At the 2026 OCP Global Summit, NVIDIA contributed its Vera Rubin rack and compute-tray designs to the Open Compute Project through MGX 1.2, the third generation of its MGX reference architecture. The headline numbers are a liquid-cooled 5,000 A busbar and racks of up to 250 kW. For connector engineers, the more important change is serviceability: NVIDIA is standardizing how a tray connects to power and cooling at the back of the rack.
This article separates what NVIDIA has published from our own engineering analysis, and it lists what is still missing from the public record.
In short:
- MGX 1.2 adds rack-scale blind-mate liquid cooling, a liquid-cooled 5,000 A busbar, fanless liquid-cooled compute trays and support for racks up to 250 kW.
- NVIDIA has not published the busbar voltage, coolant specifications, connector part numbers or where the full specification can be downloaded.
- Our analysis suggests the busbar is a 48 V to 54 V class bus, and that contact resistance and thermal design matter more at this current than almost any other connector parameter.
What NVIDIA Has Confirmed
From NVIDIA's announcement and the coverage that followed:
- What was contributed: the Vera Rubin rack and compute-tray designs, submitted through MGX 1.2 and OCP's Open Data Center initiative. NVIDIA also submitted a separate Vera CPU rack for agentic workloads such as tool calling and orchestration.
- Cooling: rack-scale blind-mate liquid cooling, with liquid-cooling technologies named MQD and MQDB. Our UQD and UQDB guide compares OCP's coupling specification with NVIDIA's MQDB family.
- Power: a liquid-cooled busbar rated at 5,000 A, and support for racks up to 250 kW.
- Trays: fanless, liquid-cooled compute trays with configurable I/O bays.
- Rack format: native 19-inch equipment support, standardized power and cooling interfaces and interoperable rack-scale interconnect infrastructure.
- Reference system: Tom's Hardware reports that a Vera Rubin NVL72 reference rack has 72 Rubin GPUs and 36 Vera CPUs and can draw more than 200 kW.
- Timing: NVIDIA says Vera Rubin is in full production with rollout in the second half of 2026. Its move to 800 VDC is planned to begin in 2027 and is not part of the MGX 1.2 specification itself.
- Partners: MGX system makers named include Dell, HPE, Lenovo, Supermicro, Asus, Foxconn, Gigabyte, Pegatron, QCT, Wistron and Wiwynn.
What Has Not Been Disclosed
- Busbar voltage: NVIDIA has not said what voltage the 5,000 A rating applies to, or whether it is the whole-rack total or a per-zone limit.
- Coolant details: flow rates, pressures and temperatures for MGX 1.2 are unpublished, and the announcement does not explain what MQD and MQDB stand for.
- Connector hardware: no part numbers, contact designs or mating-force figures have been given for the power, signal or fluid interfaces.
- Specification access: we could not find where the MGX 1.2 specification can be downloaded or when it was published.
- Cabling: some descriptions of the design call the trays cable-free. We found no such statement in NVIDIA's announcement or the coverage we reviewed, so we do not treat it as confirmed.
Our Analysis: What It Means for Connectors
The points below are our engineering assessment, not NVIDIA data.
1. What 5,000 A and 250 kW imply about voltage
Power is voltage times current, so 250 kW at 5,000 A works out to 50 V. NVIDIA has not stated the busbar voltage, but that arithmetic is consistent with a 48 V to 54 V class rack bus. It also fits NVIDIA's earlier description of 54 VDC in-rack components in current rack designs, as covered in our guide to the 800 VDC power delivery shift. This is an inference from the published figures, not a confirmed design detail.
2. Why contact resistance dominates at this current
Heat in a conductor is current squared times resistance. At 5,000 A, every 10 microohms of resistance in the current path dissipates 250 W (5,000 × 5,000 × 0.00001). That is why the busbar itself is liquid cooled, and why the blind-mate contacts between the busbar and each tray need very low and very stable resistance. No single tray carries the full 5,000 A, since the current is shared across many trays, but each tray's power contacts still carry large currents through a connector that is mated blind and re-mated many times over the life of the rack.
For the same reason, connector current ratings need to be read at the real operating temperature, not at room temperature, because a rating drops as the surrounding temperature rises.
3. Fanless trays change the thermal picture around connectors
Removing fans removes the airflow that used to cool the board, cage-mounted optics, memory and connector housings. In a fanless tray, heat has to reach the liquid through cold plates and conduction paths, so the thermal design near connectors matters more. This is our reading of what "fanless" implies, as NVIDIA has not published tray thermal details.
4. Several interfaces on one rear face
A blind-mate tray typically has to engage power contacts, signal connectors and liquid couplings in one insertion. These have different insertion forces and tolerance needs, so designers normally position the tray coarsely with guide features and let each connector float to absorb the remaining misalignment. For the open reference on the liquid side, see the OCP ORv3 Blind Mate Manifold Specification, which we summarize in our IcePack and IceCube article. Two cautions apply. NVIDIA has not said MGX 1.2 uses that exact hardware, and MGX 1.2 supports native 19-inch equipment while ORv3 racks use a 21-inch frame, so the two are not interchangeable.
5. Questions to ask when specifying parts for an MGX 1.2 class rack
Until the full specification is public, these are the questions we would put to a connector supplier:
- What is the contact resistance at the rated current, and how does it change after repeated mating cycles?
- What is the rated current at the real operating temperature, not the room-temperature rating?
- How much misalignment can each blind-mate interface absorb, and how does it float?
- Are the seals and plating compatible with the coolant in use, and what happens if a coupling leaks near a power contact?
- Is the part intended to be mated and un-mated live, or only with power off?
For the compute-rack side of the same OCP wave, see AMD Helion: what it means for connectors. For the optics that NVIDIA and Cisco announced alongside MGX 1.2, see our USFP explainer. For the 48V busbar contacts in current Open Rack designs, see ORv3 busbar power connectors. More coverage is in our OCP server interconnects hub.
Last updated October 15, 2026. We will update this article when NVIDIA publishes the MGX 1.2 specification.