OCP Rack & Power Project Call (Feb 11, 2026)

Why It Matters

The 160 kW shelf enables denser AI compute deployments, but its success hinges on careful power‑distribution engineering to ensure reliability and efficiency at scale.

Key Takeaways

• Meta unveils 160 kW power shelf for AI GPU workloads.
• Shelf uses eleven 18 kW PSUs with 98.5% peak efficiency.
• Supports up to 400 kW per rack, scalable to 800 kW.
• Voltage drop limited to ~1.5 V under worst‑case load.
• Monitoring via PMM2 module and active current‑sharing daisy chain.

Summary

The OCP Rack & Power Project call on February 11, 2026 featured Meta’s rollout of a 160 kW power shelf designed for high‑density AI GPU racks, followed by Oracle’s briefing on utility‑provider fault‑condition requirements for data centers.

The shelf integrates eleven 18 kW power‑supply units, delivering 98.5% peak efficiency and 97.5% at full load. It operates from a +‑400 V DC bus supplied by the HPRV4 rack, converting to 50 V DC on the shelf, and can support up to 400 kW per rack—potentially 800 kW with specific configurations. The design includes a 4,000‑amp bolted BK‑5000 connector, active current‑sharing via RJ45 daisy‑chain, and a PMM2 module for real‑time monitoring.

During the Q&A, engineers highlighted a worst‑case voltage drop of roughly 1.5 V across the bus bar and a 6 ms hold‑up time dictated by upstream HPRV4 specs. Concerns were raised about return‑current paths through rack steel and cable shields, prompting references to prior RV3 testing that demonstrated manageable voltage differentials when proper return‑path sizing is applied.

The announcement signals a significant step toward accommodating the escalating power demands of AI workloads, offering data‑center operators a scalable, high‑efficiency solution while underscoring the need for meticulous electrical design to mitigate voltage‑drop and grounding issues.