire loop rather than only at the cold plate. This is particularly relevant in high-density rows where the thermal load concentrates in a relatively small hydraulic network. Once the network approaches a pressure-limited condition, pump speed, valve position, and branch balancing become active parts of thermal management. The hidden energy cost is therefore created by the interaction between thermal design and hydraulic design rather than by either discipline alone.

Flow multiplication can overwhelm component efficiency

A common analytical mistake is to compare pump efficiency without first comparing the flow path that each pump serves. A highly efficient pump moving unnecessary fluid can consume more energy than a less sophisticated pump serving a smaller and better-matched hydraulic requirement. The correct comparison begins with the heat load and the allowable temperature rise, then derives the required flow before calculating pressure loss and electrical power. That sequence prevents the analysis from treating pump selection as an isolated equipment decision. It also exposes the effect of adding another exchanger because the exchanger can change both the required temperature condition and the pressure drop that the pump must overcome.

A CDU-centric system can multiply flow because the facility side and equipment side must each circulate fluid independently. The two flows may carry the same heat but operate under different temperature and pressure conditions, requiring separate pumps and controls. The exchanger couples their heat balance without allowing their fluid movement to become one continuous hydraulic path. Direct architecture can collapse those functions when the fluids are compatible and the rack equipment can accept the facility loop directly. That eliminates the need to circulate an isolated secondary fluid solely to preserve a separation that the architecture no longer requires.

The kW Hiding in Your Secondary Side

The secondary side is often described as the part of the system that delivers coolant to the rack, but that description hides the number of components through which the fluid must pass. Quick-disconnects allow serviceability, filters protect sensitive channels, manifolds distribute flow, valves regulate branches, sensors create control points, and flexible connections accommodate rack movement or maintenance. Every component changes the pressure profile of the loop. The pressure drop may appear individually small in a specification, yet the system experiences the combined resistance of every component installed along the path. A secondary loop can therefore consume meaningful pump head even when the major heat exchanger receives most of the design attention.

Secondary hydraulics rarely stop at the pump

Quick-disconnects deserve particular attention because their value comes from operational flexibility rather than heat transfer. They introduce local geometry changes that can increase resistance compared with a continuous pipe, and the total effect depends on flow rate, internal design, connection size, and the number of interfaces in the path. Filters create a different problem because their resistance can change as the fluid condition changes. A filter that looks acceptable during commissioning can require greater differential pressure after extended operation, forcing the pump to work harder to maintain the same flow. The energy model therefore needs both clean and operating-state pressure-drop assumptions rather than relying solely on the initial equipment specification.

Manifolds can also create hidden hydraulic work because distribution requires balancing pressure across branches rather than simply moving fluid from one point to another. A poorly matched branch can consume more head than the rack actually needs, while a throttled branch can waste pump energy by converting pressure into local loss. The control system may compensate by increasing pump speed, which raises pressure across every branch and can increase losses throughout the network. A direct loop can still require manifolds and branch controls, but removing an intermediate CDU loop can shorten the path between the facility distribution network and the rack manifold. The reduction in path length does not automatically guarantee lower pressure drop, yet it removes an entire class of components from the secondary circulation architecture.

Recirculation can hide a permanent electrical load

Low-temperature-difference operation can create another form of parasitic demand because the loop must move more coolant to transport the same heat. Recirculation becomes particularly important when control logic maintains a narrow temperature band rather than allowing the return temperature to rise naturally. The system can then circulate fluid at a rate determined by temperature-control requirements rather than by the minimum flow required for heat transport. The additional flow raises pressure loss across every hydraulic restriction in the path. Pump power consequently increases even though the rack’s thermal load has not changed.

A secondary loop can also circulate more fluid than necessary because it must maintain minimum flow through equipment or protect against local temperature excursions. Those requirements may be valid, but they should appear explicitly in the energy model rather than being treated as invisible operating conditions. If a minimum-flow requirement exists, the pump must provide the hydraulic work associated with that flow even when the actual heat load remains below the thermal capacity of the loop. The resulting electrical load becomes part of the architecture’s fixed or semi-fixed parasitic demand. A direct architecture can reduce some of that demand by allowing the facility system and rack system to share a hydraulic operating point, provided the resulting pressure, chemistry, and reliability conditions remain acceptable.

What You Get Back When the Middleman Goes

Removing a CDU does not mean simply deleting a box from a schematic and connecting two pipes together. The facility loop must become capable of supplying the rack with the required coolant chemistry, pressure, temperature, flow stability, filtration, isolation, monitoring, and serviceability. The rack-side components must tolerate the resulting operating conditions, and the distribution network must maintain acceptable pressure across all branches. A successful direct architecture therefore moves functions that previously sat inside the CDU into the wider cooling system or eliminates functions that no longer serve a purpose. The engineering exercise starts by listing every CDU function and asking whether the function remains necessary once the fluid paths become continuous.

Direct architecture changes the equation

The first recoverable quantity is the approach temperature associated with the removed exchanger. The second is the hydraulic pressure drop associated with both sides of that exchanger. The third is the electrical work consumed by the pump that maintains the isolated secondary circuit. Additional recovery may come from filters, valves, manifolds, controls, and other equipment that existed specifically to support the separated loop. The complete savings cannot be assumed from exchanger removal alone because the direct loop may require different distribution pressure or additional filtration elsewhere. A credible comparison therefore calculates the before-and-after pressure budget and temperature budget using the same rack heat load and the same environmental conditions.

The PUE effect then follows from the reduction in auxiliary electrical consumption relative to unchanged IT power. If the rack load remains constant while cooling-system electrical input declines, total site energy falls without any change to computing output. The corresponding PUE change can be represented algebraically as a reduction in non-IT power divided by the unchanged IT power, rather than by assigning a generic efficiency percentage to the architecture. That method matters because the actual PUE effect depends on what happens to the heat-rejection plant after the direct loop changes its operating temperature. If warmer water allows less mechanical refrigeration, the benefit can extend beyond pump savings, while a poorly matched direct loop could shift pressure or flow requirements elsewhere.

Turning recovered electrical work into operating economics

The economic calculation begins with the electrical power removed from the cooling architecture rather than with a claimed percentage improvement. Let the avoided auxiliary electrical load be represented as ΔP and the applicable electricity price as C, then the operating-cost reduction follows from ΔP multiplied by the number of operating hours and the applicable electricity price. This framework can be applied to pump power, refrigeration power, fan power, or the combined cooling load without mixing thermal capacity and electrical consumption. It also allows the analysis to use the site’s actual tariff structure instead of assuming a universal energy price. The resulting value can then be compared against the capital and maintenance consequences of changing the cooling architecture.

The same calculation can be expressed on a heat-removal basis when comparing different rack loads or operating periods. If the cooling system removes a thermal load Q while consuming auxiliary electrical power Pcool, then the relevant operating ratio becomes Pcool divided by Q. That ratio reveals how much electrical work the cooling architecture requires to transport and reject each unit of heat. A direct loop can improve the ratio by removing pump stages, reducing pressure loss, preserving temperature difference, or reducing mechanical refrigeration, but the result must come from measured or validated system conditions. This is why component-level efficiency claims cannot substitute for a site-level thermal and electrical balance.

Fewer Transfers, Truer Efficiency

The most important effect of removing an intermediate exchanger may not be the disappearance of the exchanger itself but the disappearance of an accounting blind spot. A conventional cooling diagram can make the CDU look like a single thermal interface even though it contains pumps, filters, controls, valves, heat-transfer surfaces, sensors, and often multiple hydraulic paths. Each of those elements has an electrical or thermal consequence, but the consequences can become distributed across different meters and control systems. Direct architecture brings more of the cooling work into one continuous hydraulic and thermal path. That makes the relationship between rack heat, coolant temperature, fluid flow, pressure drop, pump power, and heat rejection easier to observe.

Removing the CDU changes what gets measured

The measurement correction begins with a simple question: how much electricity does the complete cooling path consume to remove a defined heat load? The answer requires measurements at the rack, pumps, heat-rejection equipment, and major auxiliary components rather than relying on a CDU nameplate. Temperature measurements should establish the usable thermal rise and approach across every remaining exchanger, while pressure measurements should establish the head consumed by the hydraulic path. Electrical measurements should then capture the pump and heat-rejection loads associated with those operating conditions. The resulting balance makes it possible to identify whether the architecture spends electricity moving heat, changing its temperature, overcoming pressure loss, or rejecting it to the environment.

This also changes how efficiency improvements should be evaluated over time. A pump replacement may reduce electrical consumption while leaving the architecture’s thermal constraints unchanged, whereas removing an exchanger can alter both the hydraulic and thermal operating points. A control change can reduce flow while increasing approach temperature, producing an apparent pump saving that shifts the burden to the heat-rejection system. A higher coolant temperature can reduce refrigeration work while changing fluid properties and exchanger performance. The correct measurement therefore follows the entire system after every major architectural change rather than declaring success from a single equipment meter.

The real optimization target is the complete heat path

The direct-loop argument becomes strongest when it is framed around the complete heat path rather than around the removal of one component. Heat leaves the chip through its thermal interface, enters the coolant, travels through the rack distribution system, moves through the facility network, and finally reaches the heat-rejection system. Every intermediate device can change the temperature, pressure, or flow conditions that determine how much electrical work the next device requires. The objective is therefore to minimize the auxiliary energy needed to carry the heat from its source to its final sink while preserving all required operating constraints. A shorter thermal path can help because it removes opportunities for thermal resistance and hydraulic loss, but the engineering design still has to verify that the remaining components can handle the resulting duty.

The same principle applies to future high-density cooling systems because higher thermal loads make every unnecessary pressure and temperature penalty more consequential. Current research programs are explicitly pursuing cooling systems that reduce thermal resistance and pumping requirements while supporting much higher heat loads, showing that cooling efficiency increasingly depends on the complete thermal architecture rather than on refrigeration efficiency alone. Recent development work also places greater emphasis on validating both primary and secondary cooling systems under high-density operating conditions. That direction reinforces the need to evaluate the full hydraulic path, including the energy required to move coolant through interfaces that do not themselves remove heat from the computing hardware.

The site should pay for cooling once

A well-designed cooling system should move heat through as few unnecessary transformations as its operating requirements allow. That does not mean eliminating every exchanger, pump, filter, or control element because each may perform a necessary function under a specific architecture. It means recognizing that every component has both a thermal role and an energy cost, and that the energy cost can propagate beyond the component itself. A heat exchanger consumes temperature margin and pressure head, a pump consumes electricity to create that head, and the resulting temperature condition can determine how hard the heat-rejection system must work. The architecture becomes efficient when those relationships are designed together rather than optimized separately.

The resulting picture is not a case against CDUs but a case for counting everything they make necessary. A CDU can provide fluid isolation, temperature regulation, pressure control, chemistry management, and distribution functions that may be essential for a particular site, yet those functions should remain visible in the energy model rather than disappear behind a component boundary. When a direct loop can provide the same required conditions with fewer thermal and hydraulic steps, the saved electrical work becomes a measurable architectural benefit. The strongest analysis therefore compares complete thermal paths, complete pressure budgets, and complete electrical loads under the same rack conditions. That is how cooling stops being a collection of boxes on a schematic and becomes what it actually is: a continuous energy pathway from silicon to heat sink.