Compare GB200 and GB300 NVL72 offers as identified rack systems with complete bills of materials, facility requirements and acceptance obligations. Confirm the OEM implementation, compute and interconnect configuration, power and cooling boundaries, software and support. A GPU count or headline rack price cannot establish that the offered system is complete or deployable.
An integrated rack purchase connects compute procurement with electrical, mechanical and operational decisions. The seller's equipment boundary and the facility team's delivery boundary need to meet. If they do not, important components or commissioning work can remain unassigned despite an apparently complete quotation.
Use SecondWatt's GPU catalogue, GB200 NVL72 dossier and GB300 NVL72 dossier to identify the platform. Then request the actual rack configuration and the documents that support its deployment.
Key Takeaways
- Specify the exact OEM rack implementation and complete supplied configuration.
- Distinguish compute trays, internal interconnect and external cluster networking.
- Keep operating demand, peak demand and electrical provisioning separately labeled.
- Confirm which cooling equipment, manifolds and facility connections are included.
- Agree site readiness, system tests and support responsibilities before award.
Identify the platform and the purchase boundary#
NVIDIA describes GB200 NVL72 and GB300 NVL72 as liquid-cooled rack-scale systems. Their common NVL72 designation is useful context, but it does not establish that every OEM offer has identical external interfaces or commercial inclusions.
Platform identity at the reference-design level#
| Reference platform | CPU/GPU arrangement described by NVIDIA | Procurement distinction |
|---|---|---|
| GB200 NVL72 | 36 Grace CPUs and 72 Blackwell GPUs | Identify the actual GB200 OEM rack and supplied scope |
| GB300 NVL72 | 36 Grace CPUs and 72 Blackwell Ultra GPUs | Identify the actual GB300 OEM rack and supplied scope |
The table identifies platforms; it is not a benchmark comparison or a promise of availability. Ask for the OEM system model, configuration revision and bill of materials. Record whether the offer is for an assembled rack, a set of components, an upgrade scope or equipment requiring additional integration.
Do not treat a compute tray as a fractionally complete rack without examining the rest of the system. A tray-level quotation may be valid for a buyer replacing components within an existing supported installation. It answers a different question from a request for a deployable rack system.
For a used or surplus offer, request equipment identity, ownership evidence, condition history and the seller's authority to supply the configuration. Keep proposed repairs or replacements separate from hardware already present. The final accepted system should be traceable to the inventory agreed before shipment.
Reconcile the complete bill of materials#
Create one schedule covering the rack, compute, internal interconnect, power and cooling components. Assign each item an inclusion status and a responsible supplier. Use exact part or assembly identities where they are needed to establish compatibility.
NVL72 rack offer scope worksheet#
| Scope element | What to establish in the offer |
|---|---|
| Rack and mechanical assembly | Identified rack, mounting and service configuration |
| Compute trays | Quantity, CPU/GPU configuration and included storage |
| Internal interconnect | Switch trays, required cables and complete topology |
| Power components | Shelves, busbars, whips and required controls |
| Rack cooling | Manifolds, hoses, connectors and included heat-removal equipment |
| External cooling | CDU, primary-loop interfaces and exclusion boundary |
| External networking | Switches, optics, cables and cluster connection scope |
| Management and software | Required systems, rights and configuration work |
| Services | Assembly, installation, testing and support obligations |
Do not allow included cooling to remain undefined. It might mean cold plates and rack manifolds while excluding the CDU and facility connection. Similarly, included networking might cover management switches but leave the workload fabric to the buyer.
Ask the seller to identify any parts that arrive separately or require field installation. Record whether the quoted completion milestone includes those tasks. A rack delivered to the building can still be incomplete against the operational scope if essential parts or services follow later.
Where alternatives are proposed, obtain a revised schedule rather than accepting substitutions through informal correspondence. This is especially useful when storage, network or power configurations change during negotiation. The commercial total and acceptance plan should refer to the same revision of the bill of materials.
Separate power demand from provisioning#
The facility team needs the OEM's electrical planning information for the exact rack. Distinguish thermal design power, expected operating demand, electrical peak and distribution provisioning. These figures serve different purposes and should remain labeled throughout the budget and design review.
HPE's GB300 NVL72 QuickSpecs, version 3 dated September 8, 2026, provide a concrete example of these distinctions. The following figures describe that HPE implementation; they are not universal requirements for every NVL72 rack.
HPE GB300 NVL72 power fields as published#
| Published field | HPE figure | How to use the field |
|---|---|---|
| Nominal thermal design power | 132 kW | Preserve the nominal TDP label |
| Electrical design power, peak | Approximately 155 kW | Review the stated peak with the facility engineer |
| Suggested busway provisioning | 192 kW | Treat as HPE's provisioning guidance for this implementation |
Do not turn the first row into a guaranteed average workload demand or the third into a measured rack consumption figure. Obtain the applicable design conditions, distribution arrangement and response to supply loss from the rack vendor before finalizing the facility interface.
NVIDIA's GB300 power-smoothing explanation describes platform features intended to manage power variation. Those features do not establish an external UPS runtime or eliminate the need for a project-specific electrical review. Request the behavior relevant to the actual supplied configuration.
Use the UPS procurement guide to organize the external power-support scope. Any UPS, distribution or generator selection must follow the complete system requirement and intended operating arrangement, rather than one unlabeled rack number.
Match the cooling system to the actual rack#
Obtain the rack vendor's coolant, temperature, flow, pressure and water-quality requirements. Separate the server-side loop from the facility-side heat-rejection system. Identify which party owns each connection and which operating conditions that party must deliver.
QCT's GB200 NVL72 cooling paper describes a vendor-specific cooling arrangement. It should not be combined with another OEM's power or hydraulic limits to create an apparently complete but undocumented hybrid specification.
Cooling boundaries to confirm#
| Boundary | Required information |
|---|---|
| Compute hardware | Approved coolant and component operating conditions |
| Rack distribution | Manifolds, hoses, connections and pressure requirements |
| CDU | Included unit, performance conditions and controls |
| Facility loop | Available supply/return conditions and heat rejection |
| Residual air cooling | Heat not captured by the liquid system |
| Commissioning | Filling, cleanliness verification, monitoring and acceptance ownership |
CoolIT's CDU overview distinguishes liquid-to-liquid and liquid-to-air approaches. A CDU therefore needs an identified heat-rejection path; the word cooling in a quotation does not establish that the receiving building can remove the resulting heat.
Ask whether the cooling equipment is dedicated to one rack or shared across several. If shared, review the remaining capacity and operating behavior during the defined maintenance or failure case. A nominal capacity total does not establish redundant service across every connected rack.
Coordinate any separate purchase through SecondWatt's liquid-cooling catalogue. Keep the rack and CDU specifications aligned so two individually documented purchases do not create an unresolved interface between them.
Define workload, networking and software acceptance#
The complete rack should be evaluated against the intended workload and operating environment. Specify the application, software versions, parallelism, network arrangement and measurable acceptance target. The internal interconnect and the external cluster fabric need their own clear boundaries.
Do not assume that a populated rack includes every switch, optic or cable needed to connect it to storage and other racks. Request a network bill of materials and a topology drawing. Identify which links are installed and tested by the rack supplier and which belong to the receiving project.
System acceptance evidence#
| Test area | Evidence to agree before purchase |
|---|---|
| Identity and completeness | Inventory matched to the final bill of materials |
| Internal connectivity | Documented topology and relevant diagnostic results |
| External networking | Agreed interfaces and end-to-end test scope |
| Workload operation | Reproducible application test and acceptance measure |
| Thermal behavior | Recorded operating conditions and relevant telemetry |
| Power behavior | Observed demand and response under agreed conditions |
| Management | Monitoring, access and event-reporting checks |
Record limitations in the test. A factory demonstration may not include the buyer's external storage or network, while a site test may depend on infrastructure that is not yet complete. Distinguish supplier acceptance from full deployment acceptance and identify the evidence each milestone requires.
For a multi-rack purchase, identify whether acceptance covers each rack individually, the connected cluster or both. A complete rack test cannot demonstrate an external fabric that was not connected during the test. Conversely, a successful cluster workload should remain traceable to the delivered rack inventory and recorded configuration.
Agree how a late component substitution affects the evidence. A changed network device, firmware revision or cooling arrangement may require targeted retesting. The aim is to connect the accepted result to the final supplied system, with a clear record of changes and their disposition. Otherwise, a useful test report can become detached from the equipment it is supposed to support.
Verify software entitlement and support for the actual buyer, system and destination. Installed software does not by itself establish a transferable right to use or receive service for it. Obtain the applicable agreements and identify any subscriptions or configuration work the buyer must arrange separately.
Compare rack economics and deployment dates#
Normalize the offers after the technical boundary is clear. Compare the same rack quantity, included components, condition, support period and acceptance scope. Keep optional equipment separate from mandatory deployment items that have simply been omitted from one quotation.
This article provides no current rack-price range. A useful price needs a dated offer tied to an exact configuration and delivery boundary. A per-GPU or per-tray amount cannot substitute for the complete rack deployment total without the missing scope being identified and priced.
Complete deployment comparison#
| Comparison line | Scope to include or explicitly exclude |
|---|---|
| Rack hardware | Final documented configuration |
| Cooling additions | CDU, facility connection and related installation |
| Electrical additions | Distribution and required supporting infrastructure |
| Network additions | External fabric, optics, cabling and configuration |
| Software and support | Entitlements, service terms and deployment assistance |
| Logistics | Packaging, delivery, unloading and installation |
| Acceptance | Required tests, corrective work and retesting |
Review delivery and operation as different milestones. Confirm access, handling route, floor loading, service space, electrical readiness and cooling readiness before assigning a commissioning date. The power procurement and interconnection roadmap helps connect those dependencies to the wider project.
Ask the rack vendor for its shipment and handling requirements early enough for the receiving team to review the route. Identify unloading responsibility, temporary storage conditions and who checks the equipment before it moves into the final position. This is a planning task for the identified rack; generic server-rack handling assumptions are insufficient.
Create a discrepancy process for receipt. Record missing parts, packaging damage and identity differences against the shipment schedule, then agree whether installation can proceed and which acceptance activities must wait. The contractual remedy should refer to the same equipment boundary as the inspection record.
For staged delivery, list the minimum components and services required to make the first rack operational. A supplier may deliver compute hardware before the shared cooling or networking system. That can be a valid schedule choice, but it should not be described as usable first-phase capacity until the necessary interfaces and acceptance work are complete.
Where added site capacity is required, coordinate the transformer-sizing review with the rack schedule. Equipment delivery alone cannot establish that upstream electrical work, cooling plant or controls integration will be ready at the same time.
Use a readiness matrix before committing#
Give each interface an owner, required document and closure condition. The buyer should be able to see which items are confirmed, under engineering review or awaiting a supplier response. A readiness percentage is less useful than identifying the specific unresolved connection that prevents deployment.
NVL72 readiness and responsibility matrix#
| Gate | Required evidence | Owner to assign |
|---|---|---|
| Equipment | Final rack configuration and offered condition | Buyer and rack supplier |
| Electrical | Approved input and distribution interface | Facility electrical team |
| Cooling | Matched hydraulic and thermal conditions | Facility mechanical team and supplier |
| Network | Complete topology and connection scope | Network team |
| Software | Supported stack and applicable entitlements | Application/operations team |
| Site acceptance | Agreed test plan and remedies | Commissioning lead |
Use the matrix to compare GB200 and GB300 proposals on deployment fit. If one option requires a material facility change, identify its effect on cost and schedule before declaring a performance advantage commercially decisive. If both fit, compare the measured workload outcome and complete support arrangement.
SecondWatt acts as an independent intermediary. To request an NVL72 rack offer, specify GB200 or GB300, OEM preference, rack quantity, facility readiness, destination and target date. Include the desired workload and acceptance boundary so the response can address a complete, reviewable purchase.
FAQ: Buying GB200 and GB300 NVL72 racks#
Is an NVL72 rack the same as 72 separate GPU cards?#
No. NVL72 describes an integrated system arrangement, not a collection of independently interchangeable cards. Confirm the compute, internal interconnect, power, cooling and management configuration in the actual OEM offer. A GPU count alone cannot establish that the required rack components are present, compatible or covered by the proposed support terms.
Does a rack quote normally include the CDU?#
The quotation must answer that explicitly. Some boundaries may include rack manifolds while leaving the CDU, facility loop or installation to another supplier. Use the bill-of-materials worksheet to identify each cooling component and responsibility. Do not infer inclusion from a phrase such as liquid-cooled rack or complete cooling.
Which power number should the facility use?#
The facility engineer needs the vendor's complete planning information, including operating, peak and provisioning fields with their conditions. Those numbers serve different purposes. The HPE example in this article preserves its published labels and applies only to that implementation; it is not a universal electrical design rule for every NVL72 rack.
Can I reuse GB200 facility requirements for a GB300 offer?#
Only after confirming the actual rack vendor's requirements and reviewing the interfaces. Similar platform naming does not establish identical electrical, cooling, mechanical or control conditions. Request the current documentation for the proposed configuration and have the receiving teams identify any changes before the purchase and installation scopes are finalized.
Does power smoothing replace the UPS?#
No such conclusion follows from the platform feature alone. Power smoothing addresses specified power behavior within its design scope. The project's backup, continuity and recovery requirements need their own electrical assessment and supporting equipment. Obtain the actual rack's behavior and evaluate it within the complete facility operating arrangement.
What should I send with an NVL72 sourcing enquiry?#
Provide the preferred platform and OEM, rack quantity, workload, destination and target dates. Include facility electrical and cooling information, accepted equipment condition, network boundary and required support. Ask for a complete bill of materials, deviations and acceptance obligations so the offer can be assessed as a deployment rather than a headline rack price.