A Unified Open Blueprint: Open Data Center Specification v0.7
At the OCP Global Summit in San Jose, NVIDIA, AMD, Google, and Meta announced a single open document for designing next-generation AI data centers — Open Data Center Specification v0.7. The document was submitted to the Open Compute Project (OCP) as an official contribution and published as version v0.7.0.
The specification covers power architecture, cooling infrastructure, and physical layout for large-scale compute facilities known as AI "factories." It standardizes rack height, weight limits, modular power density, and unified power and cooling limits. According to the developers, the document will let hardware manufacturers, colocation providers, and operators design, build, and deliver infrastructure without waiting for each customer's closed specifications.
According to NVIDIA, the specification is compatible with the company's DSX reference design — providing a direct path from the open standard to a working system. As the company's blog notes, once the standards are openly published, manufacturers worldwide can start building toward a common target. The specification was announced as v0.7.0 on the OCP's Open Data Center project page, where, alongside rack dimensions and weight limits, it sets requirements for modular power density.
Grid Outage Resilience: Grid Resilience Specification
The second major summit announcement was the OCP Grid Resilience Specification, jointly published by Google, Meta, Microsoft, and NVIDIA. It defines how large AI facilities should behave during utility grid outages: stay connected to the grid, respond automatically, and return to nearly full load in under two seconds.
Gigawatt-scale AI factories place enormous loads on the power grid, so their resilience to outages is becoming a critical part of infrastructure reliability. The new specification targets exactly this problem: facilities automatically adapt to grid faults and keep operating with virtually no interruption. According to NVIDIA, both specifications are compatible with the company's DSX reference design — offering a proven, deployment-ready implementation spanning facility layout and power architecture to software that monitors grid signals in real time.
Vera Rubin: Open Rack Design and Meta's Decision
Within OCP, NVIDIA contributed its Vera Rubin rack and compute tray designs to OCP through the MGX architecture. This gives system builders a common foundation for developing next-generation AI factory infrastructure. The updated MGX 1.2 specification includes fanless, liquid-cooled compute trays and standardized power and cooling interfaces, supporting up to 250 kW of power at the rack level.
Meta, in turn, announced it will use NVIDIA's Spectrum-6 Ethernet switches, delivering 102.4 Tb/s of throughput, in its Vera Rubin AI factories. According to NVIDIA, Spectrum-6 delivers twice the switching capacity of Spectrum-4 and twice the per-lane bandwidth, enabling a lossless Ethernet AI fabric at 200G lane speeds.
"We are pleased to work closely with NVIDIA to make multiplane networks operationally scalable and efficient — a key requirement both for building our own infrastructure and for our joint cloud partners," said Omar Baldonado, Meta's Senior Director of Data Centers and AI Networks (NVIDIA Blog).
Meta's Spectrum-6-based N-IceCube switch will be installed in the company's 100T liquid-cooled IcePack rack and run on Meta's open-source FBOSS network operating system. FBOSS is based on OCP's Switch Abstraction Interface standard, which allows managing switch hardware independently of the vendor.
USFP: A New Optical Standard from NVIDIA and Cisco
At the summit, NVIDIA and Cisco also announced plans to work on a new optical pluggable transceiver module form factor — USFP (Ultra-Dense Small Form Factor Pluggable). The new standard paves the way for 400Gb/s per-lane speeds at nearly the same size as the widely used 200Gb/s OSFP form factor.
USFP includes 16 full-duplex lanes, with plans to expand to 32 — delivering four times the bandwidth density of traditional 200Gb/s pluggable optics. The new form factor also provides roughly 1.3x higher energy efficiency, letting operators move more data with less energy.
To grow the open ecosystem, NVIDIA and Cisco are collaborating with TE Connectivity and Amphenol on housings, connectors, and modules. Under the plan, USFP will evolve as a community-led, multi-source initiative.
For more details, see the announcement on NVIDIA's official blog; for other AI infrastructure news, browse the aishunos.uz archive.
