AS/400 and IBM i AI Infrastructure

Pick the hardware before you pick the AI tool

IBM Bob and watsonx are software choices. Where that work actually runs is a separate hardware choice with its own bill. Power 11 on-chip Matrix Math Acceleration, a Linux inference partition, and Spyre accelerator cards each answer a different question, and none of them require a new server by default.

Compare infrastructure options Check infrastructure readiness

Start with the workload

Identify the AI task, the runtime, and the data connection before choosing a server or accelerator.

AI workload placement

Compare cloud, an existing Power system, and a Linux inference partition before assuming AI must run inside IBM i.

Power S1112 local inference

IBM's one-socket Power 11 announcement puts on-chip MMA and adjacent Linux partitions in the conversation. Check the workload fit.

Spyre Accelerator prerequisites

A PCIe-attached accelerator for supported Power 11 configurations, with Red Hat and expansion-drawer requirements. It is not an IBM i-native card.

Infrastructure readiness

Inventory system headroom, partitions, data movement, storage, resilience, and licensing before a pilot.

IBM AI silicon specifications

Published figures for the Spyre Accelerator, on-chip Matrix Math Acceleration, Telum II, the AIU, NorthPole, and the 0.7 nm NanoStack research node.

The runtime the card requires

IBM names Red Hat AI Inference Server or Red Hat OpenShift AI as required for Spyre on Power 11. Here is that stack, plus the open Granite models it serves.

When another specialist site is the better answer

Hardware, software, and storage overlap in real projects. The coverage should not.