Rubin’s cableless architecture and ASIC high-layer HDI designs push PCBs to center of AI compute power

TrendForce’s latest research points out that AI server design is undergoing a fundamental structural shift. From NVIDIA’s Rubin platform featuring a fully cableless architecture, to hyperscalers’ in-house ASIC servers adopting ultra-high-layer HDI designs, PCBs are no longer merely passive circuit carriers—they are becoming a core enabler of compute performance. The PCB industry is officially entering an era defined by high frequency, high power, and high density.

Rubin-generation servers mark a shift with their cableless interconnect design. Unlike in the past, when high-speed connections between GPUs and switches relied on cables, they are now routed directly through multilayer PCBs such as switch trays, midplanes, and CX9/CPX boards. Consequently, signal integrity and transmission stability now take center stage in the design process.

Rubin upgrades its materials across the entire system to achieve lower loss and lower latency. The switch tray is built with M8U-class materials (Low-Dk2 + HVLP4) and a 24-layer HDI structure, while the midplane and CX9/CPX boards use M9-class materials (Q-glass + HVLP4) with up to 104 layers. 

This drives the PCB value per server to more than double that of the previous generation and shifts design emphasis from simple routing toward whole-system interconnect and thermal co-optimization. Rubin’s design philosophy has now become an industry template: AI servers such as Google’s TPU v7 and AWS Trainium 3 are likewise adopting high-layer HDI, low-Dk materials, and ultra-smooth copper foil.

Upstream materials reclaim bargaining power as the upgrade cycle begins
The performance requirements of AI servers are transforming the PCB materials landscape, with dielectric stability and thermal reliability now serving as key indicators. Glass-fiber fabrics and copper foil—once standard commodities—have become critical determinants of system-level performance.

Japan’s Nittobo has invested ¥15 billion to expand production of its T-glass, a material currently in structural shortage. Mass production is expected by the end of 2026 with output tripling current levels. T-glass offers low thermal expansion and high modulus, making it indispensable for ABF and BT substrates, and is priced several multiples higher than E-glass. Meanwhile, Q-glass and Low-Dk2 used in CCLs are emerging as next-generation materials due to their extremely low dielectric constants and reduced loss.

On the copper-foil side, as high-speed signaling and skin-effect constraints grow more stringent, ultra-low-roughness HVLP4 copper foil has emerged as the standard choice. However, upgrading to higher performance levels notably decreases yield—by approximately 50%—leading to ongoing supply shortages and a shift in bargaining power toward upstream suppliers. 

TrendForce predicts that 2026 will be a turning point when the value of PCBs is determined more by technological innovation than by production volume. For Taiwan’s supply chain, gaining dominance in advanced PCB materials and high-layer HD technology will be crucial to capitalize on the growth opportunities presented by the AI server super-cycle. 

Source: TrendForce, Taiwan.