The 48V rack bus is the new price bottleneck in AI data centers—SiC/GaN power silicon may capture the “dollars-per-watt” upside that electrical OEMs can’t

AI supply chains have spent the last year pricing grid upgrades and substation capacity. But rack-scale architectures are now forcing the next constraint: how efficiently and compactly you convert and regulate power once it’s inside the building.

Supermicro’s disclosed “DCBBS Blueprints” for the NVIDIA Vera Rubin NVL72 explicitly describe that each NVL72 rack includes four 110 kW power shelves with redundant 18.3 kW power supply units. That turns rack power into a modular, shelf-repeated design problem—where every percentage point of efficiency improvement and every watt saved becomes “real dollars per rack.”

The in-rack stack the market is underpricing usually looks like this:

• Rack entry conversion (often AC/DC with battery backup / ride-through features)
• Intermediate distribution bus (commonly moving to higher-voltage DC like 48V to cut current)
• Point-of-load regulation (VRMs and power-stage conversion)
• High-efficiency switching devices (SiC for higher-voltage/efficiency at high temperature; GaN in some high-frequency roles)

onsemi’s data-center positioning ties these changes to HVDC architecture choices and higher rack densities, noting the shift to 800V HVDC architecture for megawatt-scale AI infrastructure and the efficiency/power-density benefits of higher-voltage DC designs. In other words, the voltage level you choose upstream determines how many conversion losses you pay inside the rack—and what silicon types become viable at the power stage where VRMs and shelf electronics live.

This is also why the “dollars-per-watt war” starts looking less like a one-time design win (grid) and more like a compounding qualification treadmill (rack): higher-voltage architectures and denser shelves increase heat flux, which increases the value of switching devices that can maintain efficiency at higher operating temperatures and with tighter thermal constraints.

Eaton’s disclosed demand signal supports the idea that grid-side electrical build-out is active: in its reporting for Q2 2026, Eaton states strong year-over-year total backlog growth of 43% in the Electrical sector.

That backdrop matters because it funds the availability of high-power electrical infrastructure. But it doesn’t automatically monetize the in-rack conversion efficiency per watt—because the value capture depends on who supplies the shelf/VRM conversion electronics that must run efficiently at very high rack power densities.

Net takeaway: grid build-out enables deployments; rack conversion efficiency controls how much silicon you need to deliver a given compute-per-watt. That makes power silicon candidates—especially SiC and GaN—structurally better positioned to monetize the upgrade cycle than organizations whose product is primarily a one-time grid interface.

onsemi has been profitable and generates operating cash flow, but its reported financials also show semiconductor-cycle volatility. For investors focusing on rack-power silicon as the monetization layer, the key isn’t “whether power demand exists”—it’s whether silicon vendors can fund capacity and R&D through cycle bottoms.

From reported annual results: revenue fell from $8.253B (FY2023) to $7.082B (FY2024) and then to $5.995B (FY2025), while operating income declined from $2.538B (FY2023) to $1.768B (FY2024) and further to $0.748B (FY2025). That pattern is consistent with a broader semiconductor downcycle, yet operating cash flow remained positive at $1.978B (FY2025) and free cash flow stayed positive at $1.419B (FY2025).