The Power Solution Has Changed - Efficiency Must Follow the Load

Author: Philip Zuk, SVP and GM AI/Datacenter BU, OmniOn Power Collection: 2026-OCP-EMEA-Summit-Executive-Sessions

Summary

Philip Zuk from OmniOn Power argued that the traditional data center power efficiency metric — peak efficiency at a single operating point — is insufficient for AI workloads, which are fundamentally dynamic. GPU compute activity during LLM inference fluctuates between 15–45% of rated GPU load, which (due to non-linear power mapping) translates to DC/DC converter operation between approximately 45–75% of rated capacity, with short peak excursions extending higher. The key implication is that a converter optimized for peak efficiency at a single point (the traditional "platinum/titanium" approach) will accumulate significant time-weighted losses across the actual operating envelope. OmniOn's design philosophy instead targets three simultaneous metrics: (1) peak efficiency, (2) peak power output at full load, and (3) flatness of the efficiency curve across the 25–100% active operating range.

The architecture context spans the transition from traditional 54V intermediate bus (ORv3) to next-generation 800V DC bus. The 800V path reduces copper losses, enables higher-efficiency power stages throughout the chain, and removes the per-rack 200 kW / 54V copper constraint that limits ORv3 deployments. Chain efficiency comparison: ORv3 chain total 83.8% vs. next-gen 800V chain total 83.8% nominally, but the 800V path removes the GPU VRM loss floor (-10.5 percentage points at final stage in ORv3) and provides 1.9 pp overall efficiency gain plus $250K savings per 100 MW DC deployment. OmniOn uses GaN in board-mounted power and SiC in rectifiers and DC/DC converter modules.

The Osprey product line (BMP — Board-Mounted Power) is OmniOn's commercial response: non-isolated DC/DC converters available at 1.3 kW, 1.6 kW, 2.0 kW, and 2.6 kW power levels, with a 3.5 kW variant in development. The 2.6 kW variant (QODN217A0B) achieves ≥98.5% peak efficiency at 51V input, 0.30% mid-range efficiency variation (25–75% load), and 4 kW peak power at 3ms burst. Competitive benchmarking shows OmniOn's 1.3 kW module achieves 50% better time-weighted efficiency flatness than the nearest competitor in the 45–75% GPU dynamic operating region, yielding approximately 25% additional OPEX reduction on top of peak-efficiency savings — equivalent to $250K–$400K/year savings per 1,000 racks at 132 kW/rack.