NVIDIA Opens Vera Rubin Rack Design at OCP Summit, Vying to Set AI Data Center Standards
nashnova research
Nvidia unveiled open-source rack and tray blueprints for its next-gen Vera Rubin AI platform at the OCP Global Summit, covering power, liquid cooling, and networking specs — a move aimed at embedding its standards into the foundation layer of every AI data center.
What exactly did Nvidia open up, and why now?
Nvidia released server-rack and compute-tray design blueprints for its Vera Rubin platform, spanning power delivery, liquid cooling, and network connectivity.
VP Ian Buck announced the plan on stage; AI-infrastructure head Joe DeLaere said the real bottleneck is not chip performance but how fast AI factories can be built and deployed. Opening the specs lowers R&D barriers for server makers and shortens build cycles.
This means → Nvidia's influence is extending beyond the GPU itself into the power, cooling, and cabling layers of the data center. Whoever defines these specs controls the supply chain.
Why push rack voltage to 800 volts?
Nvidia proposed raising in-rack DC voltage from the traditional 48–54 V to 800 V. In plain terms = the old low voltage is like watering a large field through a thin hose — pressure too low, flow can't keep up. At 800 V the same power moves on far less current, so cables and connectors stop being the constraint.
The trigger: per-rack power draw has climbed to roughly 1 megawatt. At low voltage the resulting current is too high — a hard physical ceiling on expansion.
This means → 800 V DC is not an optional upgrade. It is the infrastructure shift AI data centers cannot avoid as power density keeps rising.
How big is the Vera Rubin NVL72 performance leap?
Ingrasys (a Foxconn subsidiary) showcased the Vera Rubin NVL72 platform: 72 GPUs + 36 CPUs, targeting large-scale model training and AI inference.
Ingrasys claims per-watt inference performance up to 10× the prior-generation Blackwell platform.
This means → if the number holds, the same power budget can handle an order-of-magnitude more inference work. Energy efficiency matters more than raw compute — power is the scarcest resource in a data center.
What problem does the Vera CPU rack solve?
Nvidia also released a dedicated rack spec built around its in-house Vera CPU — up to 256 Vera CPUs per rack, supporting 22,500 sandboxes (sandboxes: isolated environments that let different AI programs run simultaneously without interference).
The positioning differs from NVL72: NVL72 handles massive GPU training; the Vera CPU rack focuses on running and orchestrating multiple AI agents — a pure-CPU workload.
This reflects a deliberate product-line fork: training runs on GPU clusters, inference and agent orchestration run on CPU clusters, each with its own rack blueprint.
Is the supporting ecosystem keeping pace?
Foxconn Interconnect Technology (FIT) demonstrated a 1.6T eight-lane OSFP platform and a 12.8T ultra-dense XPO optical platform, addressing high-speed rack-to-rack interconnect.
FIT also launched a liquid-cooling busbar system rated up to 480 kW, designed to pair directly with Nvidia's high-power rack designs.
In plain terms = Nvidia is publishing the blueprints; Foxconn's component suppliers are putting the matching parts on the shelf at the same time. Standards and hardware arriving together sharply lowers the adoption barrier.
How large is this market, and how much can Nvidia lock in?
IDC projects OCP-standard equipment spending will reach roughly $800 billion by 2029 — 2.7× its forecast of $295 billion made just one year earlier. By 2030 the figure rises to $892 billion, about 5.8× the $154 billion actually spent in 2025.
This means → the market is expanding nearly sixfold in five years. Whoever's design specs become the industry baseline gains first-mover advantage across the supply chain.
Whether Nvidia can lock in that position during the standard-setting window depends on how fast ODMs and data-center operators actually adopt the blueprints — publishing them is the starting point, not the finish line.
市场有风险，内容仅供研究参考，不构成投资建议。