The AI wave seen from the board’s side: three simultaneous pressures — signal integrity, power integrity and heat — and the fabrication capabilities a supplier needs to answer them.
The AI systems that dominate the technology conversation are usually described in terms of models, parameter counts and compute budgets. Underneath all of it sits a layer nobody photographs: the printed circuit board that carries every bit between the processors, memory and network interfaces. When that board cannot keep up, no amount of software sophistication compensates. This article looks at the AI wave from the board’s point of view, taking the three pressures it creates one at a time and ending with what a fabricator has to be able to do in response.
The first pressure is simply the amount of data that has to move. SerDes links inside AI servers and switches now run at 112 Gbps per lane, with 224 Gbps in development, and every one of those lanes is a signal that has to arrive with its shape intact after travelling across a large board. At these rates three things happen at once. Attenuation accumulates until the receiver can no longer resolve the eye. The dielectric constant of the substrate starts to determine whether the impedance along the line matches what was designed. And conductor loss rises as current crowds toward the surface of the copper, so the texture of the foil becomes a meaningful part of the budget rather than a detail.
The practical consequence is that material selection stops being a cost decision and becomes an engineering one. Laminates with a very low dissipation factor, a dielectric constant that holds its value across the band and low-profile copper foil are the entry condition for these designs, not an optimisation applied afterwards. A broader view of how these grades compare is available in our survey of laminates used in AI servers and high-speed switches.
The second pressure is invisible in a signal diagram. A modern accelerator draws hundreds of watts, and all of it arrives through the copper of the same board that carries the data. That creates a requirement that did not exist at lower power levels: the power delivery network has to supply large, rapidly changing currents without letting the resulting noise disturb the signals running alongside it.
Meeting it means thick copper power planes, low dielectric loss in the material so that the planes behave as intended, and a layout that keeps return paths short and predictable. The board stops being a simple carrier and becomes part of the power system. Any discussion of decoupling and plane design is ultimately bounded by what the laminate and the copper weight can deliver.
The third pressure is heat, and it is where the PCB’s role changes most visibly. Processors consuming hundreds of watts concentrate a large amount of energy in a small area, and it has to leave immediately. That has pushed board technology in several directions at once: substrates with higher thermal conductivity, metal cores and embedded heat pipes that move heat laterally, and thick copper planes that spread it. In these designs the board is a component of the cooling system, and its thermal design is as important as its electrical design. The interaction between the two is examined from a device perspective in our overview of the current state of high-frequency and high-speed PCBs in AI computing.
Those three pressures translate into a short list of capabilities. Without them, a design that simulates perfectly will not survive fabrication.
Impedance tolerance of ±5% or tighter on trace width, spacing and dielectric thickness is now normal for these designs. Holding it depends on imaging accuracy, etch control and lamination consistency, and the general approach is the same discipline described in our notes on the production process for microwave PCBs.
A package with thousands of pins needs thousands of vertical connections in a small area. Laser-drilled microvias around 50 µm handle the geometry; the difficult half is plating those deep, narrow holes uniformly enough that the resistance and reliability are the same in every one.
Every through-hole via leaves a stub of unused copper that resonates at particular frequencies like a small antenna. Removing it precisely has moved from a specialty service to a standard step, and it is one of the clearest signals of whether a shop actually serves this market.
Hot air solder levelling leaves a surface too uneven for these signals. Flatter finishes such as ENIG and immersion silver give a uniform surface for assembly and better behaviour at high frequency, and the differences between them are worth deciding early. The same considerations appear in our look at RO4350B laminates in AI system designs.
The final point is about sequence rather than capability. Thermal paths, surface finish, via strategy and material grade all have to be settled before fabrication starts, and each of them depends on what the shop can actually hold. A design team that works with the fabricator from the first layout, and simulates signal and power integrity against real material data rather than nominal figures, avoids the iteration cycle that otherwise appears as a second or third board spin. That is the difference between meeting the AI industry’s demands and discovering them one build at a time. A structured view of the same requirements, organised by specification rather than by pressure, is in our companion article on the extreme demands AI places on PCB manufacturing.
AI hardware asks the ordinary PCB to do three difficult things simultaneously: carry very fast signals without corrupting them, deliver very large currents without adding noise, and move a great deal of heat without failing. Every one of those requirements is decided by the material, the geometry and the process, long before the software that made them necessary runs for the first time. In that sense the board really is the foundation of the AI wave: quiet, unglamorous, and the thing that sets the limit on what everything above it can achieve.
Frequently Asked Questions
Because every bit and every watt passes through it. Processors can only run as fast as the interconnect allows, and the board's loss, impedance tolerance and thermal path set that ceiling before any software is involved.
Thermal management, because it changes the board's role rather than only its specification. Moving from a passive carrier to part of the cooling system brings metal cores, embedded heat pipes and thick copper planes into play, and those decisions constrain everything else.
They do. At 112 Gbps and above, attenuation and dielectric constant stability become the selection criteria, and copper foil roughness turns into a measurable part of the loss budget rather than a manufacturing detail.
Because the stub left by an unused via behaves as a resonant antenna and degrades the channel at specific frequencies. Removing it is the only practical fix, and whether a shop offers it reliably says a lot about its capability.
Because material grade, via strategy, thermal path and surface finish all depend on what the process can actually hold. Deciding them with capability data in hand avoids the extra board spins that otherwise appear as unexpected programme cost.
Share your data rates, power density and thermal constraints and we will tell you which laminate, stack-up and process window the design needs.
Get a Fast Quote Fast Reply on WhatsApp