NVIDIA Opens Vera Rubin Architecture to Open Compute Project: Open Rack & Compute-Tray Standards via MGX 1.2
NVIDIA announced it is opening key rack and compute-tray designs from its Vera Rubin platform to the Open Compute Project (OCP) under the MGX 1.2 specification, providing standardized liquid cooling, 5,000A busbars, and open-source DSX OS modules.
NVIDIA announced on October 13, 2026, that it is contributing core rack and compute-tray architectural specifications from its next-generation Vera Rubin computing platform to the Open Compute Project (OCP).
By publishing these designs through an updated NVIDIA MGX 1.2 specification, the company establishes an open standard for hyperscale data center builders, contract manufacturers, and server OEMs deploying high-density liquid-cooled AI infrastructure.
MGX 1.2 Technical Specifications and Open Hardware Designs
Historically, leading-edge accelerator platforms required proprietary chassis dimensions and custom plumbing layouts that locked operators into single-vendor mechanical tooling. The MGX 1.2 release standardizes physical form factors across multiple key dimensions:
| Infrastructure Component | Prior Generation (Blackwell MGX 1.0) | Vera Rubin MGX 1.2 Standard |
|---|---|---|
| Rack Width Standard | 21-inch OCP ORV3 Exclusive | Native 19-inch & 21-inch Compatible |
| Max Power per Rack | 120 kW to 140 kW | Up to 250 kW per Rack |
| Power Distribution | 48V DC / 1,400A Copper Busbar | Liquid-Cooled 5,000A DC Busbar |
| Cooling Connection | Flexible Manual Hose Disconnects | Blind-Mate Zero-Drip Hydraulic Manifolds |
| Tray Cabling | Internal Copper & Fiber Cable Harnesses | Zero-Cable Blind-Mate Backplane Interface |
| Thermal Dissipation | 80% Liquid / 20% Ambient Air | 100% Direct-to-Chip Liquid Cooling |
The elimination of internal hoses and fans within compute trays solves one of the highest maintenance failure points in production clusters. Data center technicians can swap an entire Vera Rubin compute tray containing two Vera CPUs and four Rubin GPUs in under four minutes without opening coolant loops or detaching cable bundles.
Electrical and Thermal Architecture
At 250 kilowatts per rack, conventional air conditioning and copper busbars fail due to resistive thermal loss. The MGX 1.2 specification integrates a liquid-cooled 5,000-ampere busbar running down the central spine of the cabinet:
Vera Rubin Open MGX 1.2 Rack Architecture
┌────────────────────────────────────────────────────────┐
│ 250 kW Standardized OCP Frame (Native 19" / 21") │
│ ├── Liquid-Cooled 5,000A DC Central Power Spine │
│ ├── Blind-Mate Manifold (Facility Water Return) │
│ └── Modular Compute Trays (No internal hoses or fans) │
└──────────────────────────┬─────────────────────────────┘
│
▼
┌────────────────────────────────────────────────────────┐
│ Open DSX OS Operating & Orchestration Layer │
│ ├── Config Manager (Firmware Baseline Automation) │
│ ├── Cluster Readiness Engine (Validation Tests) │
│ ├── Rack Management Service (Power / Thermal Telemetry)│
│ └── Cluster API Provider (Kubernetes Provisioning) │
└────────────────────────────────────────────────────────┘
Cooling fluid flows directly through the busbar housing before reaching server cold plates, ensuring that power distribution components remain well below 55 degrees Celsius even during continuous full-power matrix multiplication operations.
Open-Source DSX OS Cluster Software Modules
Hardware standardization provides little benefit if software provisioning tools remain proprietary. Alongside physical blueprints, NVIDIA open-sourced four critical software services under the Apache 2.0 license:
- Config Manager: Validates firmware revisions, microcode baselines, and PCIe configuration spaces across tens of thousands of nodes simultaneously.
- Cluster Readiness Engine: Runs synthetic stress tests prior to scheduling production jobs, identifying marginal optical transceivers or degrading memory channels before they cause training checkpoint failures.
- Rack Management Service: Exposes redfish-compliant REST APIs for real-time kilowatt draw, inlet coolant temperatures, and flow rates.
- Cluster API Provider: Enables cloud operators to provision bare-metal Rubin server nodes directly via native Kubernetes custom resource definitions (CRDs).
Global Ecosystem Adoption
Hyperscale operators moved swiftly to adopt the open MGX 1.2 blueprints:
- Meta Platforms: Confirmed it will build its next wave of frontier research clusters using OCP Vera Rubin compute trays paired with NVIDIA Spectrum-6 800 Gbps networking.
- Cisco Systems: Partnered with NVIDIA to publish the Ultra-High-Density Small Form-factor Pluggable (USFP) optical transceiver standard, cutting optical interface power by 28%.
- Google Cloud & Microsoft Azure: Committed to supporting MGX 1.2 compliant chassis in upcoming renewable-powered data center campuses beginning in late 2026.