LEVIATHAN SYSTEMS

Platforms_

Vera Rubin NVL72 Readiness: Confirm the System and Its Interfaces

·Technical content·

Compare Vera Rubin NVL72 with GB300 NVL72, distinguish dated 800 VDC architecture announcements from a purchased system, and define the OEM and site inputs needed before physical work is released.

Key facts

  • NVIDIA describes both Vera Rubin NVL72 and GB300 NVL72 as 72-GPU, 36-CPU platforms. Their CPU, GPU, networking and NVLink generations differ; matching counts do not establish interchangeable equipment or facility interfaces. [N1, N2]
  • NVIDIA's August 11, 2026 account separates a hybrid power rack planned for the second half of 2026, a row power center expected in 2027, and a longer-term facility power block. [N3]
  • An architecture announcement, an OEM configuration and an approved site design answer different questions. Preserve the model, revision, applicable equipment boundary and decision owner when comparing them.
  • A survey brief needs the actual delivery configuration, facility interface documents, installed and shipping information, service views, and physical network endpoint requirements. Missing information remains an explicit input request.
  • A modular or cable-free tray description does not establish a universal aisle clearance, handling procedure, external cable list or permission to service a particular OEM system. [N1, N4]

Compare the named systems before carrying a plan forward

NVIDIA's current Vera Rubin NVL72 page identifies 72 Rubin GPUs and 36 Vera CPUs, ConnectX-9 networking, BlueField-4 DPUs and sixth-generation NVLink. Its GB300 NVL72 page identifies 72 Blackwell Ultra GPUs and 36 Grace CPUs, with ConnectX-8 and fifth-generation NVLink. The equal GPU and CPU counts describe populations, not the same components or a compatible installation package. [N1, N2]

For installation planning, carry forward the proposed workload and location as customer requirements, then ask which equipment and interface evidence must change. A prior GB300 port map is evidence about that GB300 configuration. It does not identify the external ports of a selected Rubin OEM build. A rack footprint or cooling connection drawn for one purchased system likewise needs an explicit applicability decision before reuse. The useful comparison is the complete system and its delivered interfaces, rather than a list of GPU performance figures.

Read 800 VDC claims with their date and architecture boundary

The source history matters. NVIDIA's May 20, 2025 technical article associated full-scale 800 VDC deployment with Kyber systems in 2027. That dated architecture discussion is not an electrical specification for every product called Rubin. [N5]

In its August 11, 2026 account, NVIDIA describes an MGX-compatible hybrid power rack for the second half of 2026: existing AC supply feeds the power rack, which delivers 800 VDC to compute racks. The later row power center uses overhead 800 VDC busway distribution and was expected in 2027; a facility-scale DC power block is a longer-term path. [N3]

Conversion changes AC supply to DC; distribution carries that DC to other equipment. An AC facility supply and an 800 VDC compute-rack supply can therefore describe different sides of a conversion unit. To reconcile those labels, identify the equipment between them and the interface each drawing describes. The announcement does not establish that a particular building can accept the proposed configuration.

An October 12, 2026 NVIDIA announcement also describes Visionbay's K-1 plans for an 800 VDC AI factory for future Vera Rubin deployments. A named partner's proposed architecture has a different boundary from an OEM's supplied rack and an operator's accepted power design. [N4]

For a proposed installation, obtain the selected OEM configuration's actual input-power document and the customer's corresponding facility design. Record whether each document describes the facility, conversion or distribution equipment, or the IT equipment, including separately supplied units. Keep voltage type, interface identification, revision and applicability together. An internal bus voltage or platform name cannot identify the connection the site must provide. The appointed OEM and qualified electrical owner must resolve any missing conversion or facility interface before the dependent work is released.

Give the physical survey a usable input set

A productive survey starts with a bounded question: can the customer's proposed location and project scope be assessed against this identified delivery? The survey input set should distinguish what is supplied, what the site must provide, and what remains undecided. Use the following groups to organize the buyer's input requests.

Equipment and delivery: the manufacturer, exact system or rack configuration, revision and included equipment population; which assemblies arrive integrated or separately; and the relevant shipping and installed drawings. A platform overview can establish terminology while the actual OEM package is still requested.

Facility interfaces: the applicable OEM power and cooling interface documents, their equipment boundaries, and the customer's corresponding facility drawings and responsible reviewers. Identify auxiliary equipment as its own population where it is part of the proposed project. Do not turn a GPU specification, reference rack figure or component rating into the complete installation demand.

Physical placement and access: the proposed location, the OEM's dimensional and mass information, shipping condition, mounting package, service views and removal envelopes, together with the site restrictions the appointed reviewers must assess. Keep the access review tied to the assembly and condition that will actually arrive.

Network and management: the selected equipment's external port identification, intended compute, storage and management roles, remote endpoints and the applicable physical connection list. Distinguish internal rack interconnects from connections crossing the delivery boundary. Request the missing endpoint rather than copying a prior generation's schedule.

For each group, identify who supplies the source, who decides applicability, and which part of the physical IT scope depends on that return. An unreturned drawing or unspecified interface remains unresolved in the survey brief, even when the platform name is known.

Ask the OEM for handling and maintenance inputs separately

NVIDIA describes Vera Rubin NVL72 as a third-generation MGX rack with modular, cable-free tray designs. Its October 2026 MGX contribution announcement describes rack-scale blind-mate liquid cooling and configurable I/O bays. These descriptions explain why the exact configuration and service boundary matter; they do not supply a universal field handling or maintenance procedure. [N1, N4]

For handling planning, request the named assembly's shipping condition and packaging information, applicable mass and dimensions, permitted handling information and the OEM's procedure for that particular delivery. An installed equipment drawing and a packed assembly can describe different objects. Keep those source identities distinct instead of assigning the rack's overall description to each separately delivered item.

For maintenance access, request the relevant service view, the component and removal envelope it describes, the side or direction to which it applies, and the prerequisites or restrictions that the OEM identifies. Have the responsible site and specialist owners resolve the proposed layout against those inputs. A convenient empty space in a plan is not evidence that a selected component can be removed there. The OEM's applicable procedure and appointed handling or service specialists determine the method and prerequisites.

Work through a generation change without copying the release

EX-RF131 is a fictional planning example. A customer's draft has two GB300 NVL72 positions, A01 and A02. The customer proposes replacing the A02 selection with Vera Rubin NVL72. The public platform counts would still sum to 144 GPUs and 72 CPUs across the two selected platforms. The CPU population, however, would change from 72 Grace CPUs to 36 Grace and 36 Vera CPUs. These are arithmetic comparisons of published platform populations, not a purchased bill of materials or an approved two-rack design. [N1, N2]

Four input issues now need distinct returns. First, the equipment selection record still calls A02 GB300: the customer can correct that identity to the proposed Rubin selection. Second, the physical network request still names the old platform's ConnectX-8 interfaces: the request must be revised to obtain the selected OEM Rubin endpoint mapping. Knowing the platform uses ConnectX-9 does not populate that mapping. Third, the power attachment is a 2025 roadmap image: the actual supplied power-interface document and matching facility disposition are missing. Fourth, the access attachment is a generic rack view: the selected OEM delivery and service-view package is missing.

The result is two identified corrections to the planning requests and two missing source packages. It is not four completed field tasks or evidence that either position is ready. The customer can continue collecting the missing inputs while preserving A01's separate evidence. Neither the unchanged total count nor an existing A01 decision automatically releases A02. The next useful output is a revised A02 evidence request with an owner for each missing return, while A01 retains its own configuration and decisions.

Freeze an applicable basis and keep later discoveries visible

A field team needs a specific released basis rather than a platform slogan. A useful freeze checklist identifies the selected OEM configuration and delivery population; the applicable drawing, port-map and interface revisions; the location and work-package boundaries; the customer's appointed decision owners; and the stated prerequisites, unresolved items and allowed scope. Include the handling and service documents that actually apply to the delivered assemblies.

For EX-RF131, a decision record might permit documentary preparation while leaving A02's physical installation unreleased pending the missing power and service packages. That is a scope decision for the example, not a universal release rule. When a later return changes the equipment selection, interface, location or dependency, identify the affected work and request a revised disposition from its responsible owner. A corrected document and an authorized field task remain separate records.

The actual project sequence and acceptance conditions depend on the selected equipment, site design, appointed specialists and agreed work package. Define the physical installation handoff and the separate facility, OEM and system-commissioning decisions it depends on.

Leviathan's Rack Assembly & Preparation service can be discussed for identified deliveries, equipment preparation, approved rack assembly and physical handoff records. Use the service link below to review that scope, then bring a nonconfidential outline to the project inquiry: the proposed OEM configuration, rack positions, supplied integration, affected interface records, missing returns, appointed equipment/material supplier, target schedule and required handoff. Those inputs allow a discussion of physical tasks, site requirements and availability; they do not establish authorization to work on a Rubin system.

Primary source scope and further reading

Sources were checked on October 20, 2026. N1 and N2 identify the named platform populations and technologies. N3 distinguishes the power-architecture paths and their announced time frames; N4 describes MGX features and the K-1 proposal; N5 supplies the earlier Kyber roadmap context and notes a July 31 branding update. Their links are provided below.

These public descriptions establish platform and architecture context. The selected OEM configuration and site documents must supply the actual installation requirements. The survey input groups and EX-RF131 are original buyer planning analysis, with the example's arithmetic derived from N1 and N2. The related GB200/GB300 comparison and site-survey guide expand the endpoint and survey questions; the service and inquiry links support a discussion of the defined physical package.

Frequently asked_

Can a GB300 installation plan be reused for Vera Rubin NVL72 because both have 72 GPUs?

The counts alone do not establish applicability. NVIDIA identifies different GPU, CPU, networking and NVLink generations. Compare the selected OEM delivery, actual external interfaces, service information and customer-approved site basis before deciding which parts of the earlier plan can be retained. [N1, N2]

Does Rubin mean that every site must deliver 800 VDC directly to the rack?

No. A platform name does not identify the site's connection. Reconcile the supplied equipment's input-power documents, conversion/distribution boundaries and qualified facility design; the dated architecture accounts do not replace that package. [N3, N4, N5]

What clearance or handling method should a survey assume?

Request the exact OEM assembly, shipping condition, dimensions and mass, handling information and relevant service/removal views. The appointed site and specialist owners must assess those inputs against the proposed route and location.

What can be frozen while an OEM interface is still missing?

The customer can identify the known equipment selection and source revisions while recording the missing input and its dependent work. Any limited release must say what it permits and who decided it. In EX-RF131, documentary preparation and physical installation have separate states; correcting a request does not resolve the missing power or service package.

Does Leviathan perform or guarantee the facility power and cooling validation described in a platform announcement?

No. Leviathan does not perform liquid-side work or licensed engineering; facility electrical work and full software commissioning are outside its standard physical IT service. Agree the actual tasks, resources and handoff separately from the facility and system acceptance decisions.

Discuss Your Infrastructure Project_

Tell us about your project, and our team will follow up to discuss the scope, requirements and availability.

Discuss a ProjectEmail Project Scope