AI on IBM i FAQ
IBM i AI questions with an infrastructure focus: Power inference, partitions, accelerator requirements, and the handoff to software product coverage.
Do we need Power 11 hardware to use IBM Bob?
Not solely because you chose IBM Bob. Bob is a coding and modernization software decision, not a local inference sizing result. Check the current IBM Bob package requirements with IBM, then evaluate Power hardware only if a separate workload, capacity need, or supported local model calls for it.
Does Power S1112 run an AI model inside the IBM i partition?
IBM describes local inference through Power 11 on-chip Matrix Math Acceleration, but that does not establish that every model runs inside IBM i. IBM's July 2026 announcement gives a customer example of Linux partitions alongside IBM i. Confirm the model's supported runtime and partition before planning data access or sizing the S1112.
Is IBM Spyre the same as on-chip Matrix Math Acceleration?
No. Matrix Math Acceleration is a processor feature. Spyre Accelerator for Power is a PCIe-attached card in supported Power 11 configurations with an ENZ0 expansion drawer and a Red Hat inference environment. One does not imply support for the other, and neither is an automatic requirement for IBM Bob or watsonx adoption.
What should we measure before a Power AI upgrade?
Measure the actual model and runtime, transaction volume, response time, memory and core demand, production IBM i headroom, data-transfer latency, recovery reserve, and supported system configuration. A marketing benchmark or a product announcement is not a workload-specific capacity plan.
Is every FlashSystem.ai configuration quantum safe?
Do not infer a portfolio-wide quantum-safe guarantee from the FlashSystem.ai name or a search result. IBM's February 2026 announcement describes AI-assisted storage administration, ransomware detection, and new 5600, 7600, and 9600 systems. A quantum-safety claim needs the current IBM documentation for the exact model, firmware, encryption, protocol, and use case. Storage model and FlashSystem.ai coverage belong on the dedicated FlashSystem site.
What does a Spyre Accelerator card need before it will run anything?
Five things, and you need all five. The card itself. An ENZ0 PCIe4 expansion drawer, because the card does not go in the server body. Red Hat Enterprise Linux 9.6, 9.8, or 10.2 in a Linux partition. Red Hat AI Inference Server or Red Hat OpenShift AI, which IBM names as required software. And a model for the card to run, which for most shops means IBM's open Granite family. Most first budgets contain only the card.
What is IBM Granite and does it cost anything?
Granite is IBM's own family of open language models, published under the Apache 2.0 licence. The model weights cost nothing: you can download them, tune them, and run them in production with no per-user fee and no token meter. What costs money is everything around them, meaning the hardware, the Red Hat subscription when a Spyre card is involved, and the people who keep it running. Granite 4.0 arrived in October 2025 and Granite 4.2 in August 2026.
Why would a model's memory use change how much hardware I buy?
Because a Spyre card has 128 GB of memory and that number does not move. A model holds a running memory of each conversation, and the more memory each conversation eats, the fewer people one card can serve. IBM reports that the Granite 4.0 hybrid design cut memory use by more than 70 percent for long-context and multi-session work. In practical terms that can be the difference between a project needing one card and needing four.
Can I run a Spyre card without buying Red Hat software?
Not in a supported configuration. IBM's Spyre for Power documentation names Red Hat AI Inference Server or Red Hat OpenShift AI, running on Red Hat Enterprise Linux. The engine underneath Red Hat's product is the open source vLLM project, which you can run yourself, but that is not what IBM supports for this card. The difference matters the first time you need to open a problem report.
Can I buy an IBM NorthPole or AIU chip?
No. Both are IBM Research work with no product, no price, and no announced arrival on an IBM Power roadmap. The AIU became the Spyre Accelerator, which you can order. NorthPole remains a research architecture. If a proposal cites either as a reason to buy something this year, that is a sales technique rather than a roadmap.
Does IBM's partnership with Anthropic mean I need new hardware?
No. The October 2025 announcement covers Claude being integrated into IBM software, beginning with an IBM development environment, and by mid-2026 it was most visible inside IBM Bob. It is a software and subscription decision. The announcement does not name IBM Power, IBM Z, or IBM i anywhere in it, and nothing in it describes a server requirement.
What would an AS/400 shop use InstructLab for?
Capturing knowledge that only lives in somebody's head. InstructLab lets a person who knows the subject write a small number of examples, then generates a large synthetic training set from them and tunes an open model on the result. The work can run entirely on hardware you own, so your operational knowledge never leaves the building. It does not connect the model to Db2 for i, and somebody still has to sit down and write the first honest examples.
Should we wait for cheaper or more efficient AI chips before starting?
No, and the research is the reason why rather than an argument against. IBM announced a 0.7 nm transistor architecture in June 2026 with production possible as early as the next five years. That is genuine progress and it is also five years away at best. Waiting for better silicon is a loop with no exit. Start on the Matrix Math Acceleration already built into your Power 10 or Power 11 processor, which costs nothing extra, and find out whether the workload is real.