NVIDIA used the OCP Global Summit in San Jose this week to open up the design of its Vera Rubin racks and launch new open source software for its DSX AI Factory. The company contributed the rack and compute tray designs through its open MGX architecture, and published new standards with hyperscalers for next-generation AI factories — covering everything from grid resilience to 800V DC power.
What NVIDIA actually gave away
The MGX 1.2 specification is the meat of it. It extends the open architecture with rack-scale blind-mate liquid cooling, interoperable rack interconnect infrastructure, and a liquid-cooled 5,000A busbar. It supports deployments up to 250kW per rack. NVIDIA also published a Vera CPU rack design specification, pushing MGX beyond accelerated computing into general-purpose rack design.
📊 Market Data Snapshot
On the software side, the new DSX open source components are meant to let operators manage AI factories without a proprietary stack. Ian Buck, NVIDIA's vice president of hyperscale and HPC, framed the contributions as a way to scale AI infrastructure globally — though the company didn't offer specific adoption targets.
Meta, Google, Cisco all get a line
Meta and NVIDIA advanced open AI networking with Spectrum-X Ethernet SuperNIC and Spectrum-6 innovations. Google and NVIDIA announced Falcon Transport support on future ConnectX-10 SuperNICs and BlueField DPUs. Cisco and NVIDIA introduced the Ultra-Dense Small Form Factor Pluggable standard — the industry's first module for 400Gb/s-based networking fabrics.
Visionbay, meanwhile, introduced an 800V DC AI factory design built for the Vera Rubin era. That's a notable detail. Higher-voltage DC distribution is how you cram more compute into a rack without melting the power delivery.
The power problem is now the design problem
The 5,000A busbar and the 250kW rack ceiling tell you where the bottlenecks are. AI factories are becoming power-constrained, and NVIDIA's answer is to standardize the electrical and cooling interfaces so more vendors can build against them. Liquid cooling is no longer optional at this density.
That has a second-order effect worth watching. Crypto miners sitting on energy contracts and cooling infrastructure suddenly have a skillset that maps onto AI hosting. Whether they pivot is an open question, but the hardware direction makes the overlap harder to ignore.
Why open standards matter here
NVIDIA isn't being charitable. By commoditizing the rack-scale design, it makes it harder for cloud providers to justify fully proprietary AI infrastructure — even as those same providers design custom silicon. The MGX ecosystem keeps them anchored to NVIDIA's interconnect and networking stack while giving them enough openness to feel unconstrained.
For decentralized compute networks, the practical effect is a lower barrier to building something compatible. Open rack designs and open transport protocols mean smaller operators can plug into the same physical layer as hyperscalers, at least in theory. That's a long way from tokenized GPU markets, but it's a prerequisite.
The MGX 1.2 spec and DSX software components are available now through OCP. The first partner systems based on the Vera Rubin rack design are expected to appear in the coming quarters — that's the next concrete milestone to watch.
