Grid to chip: delivering power for megawatt-scale racks in AI data centers

A proposed solution

To free up valuable rack space for compute, a solution under consideration involves relocating the AC/DC converters outside the rack and increasing their output bus voltage — essentially, separating the legacy single rack with both compute and power shelves into separate power and server racks, then boosting the bus voltage to reduce the current and associated ohmic losses across longer distribution paths.

The Mt. Diablo specification being developed by hyperscalers under the Open Compute Project proposes increasing power distribution voltage to ±400V.

At ± 400V, the current is reduced by a factor of 16.7 compared to 48V for the same power delivery. This enables either a 16.7X longer distribution path with the same conductor size and losses, or smaller-gauge cabling. The latter results in significant copper cost savings, especially if AC/DC and backup modules are placed nearby in a dedicated “sidecar” power rack.

Key challenges

Traditionally, data centers distribute three-phase 415/480V AC directly to each compute rack, where it is converted to 48V DC. In the proposed model, racks receive ±400V. This introduces new safety and regulatory concerns due to high-voltage DC distribution. In a ±400V system, grounding the midpoint limits exposure to 400V, reducing insulation demands — but it also introduces two active rails, increasing system complexity.

Additionally, higher distribution voltages affect internal rack design. Power must be routed safely to each shelf, requiring upgraded connectors, cables, printed circuit board design, and isolated DC/DC converters that meet agency safety standards. This contradicts recent trends toward removing isolation stages.

Putting it all together

In an integrated system where AC/DC conversion, energy storage, and cooling hardware are consolidated into a sidecar rack, power system specialists like Flex are well positioned to deliver turnkey, efficiency-optimized solutions. This approach simplifies integration and eliminates the need for separate AC/DC power components within each server rack. At the same time, Flex can provide board-mounted DC/DC converters for compute blades, with high-voltage input versions already in development. Together, this represents a comprehensive and scalable grid-to-chip power delivery strategy optimized for the next generation of high-density, AI-driven data centers.