NVIDIA Opens Vera Rubin Rack Designs to Standardize Gigawatt-Scale AI Factories
At the Open Compute Project (OCP) Global Summit in San Jose on October 13, 2026, NVIDIA contributed its next-generation Vera Rubin rack and compute tray designs to the open NVIDIA MGX architecture. This move provides system builders with a standardized physical foundation for next-generation AI factories, addressing the industry’s shift from megawatt to gigawatt-scale infrastructure while NVIDIA maintains tight control over the underlying compute and software stack.
What the NVIDIA MGX 1.2 Specification Adds to AI Factories
NVIDIA joined industry leaders to contribute to the third-generation MGX AI data center reference design through the OCP’s Open DC Initiative. The updated MGX 1.2 specification extends the open architecture with rack-scale blind-mate liquid cooling and a liquid-cooled 5,000A busbar designed for higher-power systems. The architecture supports rack deployments up to 250kW, simplifying serviceability, deployment, and infrastructure reuse.
Hardware partners are already adopting these physical standards. Compal Electronics is showcasing next-generation AI factory infrastructure featuring advanced power and cooling technologies at the summit, while MSI is displaying its 48RU rack based on the MGX architecture alongside liquid-cooled GPU servers.
"As AI infrastructure scales, customers need more than GPU performance alone. Cooling, networking, and system design all have to work together," said Danny Hsu, General Manager of MSI’s Enterprise Platform Solutions.
How the Rubin GPU and Vera Rubin NVL72 Compare to Blackwell
The Vera Rubin platform succeeds the Blackwell architecture, introducing a two-core chip design featuring the Rubin GPU and Vera CPU. The Rubin GPU employs two 3nm compute chips and achieves 50 PFLOPS of total computing power at FP4 precision, a significant increase from the 20 PFLOPS in Blackwell. It is equipped with 288GB of HBM4 memory.
At the rack scale, the DGX Vera Rubin NVL72 integrates 72 Rubin GPUs and 36 Vera CPUs behind NVLink 6. The following table breaks down the core specifications across the generations and configurations:
| Platform Configuration | Architecture Details | FP4 Compute Performance | Total Memory Capacity |
|---|---|---|---|
| Blackwell GPU | Previous Generation | 20 PFLOPS | N/A |
| Rubin GPU | Dual 3nm compute chips | 50 PFLOPS | 288GB HBM4 |
| Vera Rubin NVL72 | 72 Rubin GPUs, 36 Vera CPUs | Up to 3,600 PFLOPS (Inference) | 20.7 TB |
Additionally, the Vera CPU rack design extends the MGX architecture to support the agentic AI era, providing a purpose-built platform for tool calling, orchestration, and sandboxed execution.
Why NVIDIA is Opening Rack Mechanics While Keeping Compute Closed
The decision to open the physical rack designs while retaining proprietary control over the silicon and software is a strategic response to market dynamics. Competitors like AMD, Google, and AWS are actively promoting their own rack-scale alternatives, Remio.ai reported, including the AMD MI450X Helios Racks and Google’s TPU infrastructure. Openness has become a competitive requirement for data center builders.
By standardizing the physical infrastructure—such as the 250kW rack mechanics and liquid cooling interfaces—NVIDIA reduces construction friction for hyperscalers and enterprise operators. However, the central computing, NVLink 6 switching, and software technologies remain tightly coupled. Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA, noted that open infrastructure is a "critical ingredient for scaling AI factories worldwide." This approach ensures that while the physical shell is interoperable, the performance-critical components remain firmly within the NVIDIA ecosystem.
What the DSX AI Factory and Networking Updates Mean for Developers
Beyond hardware, NVIDIA launched new open source software components for the NVIDIA DSX AI Factory platform to unify computing, networking, power, and cooling into a single optimized system. The company also announced several networking advancements with major cloud providers. Meta and NVIDIA advanced open AI networking with Spectrum-X Ethernet SuperNIC innovations, while NVIDIA and Cisco announced the Ultra-Dense Small Form Factor Pluggable (USFP) standard for 400Gb/s-based networking fabrics. Furthermore, NVIDIA and Google announced Falcon Transport support on future ConnectX-10 SuperNICs.
For IT professionals and developers, the practical implication is clear: data center operators can now mix and match physical rack and cooling vendors to optimize capital expenditure and facility constraints. However, achieving the advertised tokens-per-watt efficiency and managing the DSX software stack still requires deploying NVIDIA’s proprietary networking silicon and accelerators. The open standard establishes a mechanical baseline, but the logical architecture remains a tightly integrated, vendor-specific stack.
The integration of the 5,000A busbar and 250kW rack capacity in the MGX 1.2 specification establishes a new physical ceiling for legacy data centers, forcing facility operators to upgrade their utility grid resilience to support the Vera Rubin generation.
