Platforms
H100 to H200 Refresh: Prove Compatibility Before Reusing the Rack
An H100-to-H200 refresh starts by defining the change: replace complete servers, propose an OEM-supported conversion or add H200 systems alongside existing equipment. Obtain support for the exact system and change path, then document which rack, cable and facility interfaces may be reused. A common chassis name or similar GPU rating does not approve a field swap, establish site readiness or assign the work.
Define the change path before assigning installation work
H100 to H200 can describe three different projects. Replacing a complete server changes the installed asset. A proposed conversion changes an existing asset and needs an OEM-supported path for that exact configuration. Adding new servers changes the room's population while existing systems remain in place. Start with the intended end state and give the project one of those descriptions.
Record the existing system model, asset identifier and configuration, then the proposed equipment or approved change package. Identify which systems remain operational, which assets leave and which connections change. The customer and its appointed owners must approve the service window and dependencies before the field work is released.
The distinction matters when scoping labor. A work order to mount a replacement server is not permission to modify its GPU assembly. A proposed conversion is not ready for execution because a supplier has offered H200 parts. Resolve the supported path before treating either as an installation package.
Scroll the table horizontally on smaller screens.
| Proposed path | Evidence to obtain first | Physical handoff to define | Source |
|---|---|---|---|
| Replace complete H100 servers | Exact replacement H200 system, applicable OEM documents and support/deployment arrangement | Old/new asset locations, released mounting/cabling scope and removal responsibility | Planning analysis |
| Propose an in-place conversion | Written OEM confirmation for the actual starting configuration, supported parts and responsible service team | Precisely authorized work; do not assume a conversion is offered or field-permitted | Planning analysis |
| Add H200 systems alongside H100 | Exact new systems plus approved layout, external connection plan and facility confirmation | New asset positions, protected existing connections and agreed expansion records | Planning analysis |
Original planning distinctions. The conversion row is a question for the OEM, not an assertion that any particular system has an approved conversion kit.
Identify the exact H200 variant and system
NVIDIA's current H200 table separates H200 SXM from H200 NVL. The NVL column identifies a PCIe, dual-slot air-cooled form factor, while the SXM column identifies SXM. These are different product configurations. Keep the exact variant in the equipment list rather than writing only H200.
The next identity is the complete server: manufacturer, full model, GPU count, installed adapters, support reference and relevant document revision. That record connects a product specification to the asset the team will receive or modify. If an offer changes after the site review, return the changed configuration to the responsible reviewer instead of carrying the earlier approval forward by name alone.
A useful supplier question is whether the proposed order is a complete configured system, a replacement component under support or an approved upgrade package. Ask for the document that describes the supported starting and ending configurations. A shared generation, socket description or marketing compatibility phrase cannot substitute for that answer.
References: [N1]
Compare complete H100 and H200 systems before inferring rework
Some published values can remain the same while the equipment identity changes. NVIDIA's DGX H100/H200 guide gives a concrete complete-system comparison: eight H100 GPUs or eight H200 GPUs, with an 8U form factor and a published 10.2 kW maximum whole-system power value for the documented family. Those matching enclosure and power figures do not make the accelerator populations identical or establish an in-place conversion. They also do not describe every H100 or H200 OEM system.
Read the comparison at the level of the offered equipment. A product page for an H200 accelerator, a catalog entry for a complete H200 server and the actual support disposition for an existing H100 asset answer different questions. The current shared DGX guide is evidence about the documented system configurations. It is not a finding that an existing serial-numbered system has the same delivered options or can be converted by replacing its accelerators. The earlier change-path and support-list distinctions still control that separate question.
The table below separates what the named comparison actually shows from the site-interface decision it cannot make. A matching value belongs in the source-comparison column of the reuse register. It should not silently become reuse approved. The mounting arrangement, installed and service envelopes, external interfaces, supplied accessories and facility requirements must still be compared for the actual configurations and location.
Fictional EX-HR96 proposes replacing one complete DGX H100 with one complete DGX H200 in an existing position. Its preliminary comparison records three matching published numbers: eight GPUs per system, 8U enclosure height and the guide's 10.2 kW maximum system value. It also records a changed accelerator identity, H100 to H200. The actual ordered system, mounting package, external connection issue and applicable facility responses have not been supplied. These are invented proposal inputs, not an inspected asset or a supported replacement finding.
The correct comparison result is therefore limited: the named public configurations differ in accelerator identity and share the cited published bounds. EX-HR96 does not establish that any rail, PDU, cable or cabinet can stay. It does not prove that a change is required either. Keep each interface unresolved until its actual owner returns the applicable decision; do not invent a defect merely because a new accelerator generation is proposed. One proposed complete server for one existing position is a quantity statement, not a compatible installation package.
Installation assignment is another independent boundary. The DGX H100/H200 quickstart requires relevant site information for the Installation Partner and specifies installation by NVIDIA partner network personnel or NVIDIA field service engineers. A shared requirement across the named family is not evidence that any particular contractor is eligible or appointed. Confirm the real order and task arrangement. This comparison offers no warranty interpretation, component conversion, electrical or cooling design, handling, software migration or commissioning method.
Scroll the table horizontally on smaller screens.
| Comparison field | What the named source establishes | What still needs an applicable decision | Source |
|---|---|---|---|
| Accelerator identity and population | The documented complete systems use eight H100 or eight H200 GPUs. The count matches; the accelerator identity changes. | Exact ordered system and options, existing asset identity and any proposed OEM-supported change path. No internal field conversion follows from the count. | [N4] |
| Enclosure height | The shared complete-system guide specifies an 8U rackmount form factor. | Actual rack, mounting parts, dimensions, installed/service access and configuration applicability. A common rack-unit height is not the full interface. | [N4] |
| Maximum whole-system power | The guide gives a 10.2 kW maximum system value for the documented H100/H200 family. | Applicable complete-system requirements and the facility owner's disposition for the real installation. A published maximum is not measured operating demand, a GPU TDP or an electrical design. | [N4] |
| Installation channel | The H100/H200 quickstart identifies its required personnel and relevant site-information return. | Actual appointed arrangement, permitted task boundary and required handoffs. The document does not certify a bidder, transfer a product-controlled task or approve the site. | [N5] |
Original educational comparison from the named DGX documents, checked October 19, 2026. The common figures are not a compatibility, substitution, actual load or retained-interface approval and are not generalized to other OEM systems. EX-HR96 is fictional and reports no performed work.
Read the support list at the right level
Dell's PowerEdge XE9680 manual lists both eight-GPU H100 and eight-GPU H200 configurations. NVIDIA's system catalog also lists the XE9680 with HGX H100 and HGX H200. These sources help identify documented system configurations. They do not, by themselves, establish an authorized conversion for a particular existing server.
For that proposed change, obtain a response tied to the system's actual identity and configuration. What parts and revisions does the OEM support? Who may perform the work? What happens to support entitlement and the existing configuration record? Which OEM checks and records close the change? Keep the response and its applicable scope with the work package.
If that evidence is unavailable, label the in-place path unconfirmed. A complete replacement-system option can be evaluated on its own evidence without presenting the unconfirmed conversion as a routine drop-in upgrade. The buyer's decision becomes clearer when those paths are not combined into one assumed scope.
Respect the OEM deployment and service boundaries
The XE9680 GPU-specification page states an OEM deployment-services requirement: ProDeploy Plus or Dell Customer Deployment services, with mandatory GPU subsystem tests as part of deployment for applicable XE-series models. Confirm the actual order and service arrangement with Dell before assigning overlapping work. A third-party physical installation scope must be coordinated with that requirement.
NVIDIA's DGX H100/H200 service manual likewise places component replacement within its support process: use NVIDIA-supplied replacement parts, obtain the applicable support guidance and consult the system terms before modification. That manual is not a generic instruction to convert any H100 server to H200.
Put the boundaries into the schedule. The OEM or appointed support team may own internal system work and its acceptance, while a separately authorized team owns mounting, external cabling and physical records. Identify the interface between those teams before the equipment arrives. This guide does not reproduce component replacement, electrical, cooling or diagnostic procedures.
Create a change and reuse register for the site interfaces
Reuse should be a recorded decision about a named item. Give each proposed reused interface an identity, its existing condition, the proposed equipment requirement, supporting document and responsible reviewer. Use three states: reuse approved, change required or unresolved. Avoid a single rack compatible checkbox that hides several different decisions.
The register below is an original way to keep those decisions visible. The evidence must apply to the delivered system and the actual site. A photograph can establish an observed label or location; it cannot establish an unmeasured electrical or thermal capability.
Scroll the table horizontally on smaller screens.
| Interface | Evidence to compare | Record the disposition | Source |
|---|---|---|---|
| Rack and mounting hardware | Actual rack, rail/mounting part, adjustment range, service access and applicable OEM references | Approved reuse, specified change or unresolved reviewer question | Planning analysis |
| Facility connections | Delivered system requirements and confirmation from the appointed facility owners | Approved scope and responsible owner; no inferred adequacy from GPU ratings | Planning analysis |
| External network and management | Actual adapter/port identities, role, approved media and released endpoints | Which connections stay, move or change; preserve traceable cable identities | Planning analysis |
| Existing customer cables | Cable identifiers, routing/condition records and approved new endpoint requirements | Reuse decision and any agreed inspection or verification evidence | Planning analysis |
| Support and system configuration | OEM-supported starting/end state, service scope and named acceptance owner | Written support disposition; no assumed field conversion | Planning analysis |
| Records and labels | Before/after asset locations, rack elevations, cable schedule and exceptions | Approved revision and closure owner for each changed record | Planning analysis |
This register organizes evidence. It does not certify fit, loading, electrical capacity, thermal performance, cable compliance or an OEM upgrade path.
Review the complete system and the actual site
Do not infer rack or facility approval from equal GPU TDP figures. The site review concerns the complete delivered system and its surroundings: mounting hardware, dimensions, service access, external interfaces and the facility requirements that the appointed owners must accept. Even an unchanged equipment position needs a recorded decision about the proposed configuration.
Dell's XE9680 documentation points to a rail-sizing and rack-compatibility matrix covering rail types, adjustment ranges, depth and rack mounting flanges. It also provides a thermal-restriction matrix with configuration-specific information. Those document types illustrate why a GPU data sheet is only one input to the reuse review.
Ask the facility and OEM owners to identify the applicable documents and required changes. Keep a pending item open until the responsible person supplies the evidence. Do not turn the review into improvised breaker adjustments, coolant setpoint changes or cold-plate work. If the site cannot accept the selected configuration, change the plan before releasing the physical installation.
Preserve connection traceability during the refresh
A retained cable can have a new endpoint. Before work starts, record the old and proposed device/port identities, cable identifier and drawing revision. Identify the connections that must remain untouched and who releases changes near operating equipment. The new server's adapter and port map governs the proposed connection; the old server's GPU family does not.
Keep removed, reused and newly installed cables distinguishable in the work record. If a label or endpoint differs from the plan, record the discrepancy and obtain the appropriate disposition rather than silently changing the schedule. Capture the resulting physical route and endpoint record after the authorized work.
Any requested verification must have an agreed method, qualified resources and acceptance criteria. Labeling and physical connectivity records are different from formal optical/copper certification or application performance acceptance. Do not expand a cable-reuse decision into an unsupported claim that the whole fabric has been tested.
Work through a replacement-server example
Illustrative case: a customer proposes replacing two H100 servers with two complete H200 servers in the same cabinet positions. The team wants to retain the external cable routes and labels. The proposed server count and positions are known; rail suitability, facility acceptance, port mapping and the OEM deployment arrangement are still separate questions.
The change register might show the existing cable route as observed, the new endpoint map as awaiting approval and the mounting decision as pending the actual rail/system documentation. That is a useful status report. It does not require inventing a defect or claiming that everything is compatible because both orders contain two servers.
Once the responsible owners close those items, the installation scope can name the customer-furnished systems, permitted mounting and cabling tasks, protected connections and handoff records. OEM/system acceptance remains separately assigned. This is a fictional planning example, not a Leviathan customer project, labor estimate or evidence of any specific server's upgrade compatibility.
Close the physical handoff and the open items separately
The physical handoff should identify the actual installed assets and positions, the released plan revision, completed mounting/cabling work and any deviations or unresolved items. Where permitted, approved photographs and updated cable/rack records can make that state easier to review. Keep OEM deployment or system-acceptance evidence attached under the responsible owner's name rather than relabeling it as the installer's certification.
Record who owns the next step for each exception. A deferred software check, a support decision and a physical labeling discrepancy do not have the same closure evidence. Use the project's agreed acceptance process to track completed physical scope and separately outstanding work.
Leviathan can discuss agreed customer-equipment mounting, rack-and-stack, structured cabling, labeling and physical handoff documentation. GPU supply, internal conversion work, electrical changes, cooling-system work, workload tuning and system commissioning are not promised by this article. Share the exact proposed equipment, OEM/support arrangement, site and authorized physical tasks when asking about installation scope and availability.
References: [L1]
Common planning questions
Can an H100 server be converted to H200 by swapping GPUs?
Do not assume it. Obtain an OEM-supported path for the exact starting configuration, supported parts and authorized service team. A catalog listing both configurations is not that authorization.
Can we reuse the rack, cables and PDU for an H200 refresh?
Decide each item against the actual delivered system, site requirements and responsible-owner approval. Record approved reuse, required changes and unresolved items separately; GPU ratings alone do not settle the decision.
Does one manual covering H100 and H200 prove a drop-in upgrade?
No. A shared manual can describe multiple supported system configurations. Confirm whether the proposed change to the specific existing asset is supported, who performs it and how acceptance is completed.
What should be included in an H200 installation inquiry?
The exact OEM system and variant, whether it replaces or supplements equipment, site, OEM deployment/support arrangement, customer-furnished materials, released plans and authorized physical tasks.
Primary sources
- [N1] NVIDIA H200 product specifications: SXM and NVL variantsChecked 2026-10-15. Different SXM and PCIe dual-slot air-cooled NVL form factors. Product page retains preliminary-specification caveat. No generic H100 conversion, thermal equivalence or performance claim adopted.
- [N2] NVIDIA DGX H100/H200 service manual: introduction and service boundaryChecked 2026-10-15. Support guidance, modification terms and NVIDIA-supplied replacement parts for this specific DGX family; page states September 13, 2026 update. No component procedure or stress-test command reproduced.
- [N3] NVIDIA certification programs: listed system configurationsChecked 2026-10-15. XE9680 listed with HGX H100 and H200; not an upgrade authorization for a serial-numbered server, site qualification or certification of Leviathan.
- [D1] Dell PowerEdge XE9680 installation and service manual: GPU specificationsChecked 2026-10-15. Documented H100/H200 configurations and Dell deployment-services requirement. Actual order/support arrangement requires Dell confirmation.
- [D2] Dell PowerEdge XE9680: rail sizing and rack compatibility matrix referenceChecked 2026-10-15. Manual identifies rail adjustment, depth, rail types and rack-flange topics; no actual rack fit or rail part is certified by this guide.
- [D3] Dell PowerEdge XE9680: thermal restriction matrixChecked 2026-10-15. Evidence that configuration-specific thermal documentation must be considered. No setpoint, service time, fan procedure, temperature limit or thermal-equivalence claim adopted.
- [L1] Leviathan rack-and-stack installation scopeChecked 2026-10-15. Conditional physical installation of customer-furnished equipment under agreed scope. Purchasing and delivery are separately scoped. No internal GPU conversion, OEM authorization, electrical/cooling work, commissioning or technical certification inferred.
- [N4] NVIDIA DGX H100/H200 user guide: paired complete-system component, mechanical and power specificationsChecked 2026-10-19. Only the named eight-H100/eight-H200 system identity, common 8U form factor and published 10.2 kW maximum whole-system value. No universal OEM equivalence, actual fit, load, rail/PDU/cable reuse, supported conversion, numerical electrical design or operating procedure adopted.
- [N5] NVIDIA DGX H100/H200 quickstart: installation channel and site-information boundaryChecked 2026-10-19. Only the named family requirement for relevant site information and specified installer personnel. No actual bidder eligibility, engagement, warranty-rights interpretation, registration, command, startup, diagnostics, software or commissioning procedure adopted.