---
title: "GPU Deployment Spares: Separate Installed Redundancy from Replacement Stock"
description: "Leviathan Systems — Technical content Published 2026-10-20; updated 2026-10-20."
canonical: "https://www.leviathansystems.co/articles/gpu-deployment-spares-versus-redundancy"
last-updated: "2026-10-20"
---

# GPU Deployment Spares: Separate Installed Redundancy from Replacement Stock

Leviathan Systems — Technical content
Published 2026-10-20; updated 2026-10-20.

An installed redundant component is part of the installed equipment configuration; a separately held replacement supplies material for later appointed service. Before buying a GPU deployment spare package, identify the exact replaceable units, their applicability to the delivered equipment, the owner and location of the stock, and the reservation, issue and replenishment arrangements. Compare those results separately from unused ports, excess GPU inventory, installation materials and an OEM support agreement. No universal spare percentage establishes the right purchase.

## State which spare problem the purchase is meant to solve

A deployment quotation can use spare to mean an unused cable, an unassigned port, a replacement module or additional compute capacity. Those purchases solve different problems. An installation allowance may address a damaged assembly discovered before handoff. An operating spare may address a later product failure. Extra installed GPUs may serve a capacity plan without being independent replacement units at all.

Define the intended requirement before adding a quantity to the order. Is the buyer trying to complete the original installation, reduce dependence on a future parts delivery, or provide stock for an appointed service organization? Identify the equipment population and the actual event the proposed inventory is intended to support. A room with two suppliers and different system revisions may need a different applicability comparison from a uniform delivery.

Keep the decision connected to the operating plan. The equipment owner and its qualified service provider decide which items are appropriate to hold, based on the actual products, support arrangements, replenishment and permitted service process. This purchasing comparison provides no failure-rate model or stocking percentage.

## Leave installed redundancy in the installed population

NVIDIA's DGX B200 user guide describes six installed power-supply units configured for 5+1 redundancy. Its table entry for loss of one AC input lists five operating PSUs and no system effect. That illustrates the contribution of the installed arrangement: continued operation under a documented condition. All six units remain in the installed population; the sixth is not a boxed replacement.

Separately held stock contributes at a later phase. An applicable replacement can supply material to the appointed service owner when replacement is required, reducing dependence on obtaining that unit through a future delivery. After issue, replenishment determines when another applicable unit returns to the free pool. Buying a stock item alone establishes neither continued system operation nor a restoration time.

For a receiving or purchasing comparison, give each separate replacement its own inventory identity and applicable equipment reference. Counting the same physical object once as installed redundancy and again as available stock inflates what the buyer has actually received.

The guide prohibits operating with PSUs depopulated; the redundancy label gives no permission to borrow an installed unit for another machine. The input-loss entry is descriptive product context, not a test instruction or a prediction for every PSU fault. Actual product conditions and the responsible technical and operating owners govern diagnosis, authorized service and validation.

References: [N1](https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-to-dgxb200.html#power-specifications)

## Buy the applicable service unit rather than a similar-looking part

The DGX B200 service manual separately identifies customer-replaceable components, including power supplies, fan modules and named storage and network items. It directs returns through NVIDIA Enterprise Support for an RMA and says replacement components must be supplied by NVIDIA. This is a product-specific service boundary, not a general authorization for a contractor to repair every GPU system or to source interchangeable parts independently.

Ask the supplier and appointed product-support owner to confirm the actual replacement unit, applicable system and configuration, permitted source and any relevant revision or software prerequisites. A generic label such as power supply or 400G module does not establish that relationship. Nor does a larger assembly always divide into independently usable spare subcomponents. The exact serviceable unit matters to both the quotation and the physical inventory.

Keep that confirmation with the item record. Where applicability remains unresolved, describe the item as received but unconfirmed for the proposed use. New, sealed packaging or proximity to the equipment does not settle suitability. Equally, an applicability gap is not proof that the item is defective; it is a question for the actual supplier or product authority to settle.

References: [N2](https://docs.nvidia.com/dgx/dgxb200-service-manual/introduction.html)

## Compare one pooled spare through reservation, issue and replenishment

Fictional example SPR-01R continues SPR-01 with two DGX B200 systems, S1 and S2. Using the guide's six-per-system population gives twelve installed PSU identities. A separate delivery contains two boxed items, P1 and P2. For this example alone, P1 is stipulated to be an applicable replacement for either S1 or S2; P2's applicability remains unresolved. These premises are invented, not actual part eligibility or product approvals.

The initial inventory therefore tracks fourteen physical identities: twelve installed and two separately held. Only P1 is both applicable and unreserved. Neither the installed units nor unconfirmed P2 belongs in that one-item free pool. Now two material requests arrive together, one for each system. P1 can be reserved for either request, but assigning it to both would duplicate one physical item.

Follow the states below across the recorded custodies. Reservation changes allocation without moving a unit. Physical issue transfers P1 to the appointed service owner's custody. No installation or component change occurs anywhere in this continuation, so the twelve installed identities remain twelve. The totals describe tracked units across those custodies, not an onsite warehouse count.

A replenishment request, order or dispatch record does not establish receipt of another unit. The later P3 receipt is a separate fictional event, with applicability to both systems expressly stipulated. It adds a new physical identity; reserving P3 against outstanding S2 then consumes the one free allocation. Compare proposed pools using their applicable, unreserved units and replenishment state rather than the delivery total. This is allocation reasoning, not a failure forecast or a stocking recommendation.

References: [N1](https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-to-dgxb200.html#power-specifications)

SPR-01R: reservation changes the free pool; a new receipt changes the physical total
| Recorded event | Physical identities across recorded custodies | Unreserved applicable held units | Source |
| --- | --- | --- | --- |
| Initial receipt | 12 installed + 2 held (P1 and P2) = 14. | 1: P1 applies to either S1 or S2. P2 remains unresolved. | Planning analysis |
| P1 reserved for S1 | 12 installed + 2 held = 14. P1 is still held, now reserved. | 0: the simultaneous S2 request has no free applicable unit. | Planning analysis |
| P1 physically issued to the appointed service owner | 12 installed + 1 issued (P1) + 1 held (P2) = 14. | 0: issued P1 is outside the held pool; P2 remains unresolved. | Planning analysis |
| Replenishment requested, ordered or reported dispatched; no new receipt | 12 installed + 1 issued (P1) + 1 held (P2) = 14. | 0: a record of a future supply action adds no received unit. | Planning analysis |
| New P3 received, stipulated applicable to S1 and S2 | 12 installed + 1 issued (P1) + 2 held (P2 and P3) = 15. | 1: P3 is unreserved; P2 remains unresolved. | Planning analysis |
| P3 then reserved for outstanding S2 | 12 installed + 1 issued (P1) + 2 held (P2 and P3) = 15. | 0: P3 is now reserved; no physical-count change occurs. | Planning analysis |
Wholly fictional identities and states. Issue and receipt establish the stipulated custody and inventory changes only; neither establishes completed service or restored equipment.

## Compare possession, access and replenishment with the part price

A matched part in another location may serve a different need from a customer-held item at the agreed site. Compare the actual place, custodian, permitted access, reservation rules and delivery responsibility alongside the quoted unit. Do not assume that a service entitlement establishes local inventory, or that an item described as available is reserved for this customer.

Ask the supplier to identify who replenishes after issue, what triggers that action, and the starting and finishing events of any stated target. Keep the reservation, physical release, associated request or system, failed-item return and replenishment records connected. A temporary transfer to another site changes the first site's stock position; without that update, two teams can plan against the same supposed reserve.

Consider wholly fictional offers Shared and Reserved for the same applicable replacement requirement and agreed location. The following terms are invented illustrations, not NVIDIA terms or current market offers. They establish no actual stock, part eligibility, local availability, price, attendance or fulfillment promise.

Shared identifies one applicable supplier-held unit shared across customers, with no customer reservation. It can therefore be allocated to another customer. The supplier is responsible for replenishment following physical issue. Its illustrative dispatch target starts at an approved replacement request and ends at dispatch one business day later; no physical-receipt target for replenishment is supplied.

Reserved assigns one identified applicable unit exclusively to this customer, excluding other customers' claims on it. The supplier replenishes when that unit is physically issued, with an illustrative target starting at issue and ending at physical receipt of the replenishment unit three business days later. This customer still has only one reserved item: it cannot meet both S1 and S2 allocations at once.

The two targets measure different events from different starts. One business day to dispatch after request approval cannot be ranked against three business days to replenishment receipt after issue as though either measured restoration speed. Reservation answers whose competing claims are excluded; quantity answers how many requests can receive distinct units; replenishment answers how the pool is refilled.

NVIDIA's support-guide opening separately identifies response, escalation and replacement shipping, with the actual solution and contract controlling the terms. Keep a support response, parts dispatch, physical receipt, appointed service and validated restoration distinct in the buyer's comparison. The material records address dispatch and receipt; separately authorized product work and evidence from the responsible technical and operating owners are needed to establish the later outcomes. The related GPU Rack Support Coverage guide addresses the wider support responsibilities.

The actual supplier and OEM terms determine return requirements, replacement conditions and product-specific storage or packaging instructions. The operating owner determines who may request, release and perform the agreed work. A component price alone settles none of those interfaces. Compare quotations with the same material commitment and service scope, including any unresolved diagnosis, qualified attendance, access, authorization or system-validation responsibility.

References: [N3](https://docs.nvidia.com/enterprise-support-and-services-user-guide/references-and-guidelines/)

## Carry the agreed spare population into the physical handoff

An agreed physical package can identify which separate items were received, their observable identifiers and condition, where they were handed over, and which applicability questions remain open. Keep installation materials and separately reserved operating stock distinguishable so completing the original build does not silently consume the customer's intended reserve. Actual custody and release are engagement-specific.

For a Leviathan discussion, use the related Rack Assembly & Preparation service and its project inquiry to identify the requested receiving, reconciliation, preparation or physical documentation task for customer-furnished equipment. Confirm the location, supplied inputs, handoff result and appointed material and service owners. This does not establish Leviathan-owned stock, a warehouse, a depot service, procurement authority, an OEM repair appointment or a guaranteed replenishment time. Product-support and operating decisions remain separately scoped.

The useful closeout is a reconciled population and explicit next owner: what is installed, what is separately held, what has confirmed applicability, what is reserved or issued, and what replenishment has actually been received. Keep P2-like unresolved items and outstanding requests visible. That gives the buyer a basis for the spare-package decision without turning arithmetic or a redundancy label into a maintenance promise.

## Primary sources

- [N1 — NVIDIA DGX B200 User Guide: Power Specifications and PSU Redundancy](https://docs.nvidia.com/dgx/dgxb200-user-guide/introduction-to-dgxb200.html#power-specifications). Checked 2026-10-20. DGX B200 User Guide, updated 17 October 2026: installed six-PSU/5+1 arrangement, the one-input-loss table entry and prohibition on depopulated operation. Useful product context for separating continued operation from replacement stock; not a general fault guarantee.
- [N2 — NVIDIA DGX B200 Service Manual: Introduction and Customer-replaceable Components](https://docs.nvidia.com/dgx/dgxb200-service-manual/introduction.html). Checked 2026-10-20. DGX B200 Service Manual introduction, updated 13 September 2026: replaceable-component classes, Enterprise Support/RMA route and NVIDIA-supplied replacements. Establishes the product-service boundary, not any fictional item's eligibility or a contractor's repair authority.
- [N3 — NVIDIA Enterprise Support and Services User Guide: References and Guidelines](https://docs.nvidia.com/enterprise-support-and-services-user-guide/references-and-guidelines/). Checked 2026-10-20. Opening terms distinguish support response, escalation and replacement shipping under solution- and contract-specific terms. Useful for comparing event commitments; no numerical service target or networking-warranty provision is applied to DGX.

## Related Leviathan resources

- [Reconcile supplied installation materials](/articles/customer-furnished-materials-installation-checklist)
- [Keep port allocations and cable inventory distinct](/articles/as-built-handoff-package-gpu-deployment)
- [Separate the GPU target from whole-system purchasing counts](/articles/h100-deployment-checklist)
- [Discuss equipment receiving and preparation](/services/gpu-assembly)
- [Identify product-support and operating responsibilities](/articles/gpu-rack-support-coverage-handoff)

## Define the received equipment and material handoff

Share the supplied equipment and separate-item list, requested physical preparation or reconciliation work, location and receiving owner. We can discuss the bounded physical package while keeping stock ownership and product-service responsibilities explicit.

[Discuss the scope](/contact?service=gpu-assembly&from=/articles/gpu-deployment-spares-versus-redundancy)