NVIDIA GB300 NVL72 Infrastructure & Deployment_
GB300 NVL72 pairs Blackwell Ultra compute with Grace CPUs in NVIDIA's rack-scale architecture. Installation scope must identify the exact OEM equipment, customer-approved cabling plan and specialist-owned interfaces.
What Is the NVIDIA GB300 NVL72?_
The NVIDIA GB300 NVL72 combines 72 Blackwell Ultra GPUs and 36 Grace CPUs. NVIDIA describes more GPU memory and 1.5 times the dense FP4 Tensor Core FLOPS of Blackwell, rather than a universal application-speed increase. Use the current datasheet and OEM SKU for memory, power and performance specifications.
Leviathan can discuss an agreed physical installation and cabling package for customer-supplied equipment. The customer-appointed specialists retain facility electrical, all liquid-side and software commissioning responsibilities for NVIDIA's rack-scale architecture. Installation scope must identify the exact OEM equipment, customer-approved cabling plan and specialist-owned interfaces.
GB300 NVL72 Infrastructure Specifications_
| Specification | GB300 NVL72 |
|---|---|
| GPUs per Rack | 72 Blackwell Ultra B300 GPUs |
| CPUs per Rack | 36 Grace CPUs |
| Memory per GPU | 279 GB HBM3e in current Blackwell Ultra datasheet; confirm OEM SKU |
| Rack Power | Approximately 120 kW in DGX GB reference; verify OEM input |
| Interconnect | NVLink 5.0, 1.8 TB/s per GPU, 130 TB/s total |
| Compute Performance | ~1.1 exaFLOPS FP4 per rack |
| Weight | Exact OEM populated/shipping weight required |
| Cooling | Liquid-cooled compute; ancillary airflow still required |
| Power Architecture | Exact OEM rack input and approved facility design |
What Changes from GB200 to GB300?_
The NVIDIA DGX GB reference rack has 18 compute trays with four GPUs and two Grace CPUs per tray for both generations. GB300 adds Blackwell Ultra GPUs and ConnectX-8 networking. Do not assume an existing rack, manifold, power feed or service procedure is interchangeable; obtain the exact OEM compatibility and installation requirements.
NVIDIA currently lists approximately 20 TB of GPU memory for GB300 NVL72 versus 13.4 TB for GB200 NVL72. Available memory and workload fit depend on the exact product and software; these totals do not predict required rack count on their own.
NVIDIA publishes separate hybrid and future 800V DC infrastructure architectures. Those roadmaps do not prove that every GB300 rack accepts 800V DC. The exact OEM input and approved facility electrical design govern the project; electrical construction and energization remain outside Leviathan scope.
Confirm each OEM rack and facility cooling interface rather than carrying GB200 specifications over by assumption. Qualified customer-appointed mechanical/OEM teams own CDU sizing, liquid connections, pressure/leak tests and thermal commissioning. Leviathan offers no liquid-side work.
Scoping Physical GB300 Installation_
Obtain approved OEM equipment documents, rack elevations, access and facility readiness confirmation from the responsible teams.
Agree assembly, placement, cable handling and labeling; existing approved equipment/PDU cord work is separately scoped.
Customer-appointed specialists own every liquid-side, electrical-construction and software commissioning activity.
Document agreed visual, cable-ID, label, physical-connectivity and defect checks, port records and handoff; instrument testing requires specific confirmation.
Read consumption, design power and provisioning guidance as different fields
A GB300 power figure needs its product, document and quantity label. NVIDIA's shared DGX GB rack guide describes approximately 120 kW of rack consumption. HPE's GB300 NVL72 QuickSpecs separately labels 132 kW as nominal thermal design power and approximately 155 kW as electrical design power — peak. It also gives a separate 192 kW busway-provisioning suggestion. These are source-qualified descriptions, not competing answers to one unspecified maximum-power question.
The distinction matters before any arithmetic. A consumption description is not automatically a peak electrical requirement. A nominal design figure is not a measured operating load. A provisioning suggestion is not another name for the rack's stated peak. Do not silently replace one label with another or transfer a value from the shared DGX reference to a purchased HPE or other OEM configuration.
The comparison below records the public wording checked October 19, 2026. It does not adopt HPE's guidance as a site rule, reconcile every possible operating condition, prescribe an electrical design or establish a supported delivery configuration. The actual equipment supplier and qualified facility owner must identify the applicable configuration, input boundary, conditions and governing requirements. No electrical connection, capacity setting or work is authorized here.
| Source boundary | Published field | Interpretation to preserve |
|---|---|---|
| NVIDIA shared DGX GB hardware guide | Approximately 120 kW rack consumption. | A reference-rack consumption description; it is not labeled as this HPE system's electrical peak. |
| GB300 NVL72 by HPE QuickSpecs | 132 kW nominal TDP. | The named OEM's nominal thermal-design-power field, not measured workload consumption. |
| The same HPE document | Approximately 155 kW EDPp. | The separately named electrical-design-power peak; retain its actual product and source applicability. |
| The same HPE document's provisioning suggestion | 192 kW EDPp support for busways. | Separate OEM guidance, not a measured result or a requirement this page assigns to an actual facility. |
Primary fields: NVIDIA DGX GB hardware guide and NVIDIA GB300 NVL72 by HPE QuickSpecs. Only these labeled comparisons are adopted; power connectors, circuit design, provisioning procedure, liquid-side work and OEM service commitments are excluded.
Fictional EX-GP140 illustrates a source-reading error. A planning line for an unconfirmed OEM configuration copies the shared guide's 120 kW figure into a column titled Verified maximum input. The copied number has not acquired a new meaning from that column heading. A second line names the HPE offering but drops the distinctions between its nominal, peak and provisioning fields. Neither line is a verified requirement for the proposed equipment.
The correction is to restore the source and quantity labels, then obtain the applicable OEM return for the actual proposed configuration. Do not average the four values, choose the lowest one or treat their differences as available facility headroom. No rack, circuit or cooling calculation can be completed from this fictional source comparison.
Use the exact rack view and separate the acceptance decisions
The rack-scale unit also matters to layout. HPE calls its named GB300 offering a 48U rack-scale system; that is a supplied-system form factor, not 48U of unused capacity. NVIDIA's DGX GB hardware guide describes rear access to cable management, liquid manifolds, cable cartridges and the power bus bar. Those source descriptions explain why a drawing that shows only compute-tray positions is insufficient for a physical pathway review. They do not establish every OEM rack's geometry or any clearance.
Fictional EX-GV141 has a proposed elevation for one rack and a cable-path sketch drawn from a front view. Its intended remote switch endpoint is identified, but the sketch does not show the applicable rear service arrangement. The drawing is incomplete for that review; it is not evidence that the route is obstructed, suitable or unsafe. The appointed OEM and site owners must resolve the applicable view, supported arrangement and permitted pathway before dependent work is released.
| Return | Useful evidence | Decision it cannot replace |
|---|---|---|
| Layout and pathway basis | Exact OEM model/view, issued equipment positions, source applicability and returned pathway decision. | Actual clearance, load, cooling, electrical, handling or route-design approval. |
| Agreed physical IT handoff | Included task and observation scope, endpoint identities, checked revision, exceptions and the accepting party's recorded disposition. | Unincluded test results, OEM installation compliance or acceptance of the entire system. |
| Specialist and operating returns | Their separately applicable electrical, liquid-side, technical and software evidence and actual responsible decisions. | A physical cable record cannot supply these returns or authorize their work. |
The agreed physical scope can have a clear handoff while other interfaces remain unresolved. Record those open decisions instead of treating one document receipt as complete rack acceptance. This page provides source-reading and buyer-scope education; actual layout, cable-route design, installation methods, criteria and qualified acceptance still require the real configuration and project-specific evidence.
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