Why AI Servers Demand Multi‑Layer PCBs and Ultra‑Low‑Loss Materials

Server PCBs traditionally consist of power backplanes, NICs, mainboards, hard‑disk backplanes, and riser cards, with the mainboard occupying the largest area and value. Conventional server boards are typically 6, 8‑16, or 18+ layer rigid PCBs, emphasizing high layer count, density, and transmission speed.

As chip performance improves, cloud data processing demands faster signal transmission, prompting a gradual increase in PCIe 5.0 penetration and corresponding upgrades in PCB layer count, material quality, and overall value.

AI servers adopt a heterogeneous CPU + GPU architecture. Using Nvidia DGX H100 as an example, they require two new PCB types: OAM (OCP Accelerator Module) for GPU modules and UBB (Universal Baseboard) for multi‑GPU interconnect. An 8‑GPU H100 training server needs eight OAM cards mounted on a single UBB, forming an 8‑card mesh.

Signal‑rate requirements in AI servers push PCB specifications higher: CPU mainboards now need 14‑24 layers, GPU OAM and UBB demand 20‑30 layer high‑density interconnect (HDI) designs. Materials shift to Very Low Loss (M6) copper for CPU boards and GPU OAM, while GPU UBB may use Ultra Low Loss (M7) substrates.

These upgrades raise PCB costs dramatically; a single AI server’s PCB value can reach up to ¥10,000. Correspondingly, 400 G/800 G switches evolve to 30+ layer PCBs, further increasing value.

Power delivery also evolves: servers now use AC‑DC converters for grid‑to‑DC conversion and DC‑DC converters for fine‑grained internal voltage regulation. DC‑DC solutions handle higher currents and employ more complex PCB processes than AC‑DC, with companies such as Shennan Circuits, Willgo, and Shengyi Electronics providing relevant products.

Intel’s 2023 Eagle Stream platform and AMD’s Zen 4‑based servers both employ 16‑20 layer Very Low Loss PCBs, reflecting industry convergence on high‑performance materials.

Comparing Nvidia DGX H100 with the newer GB200, the latter eliminates the traditional 8‑GPU interconnect board, replacing it with a Superchip that integrates one Grace CPU and two Blackwell GPUs on a single large board, removing the need for a UBB and reducing PCB count while still requiring high‑grade materials for increased performance.