Investment ideas can be seen by dissecting the H100·Blackwell GPU substrates. NVIDIA GPU dies and HBM memory are already well-known "mainstream" themes. However, the hundreds of small components surrounding the GPU on the substrate, and the substrate itself that binds these components together, have not yet been sufficiently reflected in prices.
The core logic is simple. The B200's power consumption increased by 71% from 700W to 1,200W compared to the H100. This means that the power management components (MLCC, inductor, VRM) on the substrate must increase proportionally. The number of HBM stacks increased from 6 to 8, and the area of the CoWoS silicon interposer that binds them together also increased. The GB200 NVL72 rack requires a huge HGX baseboard to connect 72 GPUs, and this baseboard can only be manufactured with 30-40 layer ultra-high multilayer PCBs.
With the Blackwell Ultra (B300) and next-generation Rubin architecture announced, these three bottlenecks are expected to continue structurally for more than 2-3 years. As of February 2026, Murata (the world's No. 1 MLCC manufacturer) is scheduled to make a price increase decision at the end of March, so the short-term catalyst is also clear.