LITEON Technology's Bob Weng presented the company's GPU power solutions roadmap at OCP Global Summit 2025, tracing the evolution from 3 kW embedded server PSUs (2017) through 33 kW GB300 power shelves to a next-generation 1.2 MW 800 VDC power rack targeting early 2026 delivery. The presentation centered on LITEON's power smoothing innovation: bulk capacitor increases of ~40% combined with digital PFC firmware that screens GPU dynamic load transients from propagating to the AC facility side. LITEON positioned itself as a consistent supplier across three GPU generations simultaneously, from Hopper through Blackwell NVL72.

Key Takeaways

• LITEON has supplied GPU power since 2017 and is a key supplier to OpenAI; it is now delivering power solutions for three GPU generations (Hopper, GB200, GB300) concurrently.
• The 33 kW V2 power shelf for GB300 features ~40% more bulk capacitance than the GB200 design, smoothing input current during GPU load swings at AI workloads.
• The 110 kW three-phase power shelf achieves only 11% power density improvement over the 33 kW shelf but delivers more than 100% increase in energy storage per watt, enabling 10% input current step changes at large load steps.
• LITEON's next-generation 1.2 MW power rack (800 VDC output, 4-to-3 redundancy, 16 × 100 kW shelves) targets early 2026 for mature sample delivery.
• The 800 VDC architecture separates power from the IT rack into a sidecar power rack, reducing current distribution losses at megawatt-scale GPU infrastructure.
• Digital PFC control firmware can screen out all GPU transients — including the 150% overshoot seen at idle-to-full-load transitions — before they reach the AC input side.