Liquid Cooling and PCB Thermal Design in AI Servers

Liquid cooling has moved from an optional upgrade to a default assumption in AI data centers during 2026. NVIDIA’s Vera Rubin platform has shifted entirely to liquid cooling, with company documentation stating that the Rubin generation of AI infrastructure operates with 100 percent liquid cooling and brings both compute and networking components into a closed loop system. Vera Rubin NVL72 extends the approach by combining liquid cooling with high density power delivery and a cable-free rack architecture. As rack level compute density and power consumption continue to rise, liquid cooling is no longer a facility level concern sitting outside the server. It has moved into the compute board, power board, switch board and the PCB structure itself.

From Electrical Carrier to Thermal Node

In the air-cooled era, the dominant PCB design concerns were signal integrity, power integrity and component placement. Heat was addressed with fans, heat sinks and chassis airflow, and the board’s role in thermal management was largely passive.

That changes once rack power enters the hundred kilowatt range. Air alone struggles to remove the heat generated, and the conduction path between the board and the cold plate becomes a direct factor in system performance. Heat generated by GPUs, CPUs, HBM stacks and power delivery modules must travel quickly through copper layers, thermal pads, thermal vias and structural elements to reach the coolant.

The design logic shifts accordingly. Localized heavy copper, large power planes and higher copper weight stop being purely electrical decisions and start serving as heat spreading structures. Stackup design moves from an exercise in electrical performance optimization to a three-way problem involving electrical, thermal and mechanical behavior simultaneously.

Vera Rubin magnifies this trend. The rack uses a liquid cooling manifold system and integrates power delivery, compute and high speed interconnect more tightly, so current density, component density and heat flux density on a single board all rise together. A board designed for that environment cannot be evaluated on signal performance alone.

High Layer Count and Heavy Copper Under Combined Load

AI server boards are moving toward 16 to 78 layer structures. Compute boards, switch boards, midplanes and backplanes each carry different high speed interconnect responsibilities, and at 112G, 224G and beyond, low loss materials from the M8 and M9 families must maintain stable dielectric constant and loss tangent while controlling trace roughness, via stub length and layer-to-layer skew.

Differential impedance control is tightening toward a plus or minus five percent window, which is a demanding target on a large panel with heavy copper present. Adding liquid cooling introduces further variables rather than reducing them.

Cold plate attachment applies localized mechanical stress to a large board. Thermal cycling continually changes the expansion state between the PCB, the BGA packages, the solder joints and the cold plate. If board warp, copper distribution and material coefficient of thermal expansion matching are not controlled, a board can pass electrical test and still develop solder joint fatigue or connection reliability problems after extended operation. Those failures appear in the field, long after the manufacturing data looked clean.

This is why high end AI PCB development is not simply a matter of adding layers. HDI, any-layer interconnect and mSAP line widths at 0.075 mm and below address interconnect density. Heavy copper and high power structures address supply and heat removal. High layer count low loss materials address high speed signal integrity. Liquid cooling pulls these previously separate technology tracks back into a single system engineering framework, and the manufacturing process has to satisfy all of them at once.

Designing the Thermal Path Deliberately

Because the board now participates in heat removal, the conduction path deserves explicit design rather than emerging as a side effect of routing.

Thermal vias beneath high power devices provide a vertical path from the component to internal copper planes, but their effectiveness depends on how they are arranged, how many are used and whether they are filled or plated through. An array that looks dense on a drawing may still have high thermal resistance if the copper it connects into is thin or discontinuous. Where a plane under a device is fragmented by signal routing, heat spreads poorly regardless of via count.

Copper balance across the stackup serves both thermal and mechanical goals. Even copper distribution reduces warp after lamination and gives heat a continuous path to follow. Asymmetric stackups may meet impedance targets while concentrating stress, which shows up later as bow or twist that interferes with cold plate contact.

Material selection also interacts with cooling. Higher thermal conductivity laminates conduct more heat but often differ in dielectric behavior, and the choice affects everything downstream. Reviewing these interactions during PCB design and layout is far cheaper than discovering them during thermal validation of an assembled rack.

What Liquid Cooling Changes in Manufacturing

For the manufacturer, the liquid cooling era raises the bar on process control rather than introducing one new specification.

Large panel handling and warp management become critical because a board that must mate with a cold plate has limited tolerance for deviation from flat. Copper thickness uniformity across the panel determines both current carrying consistency and thermal uniformity. Registration accuracy must hold through many lamination cycles on high layer count builds. Plating quality inside vias determines whether thermal and electrical paths behave as simulated.

Assembly adds its own requirements. High density SMT with large BGA packages, power modules and dense passive networks demands precise stencil design, placement accuracy and a reflow profile that suits very different package types on the same board. Because BGA joints are hidden, X-ray inspection is essential rather than supplementary. Thermal interface material application and cold plate mounting are process steps where variation translates directly into thermal performance differences between units.

Programs built around AI hardware PCBA manufacturing therefore benefit from having fabrication, assembly and inspection under one quality system. When a thermal anomaly appears during system validation, the ability to trace it back to a specific laminate lot, plating run or reflow profile determines how quickly it can be resolved. Under quality management systems that retain lot level records, that traceability is available by default rather than reconstructed after the fact.

The Same Requirements Are Spreading to Other Industries

The technology path being established in AI servers will not remain confined to data centers.

Central computing platforms in intelligent vehicles face the same combination of rising SoC compute and increasing power density, and 800 volt platforms push power system requirements higher still. Humanoid robot joint controllers must perform computation, motor drive and heat removal within a space measured in centimeters, where airflow is essentially unavailable. Electric aircraft motor controllers and avionics must combine light weight, high power and high reliability, which rules out bulky thermal solutions.

As a result, high heat flux PCB technology proven first in AI compute platforms is likely to migrate into intelligent vehicles, robotics and the low altitude economy. Heavy copper boards carry the high current. Flexible and rigid-flex circuits solve interconnect in confined spaces. Higher order HDI compresses control board area. High speed differential pairs carry data between sensors, compute nodes and communication modules.

Seen this way, liquid cooling is not an isolated cooling upgrade. It is pushing the PCB industry toward a new combination of capabilities in which high power, high speed, high density and thermal reliability stop being separate product line strengths and merge into a common requirement for advanced electronic systems.

From Design Review to Electro-Thermal Co-Engineering

What liquid cooling genuinely tests in a manufacturer is not whether heavy copper boards can be produced. It is whether thermal, mechanical and electrical coupling risks can be identified during design introduction, before tooling is committed.

Practical questions include whether the cold plate mounting position creates localized stress concentration, whether heavy copper regions will induce warp, whether the thermal gradient near a BGA exceeds what the package can tolerate, and whether adjusting copper structures around high speed differential pairs still holds impedance inside the plus or minus five percent window. Each of these spans more than one engineering discipline, which means the review cannot be conducted by a single function working in isolation.

Manufacturing support for this class of work spans high layer count fabrication, HDI, heavy copper power design, high frequency and high speed boards, and flexible or rigid-flex construction, combined with upfront manufacturability review covering differential impedance control, large panel warp management and stackup evaluation. Bringing assembly into the same process chain allows PCB fabrication capability to be validated against real assembly outcomes rather than theoretical assumptions.

This is also why competition in AI server boards is shifting from fabrication precision toward system manufacturing capability. The questions that will decide whether a supplier wins the next generation of programs are no longer just whether the process can reach 16 layers or 40 layers, or what minimum line width is achievable. They are whether the manufacturing organization understands signal behavior, power delivery, mechanical structure and heat removal together, and whether its PCB manufacturing process can hold all four under control at production volume.

What the Shift Means for Board Value

Starting with Vera Rubin, liquid cooling, cable-free rack architecture, high layer count boards, high speed optical interconnect and high density power delivery are being considered within a single rack level architecture. NVIDIA has stated that Rubin generation infrastructure uses 100 percent liquid cooling and extends that architecture toward higher density AI systems.

The industrial signal is clear. Performance competition in AI servers has expanded from the individual GPU to the rack as a system, and the PCB occupies the intersection of compute, power delivery, interconnect and heat removal at the same time.

For the PCB industry, the important change is not that a new environmental requirement was added. It is that the board has entered the thermal management system itself. As compute density continues to climb, the measure of a high end PCB will extend beyond layer count, material grade and line precision into integrated electrical, thermal and structural manufacturing capability. Suppliers who build that combination will be positioned for the next generation of AI hardware, and the same capabilities will carry directly into vehicle, robot and aircraft electronics as those markets follow the same trajectory.