Workload Placement

AI Infrastructure for IBM i Shops

An AI project touching IBM i does not necessarily execute in an IBM i partition. Decide where the model, agent, and data connection run before selecting Power hardware.

The first infrastructure question is not 'Which AI server should we buy?' It is 'Where will this workload actually execute?' An IBM i application can supply the data while an agent, model, or inference service runs in a different partition or outside the Power system entirely. Keeping those roles separate prevents a software integration project from becoming an unnecessary hardware purchase.

Separate the IBM i system from the inference runtime

Three placement decisions

ChoiceWhat to verify
External AI serviceData access, residency, network latency, governance, recurring cost, and the supported connector or API. IBM Bob or watsonx adoption alone does not imply a Power upgrade.
Power CPU inferenceSupported operating system and framework, available cores and memory, model size, latency, and whether on-chip Matrix Math Acceleration benefits the actual workload.
Power 11 with SpyreSupported Power 11 machine type, ENZ0 expansion drawer, Red Hat inference stack, Linux partition resources, accelerator allocation, and licensing. This is a distinct configuration, not on-chip MMA.

What changed in 2026

IBM announced the one-socket Power S1112 on July 15, 2026, describing local AI inference using Power 11 on-chip Matrix Math Acceleration. IBM's release also quotes a customer exploring Linux partitions alongside IBM i. That is a useful placement example, not a promise that every IBM i application or model will run natively inside IBM i. IBM Power Autonomous Operations and IBM Bob were part of the same announcement, but they address system management and development rather than the inference runtime.

Source: IBM's July 15, 2026 Power announcement. IBM listed September 23, 2026 as the expected general-availability date for Power Autonomous Operations; confirm current status before treating it as shipping.

A practical placement sequence

1. Name the workloadCode assistance, retrieval, scoring, local inference, or system automation are not interchangeable.
2. Identify the runtimeRecord the operating system, partition, framework, and model requirements separately from the IBM i data source.
3. Measure the connectionTest data movement, response time, access control, and failure handling with a narrow pilot.
4. Size only after a pilotUse measured throughput and headroom. Verify machine compatibility, support, and licensing against current IBM documents.

For the compact Power option, see Power S1112 AI inference. For a larger accelerator configuration, see Spyre prerequisites. Software product and integration questions belong on AI.as400Software.com; IBM i storage questions belong on FlashSystem for IBM i and Power.