A two-second rule for AI factories
Google, Meta, Microsoft and Nvidia published an OCP Grid Resilience Specification at the Open Compute Project’s global summit in San Jose on Monday. According to Nvidia’s account of the announcements, it defines how an AI facility should behave when the grid around it is disturbed: stay connected rather than disconnect, respond automatically, and return to near-full load in under two seconds.
That is a number aimed at a specific complaint. A gigawatt-class training cluster is a load that can swing by hundreds of megawatts in seconds, and utilities in several markets have spent the past year arguing that such loads are a reliability problem rather than just a demand one. A written specification saying the facility rides through the disturbance, and comes back fast and automatically, is the industry answering that in its own words. It is a design target, not a measurement from any deployed site.
Alongside it the same group — minus Microsoft, plus AMD — published an Open Data Center Specification v0.7, which Nvidia describes as a single blueprint for AI factory facility design covering power architecture, cooling and physical layout.
Vera Rubin goes into the open spec
Nvidia also contributed its Vera Rubin rack and compute tray designs to OCP through the MGX architecture, as the third-generation MGX 1.2 reference design under the project’s Open DC Initiative. The spec covers rack-scale blind-mate liquid cooling, a liquid-cooled 5,000-amp busbar, native 19-inch equipment support, configurable I/O bays, fanless liquid-cooled compute trays and standardised power and cooling interfaces, with rack deployments supported up to 250kW. It also carries a Vera CPU rack specification aimed at tool calling, orchestration and sandboxed execution for agentic workloads.
Meta is the clearest adopter. It is taking Nvidia’s Spectrum-6 Ethernet switches, at 102.4Tb/s, into its Vera Rubin build-out: its N-IceCube switch is built on Spectrum-6, runs Meta’s open FBOSS network operating system and fits the company’s liquid-cooled IcePack rack. Spectrum-6 doubles both the switching capacity and the per-lane bandwidth of Spectrum-4, at 200G a lane. Paired with ConnectX-9 SuperNICs, Nvidia says Spectrum-X Multiplane spreads GPU traffic across up to eight independent planes and supports clusters beyond 100,000 GPUs, shifting traffic to the remaining planes when a component fails.
“We are excited about our close partnership with NVIDIA in making multiplane fabrics operationally scalable and performant,” said Omar Baldonado, senior director of data center and AI networking at Meta, in Nvidia’s post.
A new optical module, with no ship date
Nvidia and Cisco used the same day to announce USFP, an ultra-dense small form factor pluggable optical module. Their press release describes a module built for 400Gb/s-per-lane electrical signalling with capacity up to 12.8Tb/s, packing 16 full-duplex lanes into an OSFP-sized footprint and expandable to 32 or more later. The companies claim about four times the bandwidth density and about 1.3 times the power efficiency of 200Gb/s-per-lane modules, and support for up to 100 watts of dissipation without a dedicated embedded cooling circuit. A two-dimensional array connector with two-step insertion is meant to reduce contact wear.
TE Connectivity and Amphenol are building cages, connectors and modules; the parts are designed to work with Nvidia Spectrum-X and Cisco Silicon One Ethernet. The release gives no ship date, calls the modules planned, and says an open specification will follow through a community-led multi-source project.
Two further items rounded out the day. Nvidia said it will support Google’s Falcon transport protocol on future ConnectX-10 SuperNICs and BlueField DPUs, letting network cards interoperate over a common transport while keeping standard RDMA interfaces. And it open-sourced four components of its DSX AI Factory software — a switch config manager, a GPU cluster readiness engine, a rack management service and a Cluster API provider for bare-metal provisioning.
What is worth watching is the grid specification rather than the silicon. Nvidia’s rack designs will be implemented because Nvidia sells the parts. A two-second recovery target only means something once a utility writes it into an interconnection agreement.
