The Open Standard for AI Infrastructure.
OCP ORv3 — A 21-inch open rack architecture with 48VDC power, liquid cooling, and blind-mate backplanes. Built for the density, thermal load, and modularity that AI clusters demand.
Informational Reference
Open Compute Project
From 19 Inches to 21.
From Patch Cords to Bus Bars.
The Open Compute Project (OCP) ORv3 specification redefines the rack as an integrated system, not merely a frame. At 21 inches wide (533 mm), it accommodates wider compute sleds than the legacy 19-inch standard. A centralized 48VDC busbar eliminates traditional PSU redundancy at the server level. Liquid cooling manifolds run vertically alongside the power distribution. Blind-mate backplanes allow technicians to swap sleds without touching cables or coolant lines. The result: higher density, lower PUE, and component-level serviceability for AI training and inference workloads.
Key Architectural Principles
Wider Sleds, Tighter Density
The 21-inch rack width accommodates larger GPU and accelerator sleds without forcing non-standard chassis depths. This preserves front-to-rear airflow geometry while increasing compute-per-rack-unit beyond legacy 19-inch constraints.
One Power Plane. One Standard.
ORv3 replaces distributed PSUs with a vertical 48VDC busbar running the full rack height. Sleds draw power through blind-mate taps at the rear. This reduces copper weight, improves conversion efficiency, and standardizes power delivery across compute, storage, and networking sleds.
Heat Is a Rack-Level Problem
Vertical cooling manifolds distribute liquid to high-TDP GPU and accelerator sleds. The specification defines quick-disconnect blind-mate couplings that allow hot-swap maintenance without draining the rack loop. Air cooling handles the remainder; liquid handles the heat.
No Cables. No Contamination.
Network, storage, and management connectivity migrate to blind-mate interfaces at the sled rear. Sleds slide in and lock home — no patch cord management inside the rack, no exposed connectors in the hot aisle, and no technician time spent tracing cables during a swap.
Why It Matters for AI
Swap the Sled, Not the Rack
AI training clusters run 24/7. When a GPU sled fails, ORv3 allows a technician to pull the sled, disconnect power, cooling, and network simultaneously via the blind-mate rear plane, and insert a replacement — all without powering down neighboring nodes or re-dressing cables.
700W per GPU. And Rising.
Next-generation AI accelerators exceed 700W per device. Air cooling alone cannot sustain this density. ORv3's rack-level liquid cooling manifolds provide the thermal headroom required for 800G to 1.6T to 3.2T network transitions without throttling.
One Rack. Many Vendors.
Because ORv3 is an open specification, hyperscalers can source compute sleds, networking gear, and rack infrastructure from multiple vendors without integration guesswork. The mechanical, electrical, and thermal interfaces are standardized — reducing vendor lock-in and supply chain risk.
Where Optical Connectivity Fits.
Rear-Mounted Fiber Busbar
In an ORv3 rack, the optical backplane sits at the rear alongside the 48VDC busbar and liquid cooling distribution. This is a passive, blind-mate fiber busbar — not a switch — that provides physical-layer connectivity between GPU sleds, leaf switches, and spine uplinks. It moves fiber out of the front aisle, eliminates patch cord clutter, and aligns with the blind-mate serviceability philosophy.
Front-Face CPO Photonics
Co-Packaged Optics (CPO) and Near-Package Optics (NPO) place photonic engines at the front of the sled, close to the ASIC. The front face is reserved for ELS pluggable modules, loopback adapters, and thermal management — not for trunk cables. The rear busbar handles the heavy fiber count; the front handles the active optics.
Specification Governance.
Building on ORv3?
ADTEK supplies the rear-mounted fiber busbar, CPO cassette, and blind-mate optical connectivity layer that integrates with ORv3 and MGX rack architectures. Discuss your fiber backplane strategy with our engineering team.