Power-Thermal Synergy: The Evolution of Integrated Cooling Infrastructure

• Author: Eric You ｜Advanced Mechanical Infrastructure

• Written & Interviewed by: LITEON Editorial Team (Corporate Brand Value Development Center)

• Technical review: LITEON Center of Core Competence

As next-generation AI accelerators boost thermal design power (TDP) from H100-class 700W envelopes to 1000W and beyond, the existing airflow paradigm becomes a bottleneck.

Yet, rack-scale hot-spot management gets tougher. Peak computing may be limited by thermal throttling, transient instability, and overbuilt room-level air conditioning without better heat-transfer efficiency. Since data center cooling accounts for up to 40% of energy usage, thermal design is one of the efficiency levers. Thus, AI data center cooling is turning toward liquid loops to transport heat closer to the silicon, reduce fan and chiller reliance, and drop PUE from air-cooled levels to more efficient ones.  ＊PUE = Power Usage Effectiveness = Total Facility Power ÷ IT Equipment Power

A 2024 global survey found that the average data center PUE was 1.56. So, many data centers are no longer seeing big efficiency improvements from traditional facility upgrades alone. They now need bigger changes in power architecture, cooling design, and system integration.

The bigger change is architectural. Since liquid cooling converts more watts into computation, operators improve node density, TCO, and ESG metrics, as well as reduce cooling overhead and recover rack space. But this only works when the supply chain stops buying cooling separately. Formerly, power shelves, racks, control logic, pipes, and cooling distribution units were sourced separately, causing compatibility issues and ownership confusion during deployment and service.

LITEON simplifies validation, accountability, maintenance planning, and lifecycle support in AI data center cooling by connecting long-term power R&D with rack-level liquid-cooling integration. It gives customers one coordinated path, including power delivery and heat removal.

Cooling Before Temperature Rises: Active Thermal Management Based on Power Signals

 From Temperature Feedback to Power Awareness

Traditional liquid loops are waiting for return-water or supply-water temperature to move before increasing pump speed. The controller is reacting to heat that has already entered the coolant path. The lag is not good enough as AI workloads soar. Power draw can spike near-instantaneously, while fluid temperature changes occur later, given the thermal inertia of the cooling loop. In other words, if the electrical event occurs first, AI data center cooling cannot rely on temperature feedback alone. So, advanced CDU designs are moving towards automatic flow and temperature adjustment based on actual operating conditions and not fixed mechanical assumptions.

Note that AI workloads can cause GPU power draw to spike for just a few milliseconds, sometimes forcing devices past the nominal TDP or even into absolute power limits. It makes a case for using power telemetry as an earlier control input than coolant temperature alone.

Turning Rack Current Into a Cooling Command

LITEON's approach is based on one simple but powerful idea. When the rack power system detects a fast load increase, that is an early warning of incoming heat. The control logic translates the change of current to a digital cooling command, without waiting for the sensors in the loop to confirm the increase. This allows the CDU to increase the flow rate and improve heat exchange before the coolant temperature changes significantly. This is the technical jump behind more responsive AI data center cooling solutions. The cooling system starts to behave as part of the power-control loop. LITEON has advocated system-level energy management for dynamic AI server loads, and increased flexibility for liquid-cooling and thermal-management systems.

Predictive Pump Control at the CDU Layer

The patented design, in principle, connects rack level load fluctuations with dynamic pump control. The control variable is a rapid increase of computing load, and the CDU takes hydraulic output prior to the temperature curve. This is important since smoother pre-emptive actuation reduces overshoot, avoids aggressive correction at the end, and makes pressure-flow behavior more stable during workload bursts. The outcome is a smart thermal chain with power telemetry, firmware logic, pump response, and heat-exchanger capacity all working in one sequence, as if the system is reading the next move of the workload before the coolant tells the story. This predictive control concept is reflected in LITEON’s patent application, which discloses a digital control method for adjusting CDU pump operation based on power variation. By using electrical load changes as an earlier control signal, the CDU can respond before the coolant-temperature curve fully reflects the thermal event. This strengthens the link between computing load, power telemetry, and hydraulic output, enabling more proactive thermal management for high-density AI workloads.

High-Availability Defense Line: Smooth Flow Compensation and Non-Stop Maintenance

Clustered CDUs as a Shared Hydraulic Reserve

Availability in high-density AI data center cooling is a function of how gracefully the hydraulic network degrades if one unit is degraded, not just how much heat it can remove at full capacity. LITEON's clustered CDU management is a flow pool of multiple units in coordination. If one CDU goes abnormal or is temporarily derated, the remaining units ramp up pump output in a controlled ramp, making up the missing capacity without a sudden dip in pressure. That software-defined compensation logic makes the failure handling a managed transition, reducing the chance that one weak node upsets the compute platform. LITEON has been bragging about the redundancy and online maintenance built into its high-capacity in-row CDU design.

Isolation That Protects the Water Path

The next line of defense is circuit isolation. Electronic check valves allow an interrupted cooling distribution unit to be disconnected from the active water path before reverse flow can cause turbulence, pressure imbalance, or unwanted mixing in the manifold. This is a small but important detail. This prevents the stopped branch from becoming a hydraulic shortcut, maintaining directional flow and stabilizing the rest of the pipeline. Operation continues with fewer transients, and service teams diagnose the affected unit. A common approach to protecting liquid-cooling performance for data center systems is backflow prevention by check-valve.

Maintenance Without Turning Cooling Into a Shutdown Event

Many designs are brought down by the fact that routine service is difficult, thus LITEON has moved the filter to the front side of the CDU and made it replaceable without stopping the entire cooling train. The system is combined with a 1+1 pump structure that allows pump path switching while keeping the water pressure and flow rate within the required operating band. Filter changes and pump transitions become controlled maintenance actions rather than disruption events. That is a reliability for operators that feels like a normal way of operating. Redundant pumps and smart flow monitoring are also core CDU reliability mechanisms. This serviceability feature is associated with LITEON’s invention disclosure IDF-23460, which addresses a CDU filter design that supports maintenance and replacement without interrupting overall cooling operation. By enabling front-side access and controlled service procedures, the design helps reduce maintenance-related downtime while keeping coolant flow, pressure stability, and system availability within the required operating range. This reinforces the CDU’s role not only as a cooling device, but as a maintainable infrastructure platform for high-density data center environments.

Architecture and Embedded Energy Storage: Redefining Cooling Continuity Through an 800V DC Link

800V DC Moves the CDU Closer to the Power Backbone

LITEON's next generation in-row CDU concept is a very direct approach.

• 800V high-voltage DC powers the unit.
• Removing unnecessary rectification stages.
• Reducing the number of internal voltage-conversion steps that must be completed before that energy can be provided to the pumps, controllers, and auxiliary subsystems.

Remember, each conversion stage loses power as heat. At hyperscale, 1% gain is not a rounding error, but it is a lever on long-term operating costs. 800V DC as a future AI-factory backbone can improve end-to-end efficiency and support racks scaling to megawatt-class demand.

High Voltage, Lower Current, More Usable Space

The electrical logic is simple. At the same power, higher voltage means lower current, and lower current means less copper, smaller conductors, less volume of distribution hardware, and less resistive loss. In rack rows, that implies leaner cable bundles, less congested rear service zones, and more physical room for airflow clearance, hose routing, inspection access, and cleaner power-cooling layout discipline. According to LITEON as well, its 800V DC solution is a means to reduce current, conversion losses, copper and cabling usage, deployment cost, and rack-space pressure, which fits into the infrastructure economics of AI data center cooling.

Embedded BBU Turns Cooling Into a Ride-Through Load

The more interesting step is to embed battery backup inside the architecture of the CDU itself. In this cooling distribution unit data center design, the CDU is not simply a thermal appliance awaiting the upstream power to recover. However, in the event of external AC grid failure, the internal BBU is capable of maintaining liquid circulation for the short but vital handoff until generators stabilize or backup distribution kicks in. Conceptually speaking, that patented approach protects high power racks from a dangerous thermal stall, since the coolant loop continues to flow while the electrical system re-arranges itself around the fault. This architecture is reflected in LITEON’s patent application US20250364936A1, titled “Cooling distribution device and method.” The application discloses an HVDC-powered CDU architecture with embedded BBU support, allowing the cooling system to maintain liquid circulation during power-transfer events. By integrating backup energy directly into the cooling distribution architecture, the design helps protect high-power AI racks from sudden thermal interruption when external power conditions change, supporting stronger ride-through capability and cooling continuity.

Built for Liquid Cooling: Hybrid Immersion Technology for Power and Cooling Coexistence

Dielectric Fluid as an Electrical Safety Layer

For high power density power systems, the coolant is part of the electrical safety architecture. Dielectric liquid is meant to withstand direct exposure to energized components with much lower short-circuit exposure. DI water can become dangerous once contaminants alter its conductivity profile. Therefore, LITEON is studying dielectric fluids with high insulation strength and non-conductive behavior. In AI data center cooling, then, a minor leak incident can be dealt with a larger safety margin in practice, rather than immediately leading to component damage or system risk. Along these lines, OCP immersion guidance also describes dielectric liquid as an electrically insulating liquid that can be in direct contact with electronics for heat removal.

Hybrid Immersion Inside the Power Train

A more forward looking concept is hybrid immersion for PSUs, whereby dielectric fluid can make direct contact to selected power components instead of forcing every thermal path through bulky cold-plate structures, interface materials, and mechanical clamps.

This leaves us with a denser packaging model:

• Less structural volume.
• Shorter heat paths.
• Better utilization inside the power module.
• More design freedom for higher-output, liquid-cooled PSUs.

This allows the power conversion hardware to be built as a liquid-native subsystem, rather than an air-cooled product with thermal accessories added on, for customers looking at AI data center cooling solutions in the future. LITEON emphasizes its capability to integrate power, liquid cooling, cabinets, software, and complete systems from the PSU to the cabinet level. This hybrid immersive power approach is linked to LITEON’s patent application US20260089899A1, titled “Liquid-cooled power supply cabinet and data center cooling system using the same.” The application frames liquid-cooled power hardware as part of an integrated data center cooling system, rather than as an isolated power module. By enabling closer interaction between power conversion hardware, dielectric liquid cooling, cabinet-level design, and system-level thermal management, this architecture supports a more compact and scalable path for future high-output, liquid-cooled AI infrastructure.

Reliability Has to Be Proven, Not Assumed

This is the moment when the story gets more important. Dielectric fluids, seals, hoses, manifolds, elastomers, and metallic interfaces must be robust to years of pressure and temperature cycling, exposure to chemicals, and handling during service. LITEON's long-life design benchmark, validated by long duration, high temperature, and high pressure validation, is designed to prove that sealing force, material elasticity, and fluid compatibility do not degrade under hyperscale operating conditions. Alongside U.S. university collaborations, that testing discipline gives LITEON a path to evolve hybrid immersion power infrastructure before next-generation compute loads make today's mechanical assumptions too conservative.

Moving Beyond Component Procurement to Unlock AI Computing Potential Through Power-Thermal Integration

The transition to liquid-cooling is not just replacing one thermal device with another. This is where power delivery, heat removal, firmware intelligence, rack mechanics, and service logic all come together as one infrastructure layer. This is also where LITEON's decades of power-management know-how become pertinent. By taking power-domain design into liquid cooling, and connecting load behavior and thermal response at the system level, LITEON can provide integration depth that traditional cooling-only suppliers may find difficult to replicate on their own. Its journey from Power to Liquid Cooling at the rack level allows customers to reduce interface risk, clarify ownership, and maintain reliability as architecture scales.

The time has come for teams who are preparing the next wave of accelerated computing to explore the integrated AI data center cooling solutions offered by LITEON. These solutions will provide a more dependable, scalable, and energy-efficient AI infrastructure.