Lifestyle
NVIDIA Opens Its Vera Rubin Rack Design at OCP and Unveils Open AI Factory Standards with Partners
At the OCP Global Summit, NVIDIA announced it is opening its next-generation Vera Rubin server rack design to the industry, and shared progress on networking, power and software with partners including Meta, Google and Cisco. This is data-center-level news that mainly affects equipment makers and data center operators; the content comes from NVIDIA's official blog.
About 5 min read
Key Points: What NVIDIA Announced at OCP
NVIDIA announced new collaborations and contributions at the Open Compute Project (OCP, an industry community promoting open data center hardware specifications) Global Summit, held at the McEnery Convention Center in San Jose, aiming to build a shared foundation for next-generation AI infrastructure. NVIDIA calls the large data centers that run AI workloads "AI factories."
According to NVIDIA, it is contributing the Vera Rubin rack (the frame that holds servers) and compute tray (the compute module inside the rack) designs to the OCP through the open NVIDIA MGX architecture, giving system makers a common base for developing next-generation AI factory infrastructure. NVIDIA says these results show its "extreme co-design" combined with open standards, which can speed up deployment, improve energy efficiency and raise tokens per watt (the AI output generated per unit of power).
Details on Racks, Software and Standards
NVIDIA says the MGX 1.2 specification adds rack-level blind-mate liquid cooling (equipment connects to the liquid cooling lines simply by being pushed into the rack), interoperable rack-level interconnect infrastructure, a liquid-cooled 5,000A busbar (the conductor bar that carries high current within the rack), and native 19-inch equipment support, and that the architecture supports rack deployments up to 250kW.
NVIDIA also says the Vera CPU rack design specification extends MGX beyond accelerated computing, and is designed to support tool calling, orchestration and sandboxed execution, so that AI agents (AI programs that can call tools on their own to complete tasks) can interact with applications, data and services in real time.
On software, NVIDIA says it has expanded DSX OS (the open operating layer of its AI factory platform) with open-source components covering network configuration (Config Manager), cluster validation (Cluster Readiness Engine), rack management (Rack Management Service) and bare-metal provisioning (Cluster API Provider, which handles installation and setup on physical servers that have no system installed yet). NVIDIA explains that Cluster API Provider can turn raw AI hardware into a production-ready cluster in hours rather than days, and that these components come from its experience running NVIDIA DGX Cloud.
On standards, NVIDIA says it published Open Data Center Specification v0.7 together with AMD, Google and Meta, covering power architecture, cooling and physical layout. There is also an OCP grid resilience specification published by Google, Meta, Microsoft and NVIDIA, which NVIDIA says requires large AI facilities to stay connected during grid disturbances, respond automatically, and recover to near full load in under two seconds.
Partner Progress at a Glance
| Partner | What NVIDIA describes | Key figures (NVIDIA's claims) |
|---|---|---|
| Visionbay | Announced plans for the K-1 AI factory platform, using an 800 VDC (800-volt direct current power) architecture combined with Vertiv power and cooling technology | 800 VDC can be adopted in stages |
| Meta | Adopting Spectrum-6 Ethernet switches; its own N-IceCube switch runs FBOSS (Meta's open-source network operating system) and is housed in the liquid-cooled IcePack rack | 102.4Tb/s; twice the switching capacity of Spectrum-4; Multiplane with up to eight network planes, scalable to 100,000+ GPUs |
| Cisco (with TE Connectivity and Amphenol) | Announced the USFP initiative, a new pluggable optical module (a transceiver connecting network equipment to optical fiber) specification | Package size about the same as OSFP; 16 duplex lanes, planned to expand to 32; 4x bandwidth density; about 1.3x energy efficiency |
| Falcon Transport will support future ConnectX-10 SuperNIC and BlueField DPU | Applications keep the standard RDMA (remote direct memory access) programming interface |
NVIDIA explains that Visionbay's K-1 is designed to introduce 800 VDC gradually where it brings the most benefit, rather than converting everything on day one. For the Multiplane technology Meta is adopting (which spreads GPU traffic across multiple independent network planes), NVIDIA says that when a component fails, traffic shifts to the remaining planes, so operation continues even at reduced bandwidth. NVIDIA also says Falcon is a protocol for reliably moving data over Ethernet, with multipath load balancing, programmable congestion control, hardware loss recovery and per-connection security.
What It Means for Ordinary Readers
This is data-center-level infrastructure news, not a consumer product launch. By NVIDIA's account, open standards let equipment makers, colocation providers and data center operators design and build to a common specification instead of waiting on individual customers' proprietary specs. If it works as described, it could make building AI facilities more efficient over time, but actual results remain to be seen once operators deploy it.
Frequently asked questions
What is OCP, and what did NVIDIA contribute?
OCP is the Open Compute Project, an industry community promoting open data center hardware specifications. NVIDIA says it contributed key Vera Rubin rack and compute tray designs to OCP through its open MGX architecture, as a shared foundation for system makers.
Whose claims are these figures and benefits?
This article draws only on NVIDIA's official blog, so all descriptions and figures should be treated as claims by NVIDIA and its partners.
Are USFP and Falcon support already on the market?
As NVIDIA puts it, USFP is a specification with an "announced plan," and Falcon support targets "future" ConnectX-10 and BlueField DPU products. NVIDIA has not given a launch timeline.
Will this directly affect ordinary consumers?
NVIDIA's announcement focuses on data centers and AI infrastructure and does not discuss consumer products.
Browse the latest news in this topic
Lifestyle
Intel Unveils Data Center AI Inference GPU "D-series" Specs and New Xeon Rack Designs, Targeting Enterprise AI Deployment
At the OCP Global Summit on October 12, 2026, Intel disclosed specifications for the Intel GPU D-series (formerly code-named Crescent Island) and new Xeon rack designs. Aimed at data centers and enterprises, Intel says they make AI inference easier and cheaper to deploy; all specs are Intel's claims.
Lifestyle
Meta Launches Muse Wonder and Muse Wonder Full: AI Models Built to Complete Real Tasks for Users
On October 12, 2026, Meta announced Muse Wonder and Muse Wonder Full, which it calls its most capable models yet, built for "agentic tasks" in which AI completes real work for users step by step. Developers can use Muse Wonder now, Muse Wonder Full is open to Muse Code subscribers, and Muse Wonder will reach Muse subscribers first. Meta also outlined its privacy and safety training.
Lifestyle
Google Cloud Previews TimesFM 3.0 in BigQuery: Multivariate Time-Series Forecasting in SQL
Google Cloud has announced a preview of the TimesFM 3.0 forecasting model in its BigQuery analytics service, letting data teams forecast multiple targets in one SQL query and factor in external drivers such as weather and holidays. This article summarizes what is new versus earlier versions.
Lifestyle
Google Cloud Launches Looker Data Apps Preview: Ask AI in Plain Language to Build Internal Data Tools
Google Cloud has announced Data Apps for its business intelligence tool Looker, now in preview. Google says users describe what they need in everyday language and an AI agent builds an interactive app that reuses the data definitions and permissions a company already has in Looker. This article summarizes Google's claims and what they mean for Looker customers.