Nine H100s or One: Sizing the Agentic Maximo Assistant Before You Promise It

The hardware requirement for an on-premises Maximo Assistant depends on its AI Service patch level. Start with the documented deployment boundary, then prove tool readiness, permissions, and operational control before promising a production service.

Share
Nine H100s or One: Sizing the Agentic Maximo Assistant Before You Promise It

Nine H100s or One: Sizing the Agentic Maximo Assistant Before You Promise It

The first question in an agentic Maximo Assistant proposal should not be how impressive the demonstration looks. It should be which AI Service build the proposal assumes. IBM's enablement documentation describes an unusually consequential difference: for an on-premises or air-gapped deployment of AI Service 9.2, the listed GPU requirement changes from at least nine H100 GPUs without the July 2026 patch to at least one H100 GPU with that patch. The documented node still has 16 CPUs and 184 GB of memory in either case.

That is a deployment prerequisite, not a published benchmark. It does not establish how many technicians can ask questions simultaneously, how quickly a complex asset query will return, or what a production service will cost. Those are separate questions that require measurements in the proposed environment. Treating the smaller requirement as a complete sizing answer would exchange one planning mistake for another.

The useful lesson is that patch level, model topology, tool availability, and user permissions belong in the same design review. A programme that treats them as separate workstreams can buy hardware against one assumption, demonstrate another configuration, and ask security to approve a third.

This article uses IBM's agentic assistant enablement instructions and Maximo Assistant architecture as its product evidence. The testing and operating practices below are recommendations, not additional IBM requirements or measured customer results. The aim is to turn the documented prerequisites into a defensible acceptance plan before an assistant becomes a commitment to the business.

1. Treat the patch level as a sizing input

IBM explicitly ties the hardware difference to the models used by the assistant. With the July 2026 patch, AI Service 9.2 uses gpt-oss-120b for all assistant processes. Without that patch, the documented configuration uses both gpt-oss-120b and Llama-4-Maverick-17B-128E-Instruct-FP8. The architecture page assigns the latter model to creating and verifying plans, routing requests, and processing follow-up requests. Other actions default to gpt-oss-120b.

For the on-premises or air-gapped case, the enablement page specifies a Red Hat OpenShift 4.20 cluster with one GPU secondary node. Its listed CPU and memory requirements remain the same across the two patch conditions. The GPU count is the part that changes. Do not generalize this into a claim that every Maximo AI workload now needs one GPU, or that any available accelerator is equivalent to an H100. Neither conclusion follows from the opened documentation.

The practical design artifact should therefore begin with a component manifest. Record the proposed AI Service level, whether the relevant patch is included, the model arrangement, the hosting boundary, and the hardware prerequisite being used. Attach the applicable documentation reference. A line that says only MAS 9.2 is insufficient to resolve the hardware question because the source itself distinguishes configurations within that release family.

Next separate minimum prerequisite evidence from capacity evidence. The prerequisite answers whether the proposed environment matches the documented starting point. Capacity evidence answers whether that environment supports the intended workload. A representative test should include the kinds of requests the business actually expects: a simple record lookup, a question requiring an aggregate, an asset insight request, and a follow-up that depends on conversation context. These are proposed test categories, not a vendor benchmark suite.

Measure response time, unsuccessful requests, resource use, and the effect of overlapping users. Keep the test data and configuration stable enough that comparisons mean something. If the patch level changes, record that change rather than attributing every improvement to hardware. If a tool is unavailable during one run, do not compare its general-knowledge answer with a tool-backed answer from another run as though the requests exercised identical paths.

Finally, avoid translating a hardware count directly into a procurement saving. Node availability, support arrangements, model hosting, storage, operational effort, and service expectations still need evaluation. IBM's documentation supplies an important prerequisite distinction, but it does not price the proposed service. The responsible conclusion is narrower and more useful: resolve the patch condition before accepting a sizing estimate, then validate the service against the workload it is supposed to carry.

2. Prove the service chain, not just the chat window

The architecture page describes a chain of components behind an agentic response. A user submits a prompt in the chat window. The prompt reaches the supervisory ALM agent, which runs as the alm-agent pod in OpenShift. The agent calls AI Service, which sends the request to the model in watsonx.ai. The agent determines the required tools and invokes them on the Maximo MCP server using JSON-RPC 2.0. The server routes each call to the appropriate Maximo Manage API, and results return through the same chain to the response.

This matters because a visible chat window does not prove that the full chain works. A response may exist even when the deployment has not demonstrated successful access to the intended operational data. IBM's enablement instructions explicitly say that when no AI configuration tools are ready, the assistant can answer asset lifecycle management prompts using general knowledge. When tools are ready, the assistant can still choose between them and general knowledge.

Consequently, the acceptance test should distinguish conversational availability from task completion. Ask for information whose correct answer can be checked against an authorized Maximo record. Preserve the request, expected scope, actual response, and supporting operational evidence available in the deployment. A plausible paragraph is not enough to establish that a data retrieval tool was used or that its answer is complete.

IBM's procedure gives administrators specific readiness checks. AI Service must be connected. The deployment's configuration cards must show Running or Ready before continuing. The optional OpenShift check looks for the alm-agent pod in Running status. AI configurations intended as tools must show Ready, and the assistant's Tools section must indicate that the relevant configurations are available. These checks establish different facts and should not be collapsed into a single green checkbox.

For an operating team, a useful handover document maps each symptom to its next diagnostic boundary. If chat is unavailable, start with the interface and service availability. If chat works but a record question fails, investigate the relevant tool configuration, permissions, and underlying API path. If one user succeeds and another does not, compare access before assuming a model defect. These are diagnostic recommendations based on the documented chain, not a claim that every failure has one predictable cause.

The same distinction helps business sponsors. A successful demonstration should say what was proved: this user, this configuration, this data scope, and this request. It should not silently expand into a promise about every user, every site, or every asset question. The operational chain is the unit of acceptance, while the chat window is only its visible starting point.

3. Make tool readiness and refresh behavior explicit

IBM lists four assistant tools in its architecture documentation: data retrieval, condition insights, documentation search, and schema retrieval. Data retrieval translates natural-language requests into OSLC API calls against Maximo Manage object structures and uses the nl2oslc model template. Condition insights uses insightsgenerator. Documentation search uses docsearch and searches a static, curated set of IBM Documentation articles. Schema retrieval provides API schema information and does not require an AI configuration.

These distinctions should appear in the service description. A data lookup and a condition insight are not interchangeable outcomes. The enablement page shows asset insights using both insightsgenerator and nl2oslc, while ordinary record and history questions use nl2oslc. An acceptance pack should therefore identify the tools expected for each supported use case instead of asking a collection of convenient questions and declaring the whole assistant ready.

There is also a timing boundary. IBM says the assistant is refreshed with the latest active configurations every 10 minutes. The procedure instructs administrators to wait at least 10 minutes and describes the MCP server's tool refresh on that interval. It also documents a pod-deletion method to force refresh. That operational action belongs in an approved runbook, not as a casual instruction to restart production components whenever a tester becomes impatient.

The timing matters during controlled changes. If a tool configuration has just been activated, an immediate test may not represent the final configuration. If an administrator is trying to remove a capability, it is equally unsafe to assume that saving a setting proves that every subsequent request has already adopted it. Observe the relevant readiness state and test the intended result after the documented refresh behavior has been accounted for.

IBM further notes that existing active configurations usable as tools are automatically used by the assistant. That makes configuration inventory important. Before a pilot, list the active configurations and identify which are exposed through the assistant. A team should not discover its effective tool set by watching an unexpected answer in front of a business audience. Review this inventory after changes as well as before the initial launch.

A disciplined test record links a supported question to its prerequisite tool configuration, readiness evidence, authorized data, and expected result. It should also include an intentionally unsupported request. The purpose is not to demand that the assistant answer everything. It is to learn whether the service fails intelligibly when a required capability is absent. A clear limitation can be documented and taught. An unexplained fallback that looks like an authoritative operational answer is much harder for users to recognize and manage.

4. Test permission differences and disconnected references

IBM states that Maximo Assistant operates within the same security, permissions, and approval rules as Maximo Manage and Maximo Application Suite. Two users asking the same question can receive different responses because their security group permissions and access rights differ. The enablement procedure separately requires granting assistant access and ensuring access to the data used by the assistant. Access to the interface is therefore not evidence of access to every underlying record.

The recommended test is a deliberate comparison between users with different legitimate scopes. Select a request with a known expected result for each role. Run it under those roles rather than sharing one administrator's demonstration account. Confirm that the authorized user can obtain the intended information and that a restricted user does not receive information outside the expected boundary. Record limitations in interpreting an incomplete answer, especially when the underlying question implies a count or comparison across records.

This is not a reason to weaken Maximo permissions for a more impressive demonstration. If a useful business question cannot be answered within a technician's current scope, that is an access-design discussion. It should be resolved through the normal security process. The assistant should not become the pretext for granting broad access that the same user would not receive in the ordinary application.

Disconnected deployment introduces a different limitation. IBM says docsearch is enabled by default and that its responses typically contain reference links to IBM Documentation. The enablement and architecture pages state that these links do not work in on-premises or air-gapped environments. A successful response can therefore include an unusable supporting reference in precisely the environment where the programme is trying to establish a self-contained service.

Do not overstate that limitation as proof that the entire assistant cannot work offline. The documented issue concerns the reference links. Equally, do not dismiss it as cosmetic. Users who are told to verify an answer through a link need a verification path that actually exists inside their operating boundary. The proposed acceptance plan should include opening returned references from a representative user workstation in the actual network environment.

Where references are unavailable, define an approved alternative verification process and explain the limitation during onboarding. Do not imply that IBM's curated search automatically includes the organization's own procedures or a locally maintained copy of every cited page. The opened documentation does not establish either claim. The operational objective is simple: users should know whether an answer is based on an authorized record, a curated documentation search, or general knowledge, and how they are expected to check it before relying on it.

5. Separate configuration controls from emergency guarantees

The enablement instructions say administrators must deactivate the model before editing quick starters or other configuration settings. They also state that the assistant does not currently include a feedback mechanism. These are concrete product behaviors worth including in the operating plan. They are not, by themselves, a complete incident response framework or proof of an instantaneous stop mechanism for work already in progress.

A sound control review separates ordinary administration from emergency containment. Ordinary administration covers approved configuration changes, readiness checks, and regression testing. Emergency containment asks different questions: who can suspend access, which component is affected, what happens to active requests, and what evidence confirms the intended restriction. The opened pages do not answer all of those questions, so a design review should mark them for validation rather than filling the gaps with confident language.

In particular, do not describe model deactivation as a verified kill switch without testing its scope in the proposed deployment. The documentation supports its role as a prerequisite for editing configuration. Whether it interrupts an active request, prevents only new requests, or interacts with tool refresh behavior needs evidence from the relevant supported implementation. Those distinctions matter when an incident is already underway.

Create a controlled rehearsal in a nonproduction environment. Establish a baseline supported request, apply an approved administrative restriction, and observe the result. Preserve the timing and the exact control used. Restore the test configuration through the normal change process and repeat the supported request. This is a recommendation for learning the control boundary, not an instruction to disrupt a production service or experiment with unsupported operations.

The missing feedback mechanism also deserves a practical response. Establish an internal channel for users to report incorrect, incomplete, or unverifiable answers. Ask for the request, relevant business context, expected scope, and a safe description of the discrepancy. Avoid collecting unnecessary sensitive record contents in a general support channel. Assign an owner who can distinguish a data problem from an access issue, a configuration issue, or a response-quality problem.

A useful operating review then asks whether reports lead to reproducible tests. If a technician says an asset history answer was incomplete, the team should be able to reconstruct the permitted data scope and compare the answer with the underlying record. Without that discipline, feedback becomes anecdote. With it, limitations can become documented operating rules or targeted fixes. The purpose is not to promise perfect answers. It is to ensure the organization can investigate mistakes without treating a fluent response as evidence that the system behaved correctly.

Practical Implications

For a platform owner, the immediate deliverable is a deployment manifest linked to the IBM prerequisite page. Include the AI Service patch condition, model arrangement, hosting boundary, and proposed resources. Do not approve procurement from a release-family label alone. Keep the capacity test separate so that a documented minimum is not mistaken for measured production throughput.

For a Maximo administrator, build a use-case-to-tool inventory. Record active configurations, their readiness, the assistant's available tools, and the refresh considerations that apply to changes. Use representative authorized records to prove that a response came through a functioning operational path rather than accepting chat availability as the sole success criterion.

For security and support teams, establish role-based tests and an answer-investigation process before widening access. Include a disconnected-reference test where the environment requires it. Validate the scope of administrative restrictions through approved nonproduction exercises, and document unresolved behavior instead of naming an untested control an emergency stop.

For programme sponsors, make the pilot promise narrow enough to verify. Name the users, supported questions, data scope, and environment. Expansion should follow evidence from those conditions. A smaller initial service with a clear acceptance record is more defensible than an impressive demonstration whose hardware, permissions, and tool assumptions cannot be reconstructed.

Bottom Line

The nine-H100 versus one-H100 distinction is valuable because it forces a better opening conversation. IBM's documentation makes AI Service patch level part of the on-premises and air-gapped deployment prerequisite. That decision belongs upstream of procurement and upstream of a promise to deliver an assistant to the business.

The rest of the design must be equally explicit. Prove the service chain behind the chat window. Confirm the tools and account for refresh behavior. Test users with different permissions. Verify reference links inside the actual network boundary. Distinguish documented configuration behavior from emergency controls whose effect still requires validation.

None of those practices requires inventing a benchmark or assuming that a model change guarantees production performance. They require a traceable relationship between the product documentation, the configured environment, and the tasks the organization expects people to perform. That relationship is the difference between an agentic demonstration and a service that operations can responsibly accept.