NVIDIA has shared key Vera Rubin rack designs with the Open Compute Project. At the same time, the industry is developing 800-volt direct-current power systems for denser AI facilities. The two efforts meet at a practical question: how do you build and power an AI factory that equipment makers can actually deliver and operators can maintain?
Quick Take
- What opened: At the October 12 OCP Global Summit, NVIDIA said it contributed key Vera Rubin rack and compute-tray designs through its modular MGX architecture. It did not announce that Rubin GPU chip designs had become open source.
- What changes at the rack: MGX 1.2 describes reusable interfaces for liquid cooling, trays, interconnects and high-current power delivery. NVIDIA says the architecture supports rack deployments up to 250 kW.
- Why 800 V DC matters: At a given power, distributing electricity at a higher voltage lowers current on that portion of the path. The industry is developing shared DC interfaces to handle rising rack density and broaden the supply chain.
- The essential distinction: NVIDIA's 5,000 A rack busbar and an 800 V DC facility feed describe different elements of power delivery. One specification does not prove that every Vera Rubin installation uses the other.
What Did NVIDIA Actually Open for Vera Rubin?
In its October 12 announcement, NVIDIA said it contributed key Vera Rubin rack and compute-tray designs to the Open Compute Project (OCP) through NVIDIA MGX. The updated MGX 1.2 design gives manufacturers a common starting point for the surrounding physical infrastructure: trays, service access, cooling and power connections.
This is a notable choice for a company that sells a tightly integrated AI platform. A reusable rack design can allow more manufacturers to build compatible components and help operators expand capacity without commissioning every detail from scratch. The commercial logic is straightforward: supply chains and deployment speed increasingly determine how much AI computing a customer can bring online.
“Open” has a defined scope here. A contributed rack or tray specification is not a release of NVIDIA's chip intellectual property, NVLink design or the full production software stack under an open-source licence. Buyers should ask which interfaces are published, which implementations have been tested together, and which parts still depend on NVIDIA-specific hardware or contracts.
Why Design an AI Factory Around the Rack?
A Vera Rubin NVL72 is a connected system rather than 72 unrelated accelerator cards: NVIDIA describes 72 Rubin GPUs and 36 Vera CPUs working together at rack scale. Compute trays need dependable power, liquid cooling, networking and service procedures. If any one of those lags behind the silicon, it limits the useful capacity of the whole installation.
NVIDIA says MGX 1.2 adds blind-mate liquid-cooling interfaces, configurable I/O bays, fanless liquid-cooled compute trays, native support for 19-inch equipment and a liquid-cooled 5,000 A busbar. Its stated up-to-250-kW rack support is a design capability; it is not a measured consumption figure for every Rubin rack. The separate Vera CPU rack specification uses the same modular approach for CPU-heavy agent tasks such as orchestration and sandboxed execution.
A rack that can be assembled, tested and serviced through common physical interfaces is easier for suppliers to manufacture repeatedly. For operators, that could also improve repair times and future component reuse. Those outcomes still need proof in actual deployments: an interface specification alone cannot guarantee field reliability, spare-parts availability or support for equipment from competing suppliers.
Why Is 800 V DC Entering the AI Factory Conversation?
High-density racks need large amounts of electricity. The basic electrical relationship is power = voltage × current. For an illustrative 250 kW delivered at a single distribution point, the ideal current would be about 313 A at 800 V, versus about 5,208 A at 48 V. This is arithmetic for two hypothetical distribution voltages, not a wiring diagram for a Vera Rubin rack. At the same conductor resistance, current also affects resistive losses, which scale with current squared.
Higher-voltage DC can reduce the conductor and conversion burden along the parts of a system designed to use it. It does not mean accelerators themselves operate at 800 V: downstream equipment must provide the appropriate voltages. Protection against faults, safe isolation, compatible connectors, trained maintenance teams and cooling remain essential. Facility-level efficiency has to be measured across the whole power path, not inferred from the voltage alone.
The Open Compute Project says Google, Microsoft and NVIDIA are working with suppliers on shared 800 V DC interfaces and safety requirements. The same account describes two pathways: a side power rack converting existing 480 V AC near compute rows, and a longer-term design that converts medium-voltage AC into an 800 V DC facility backbone. OCP explicitly describes DC as an additional deployment option alongside existing AC infrastructure.
The Open Infrastructure Pieces: What Each One Covers
| Initiative | Where it applies | What has been shared or proposed | Limit to remember |
|---|---|---|---|
| MGX 1.2 | Racks and compute trays | Design interfaces for modular assembly, liquid cooling and high-density power | Open rack interfaces do not make the GPU or interconnect open source. |
| Open Data Center Specification v0.7 | Facility layout and design | OCP design specification authored by AMD, Google, Meta and NVIDIA | Version 0.7 is a published design specification, not proof that every site follows it. |
| 800 V DC collaboration | Power delivery across rows or facilities | Shared interface and conversion requirements being developed through OCP | Adoption depends on site power, protection, certifications and available equipment. |
| DSX software components | AI factory operations | NVIDIA introduced open-source tools for configuration, cluster readiness, rack management and provisioning | Individual components being open source does not describe the licensing of the entire stack. |
The 5,000 A busbar is a rack-level high-current feature described for MGX 1.2. 800 V DC describes an upstream distribution architecture. A deployment can convert voltage between those stages, so their headline numbers must not be multiplied together to claim a rack power rating. NVIDIA's published rack support figure is up to 250 kW, subject to the implemented design.
Is 800 V DC Already Running in Every Vera Rubin Data Center?
No such conclusion follows from the OCP announcement. NVIDIA cited Visionbay's planned K-1 AI factory design as an example built around 800 V DC with future Vera Rubin deployments in mind. That is a design plan; it should not be reported as evidence that all Rubin installations already use an 800 V DC backbone.
NVIDIA's power architecture roadmap describes a hybrid power rack for existing AC facilities, a row power center with availability expected in 2027, and a longer-term direct-conversion approach for new sites. Dates and equipment readiness can change. An operator choosing between these routes must check the electrical system already installed, usable floor space, certification and local grid conditions.
What Should AI Factory Operators Measure Before Committing?
Map grid connection, conversions, rack inputs, protection and measured losses at the expected operating load.
Test liquid connections, leak response, component replacement time and spare-part interoperability.
Compare accepted work and workload throughput per facility kWh, alongside latency, uptime and total cost.
Verify which rack, power and cooling parts have genuinely compatible alternatives and support contracts.
There is also a grid question. Better power distribution within a data center does not create additional generating capacity or remove the need for an approved connection. For European operators, site energy and environmental requirements belong in the facility plan. If the resulting AI services process personal data or are used in regulated settings, assess GDPR and applicable EU AI Act duties for those services; a particular rack or power design does not establish legal compliance.
MaGeN-AI View
The practical shift
The October announcement is about making the physical system around AI accelerators repeatable: rack structures, trays, coolant connections, power interfaces and operating tools. Open designs can give more suppliers a common target; 800 V DC offers a route to denser power distribution where the site and safety design support it. Neither development makes an AI factory instant, fully vendor-neutral or automatically efficient. The meaningful proof will be measured deployment time, serviceability, cost and useful AI output per unit of facility energy.
Frequently Asked Questions
What did NVIDIA open about Vera Rubin in October 2026?
NVIDIA said it contributed key Vera Rubin rack and compute-tray designs through its MGX architecture to the Open Compute Project. That does not mean the Rubin GPU design or the whole software stack became open source.
Does every Vera Rubin rack use 800 V DC?
No. MGX 1.2 rack contributions and 800 V DC facility distribution are related but distinct efforts. A site may use existing AC distribution, a hybrid side power rack, or a future DC-native design, subject to its equipment and engineering.
Why use 800 V DC for an AI factory?
At equal power and comparable distribution points, higher voltage means lower current. That can reduce conductor demands and resistive losses, though actual savings depend on conversions, protection systems, cooling and site design.
Is a 5,000 A busbar the same thing as an 800 V DC feed?
No. NVIDIA describes a liquid-cooled 5,000 A busbar as part of MGX 1.2 rack design. The 800 V DC work concerns an upstream power-distribution architecture. They need not carry the same voltage or current.
Will open rack standards make NVIDIA hardware vendor-neutral?
Open interfaces can broaden the supplier base for racks, cooling and power equipment. They do not automatically make accelerator chips, interconnects or software interchangeable; buyers still need to test compatibility and commercial terms.
