OCP Global Summit: Amid Ecosystem Adoption, NVIDIA Opens Vera Rubin Architecture to Help Scale AI Factories Worldwide

At the Open Compute Project Global Summit in San Jose, NVIDIA contributed Vera Rubin rack and compute tray designs to OCP through its MGX architecture, alongside a stack of partner announcements: open facility and grid standards with hyperscalers, new open source software for its DSX AI factory platform, a Meta networking deal, a new optics form factor with Cisco, and Google's Falcon transport on future NVIDIA network cards.

This matters because the bottleneck in AI has moved from chips to buildings. Racks are heading toward 250kW, busbars are carrying 5,000 amps, and data centers are being designed around 800-volt DC power. When those designs are published as common specs, equipment makers and colocation operators can start building without waiting on one customer's private blueprint. That can speed up capacity, and it can lower the cost of getting it right.

We think this is a genuinely useful move, and we also think it is a very smart business move. Both things are true. Opening the rack, the tray, the cooling interfaces and the facility blueprint makes it easier for the whole supply chain to build AI capacity faster. It also makes NVIDIA's architecture the default shape of that capacity. Open standards that happen to be tuned for one vendor's systems are still open, but they are not neutral.

The piece I find most important is the least flashy one: the Grid Resilience Specification from Google, Meta, Microsoft and NVIDIA. It defines how giant AI loads should behave when the utility grid stumbles, staying connected and recovering to near-full load in under two seconds. As these facilities scale into gigawatts, they stop being just customers of the grid and start being a force on it. Writing down how they should behave is overdue, and we want regulators and utilities reading it closely, not just engineers.

The Open Data Center Specification v0.7, published with AMD, Google and Meta, points the same way. Having AMD's name on a facility blueprint alongside NVIDIA's is a meaningful signal that at least the building layer may stay shared, even if the silicon inside is fiercely competitive. I'm watching whether that holds as the spec matures past a 0.7 draft.

The software side deserves credit too. Tools that stress-test GPU clusters before training starts, automate switch configuration and stand up bare metal in hours rather than days are the kind of unglamorous work that keeps expensive hardware from sitting idle or crashing mid-job. Open sourcing them lowers the bar for smaller cloud providers.

Now the downside, and where it leads. Every one of these standards is designed to make bigger AI factories easier to build. That means more electricity, more water and cooling infrastructure, and more concentration of compute in a handful of operators who can afford 100,000-GPU clusters. Efficiency per token improves, but total demand keeps climbing. If the industry gets very good at building these quickly, the hard questions shift to communities and grids that did not get a seat at the summit.

Our read: this is the infrastructure layer of AI growing up, and openness here is better than a patchwork of proprietary designs. But this is NVIDIA's own account of its own week, and the test is adoption. We will believe the ecosystem claim when independent builders ship racks to these specs and when the grid commitments show up in real interconnection agreements.

The Open Compute Project is the industry group, started by Facebook, where hyperscalers share data center hardware designs. NVIDIA's MGX is its modular reference architecture for servers and racks; the updated MGX 1.2 spec adds blind-mate liquid cooling, fanless liquid-cooled trays and native 19-inch equipment support. NVIDIA also described a Vera CPU rack aimed at agent workloads like tool calling, orchestration and sandboxed execution, separate from the Vera Rubin NVL72 systems used for large-scale training and inference.