The AI infrastructure narrative has fixated on GPUs, but the more durable bottleneck is quietly forming one layer down. Supply-chain players now see analog ICs and power semiconductors staying tight for years, with further price increases likely. This matters because every rack of accelerators depends on a dense web of power-delivery components that no one hoards or speculates on the way they do compute.
The global implication is a shift in where AI infrastructure risk actually lives. Power semiconductors are commodity-adjacent, low-margin parts that fabs have historically underinvested in relative to leading-edge logic. When demand steps up structurally, capacity cannot follow quickly, so the shortage behaves less like a cycle and more like a floor. That reprices the entire buildout: data center developers face higher bills of materials, longer lead times, and thinner ability to guarantee delivery dates to hyperscaler tenants. Expect this to widen the gap between operators with locked-in supply agreements and everyone bidding on the spot market.
There is also a strategic dimension for chip players themselves. As Groq pivots toward renting Nvidia-powered compute and Nvidia deepens its capital ties across the stack, the constraint on ancillary silicon becomes a hidden governor on how fast any of these ambitions can scale. Capital is abundant; power-conversion capacity is not.
For Japan, this is one of the rare AI stories where domestic suppliers sit on the advantaged side. Japanese firms such as Rohm, Fuji Electric, Mitsubishi Electric, Toshiba and Renesas hold genuine strength in power devices, including silicon carbide, and a sustained shortage converts that legacy position into pricing power and multi-year order visibility. The opening is real, but so is the risk of under-investing in capacity while the window is open.
For Japanese SIers and enterprise IT teams, the takeaway is procurement discipline. On-premise AI and edge deployments will feel component lead times directly, and RPA-style automation projects that assumed cheap, available hardware may need rescoped timelines. The teams that treat power-component sourcing as a first-class planning variable, rather than an afterthought behind GPU allocation, will ship AI infrastructure on schedule while competitors stall waiting on parts no one thought to worry about.
