IBM Bob self-hosted for IBM i: you need an x86 cluster first

IBM Bob self-hosted deployment went generally available on September 24, and IBM announced it on October 1. For IBM i shops whose security team said no to an AI assistant that ships source code to someone else’s cloud, that sounds like the answer. For some of them it is. For most small IBM i shops it is not, and the reason is near the top of IBM’s own system requirements page: the Bob backend runs only on x86 OpenShift, and your IBM i runs on Power.

This piece covers what self-hosted Bob actually moves inside your walls, what the cluster and the model cost you in hardware, and which kind of shop should start a proof of concept versus wait.

What self-hosting moves, and what it does not

The IBM Bob Premium Package for i has two network hops. The developer’s IDE talks to the IBM i over SSH through the Code for IBM i extension, which is why IBM’s prerequisites list IBM i 7.4 or later, the SSH daemon, and 5733-SC1. Then the IDE sends prompts, with your source in them, to the Bob backend, which forwards them to a model.

Developer PC (Bob IDE + IBM i extension)
   |  SSH via Code for IBM i
   v
IBM i (7.4+, STRTCPSVR *SSHD)        <- unchanged by self-hosting
   
Developer PC
   |  HTTPS, prompts + source context
   v
Bob backend  --->  model endpoint     <- this is the hop self-hosting moves
(SaaS: IBM's cloud | self-hosted: your OpenShift + your model)

Self-hosting only changes the second hop. The first one stays exactly as it was. Bob still signs on to the IBM i with a real user profile, still compiles, still runs SQL. So a self-hosted deployment does not get you out of the profile and authority review covered in our IBM i AI security piece. It only answers the question of where the source goes after it leaves the box.

That is still the question that blocks most approvals, so it matters. But be clear with your security team about which risk you are retiring.

The x86 problem with self-hosted Bob on IBM i

IBM’s system requirements page is blunt: “The Bob on-premises backend requires an amd64 (x86_64) OCP cluster.” Mixed-architecture clusters are allowed only if you pin Bob’s workloads to x86 nodes yourself, and IBM notes that Bob does not apply those constraints automatically. Supported OpenShift versions are 4.20 through 4.22.

Here is what IBM publishes for the Bob backend alone, with no model serving included:

Configuration	CPU	Memory	Storage
Bob Core (raw)	22.1 vCPU	35.1 GiB	~30 GiB
Bob Core + RAG (raw)	38.1 vCPU	69.1 GiB	~62 GiB
Bob Core production, with headroom	~36.5 vCPU	~53.4 GiB	~50 GiB
IBM’s minimum reference cluster	~84 vCPU across 9 nodes	~168 GiB	600 GiB

The reference cluster is three control plane nodes, three infrastructure nodes and three workers. Storage has to be managed NFS or OpenShift Data Foundation, PostgreSQL wants SSD-backed block storage, and you need a separate backup target. A single-node OpenShift install is enough for a proof of concept, according to IBM’s release post.

If your company already runs OpenShift on x86 for other workloads, Bob is one more namespace and this table is a capacity request. If your whole estate is a Power server, a tape library and some Windows file servers, this table is a new platform you have to buy, learn, patch and back up. That is the line that splits IBM i shops into two groups.

Picking a model: your cloud account or your GPUs

Self-hosted Bob runs exactly one core model at a time. IBM’s gateway docs say configuring more than one “is not supported and results in undefined behavior.” The SaaS version routes between models. The self-hosted version does not.

You have two routes.

Frontier model through your own cloud account. Claude Sonnet 5 or Opus 4.8 on AWS Bedrock, Gemini 3.7 Flash on Vertex AI, or GPT 5.6 Sol on Azure OpenAI. IBM’s release post is precise about what this means: model requests, “including the code context they carry,” go to your cloud account under agreements you already have. Your source does leave the building. It just goes to a provider your legal team already approved instead of to IBM.

Open-weight model on your own GPUs. For networks with no outbound connection, IBM supports NVIDIA Nemotron 3 Ultra and Poolside Laguna S 2.1 (a third, Mistral Medium 3.5, is restricted to the Z package). The hardware gap between those two is large:

Nemotron 3 Ultra is a 550B-parameter model. NVIDIA’s NIM support matrix lists eight H200 or eight B200 GPUs for full precision, and four H200s at the smallest quantized setting it supports on that card.
Laguna S 2.1 is 118B parameters with about 8B active per token. Poolside’s model card puts full-precision weights at roughly 236 GB across multiple GPUs, with FP8, NVFP4 and INT4 versions that cut that substantially. On Poolside’s own benchmark table it scores 70.2% on Terminal-Bench 2.1 against Nemotron’s 56.4%.

Add a guardrail model on top. IBM strongly recommends one and names openai/gpt-oss-20b for air-gapped installs.

Two caveats matter more for IBM i than for anyone else. First, IBM’s own model page says frontier models give the best results on complex mainframe work and describes the open-weight models as suited to “standard coding assistance.” IBM says nothing about IBM i either way, and nobody has published how Laguna or Nemotron handle fixed-format RPG, CL or DDS. Treat that as unknown, not as bad. Second, neither local model accepts images, so pasting a screenshot of a green screen into the chat stops working.

What you give up versus SaaS Bob

IBM’s known-limitations and overview pages list the trade plainly:

No in-place upgrades. Moving to a new release means a fresh install, with hotfixes applied by re-running the installer. Budget a rebuild every release.
Your security logging. “Security event logging and monitoring for Bob self-hosted are managed at the OpenShift platform level and are not provided by Bob.” Activity logs are not in the Admin UI either. You read pod logs.
No Bobalytics. The usage analytics view is SaaS-only. On an air-gapped install the link goes nowhere.
Manual metering on air-gapped clusters. Bob reports one number to IBM, the count of active users. Connected clusters send it daily. Disconnected ones require someone to download a payload and upload it to IBM Software Central.

There is also an open question specific to IBM i. The Premium Package for i includes a documentation search tool that queries indexed IBM i manuals and Redbooks, and it is one of the better reasons to pay for the IBM i package over a generic assistant. IBM’s release post says the IBM i package works self-hosted. But the self-hosted entitlements page says “all available Premium Package add-ons extend Bob with IBM Z” and documents the backend RAG component under the Z package, describing it as covering a “Z/I-specific corpus.” Read together, the IBM i doc search probably needs the RAG add-on enabled at install time, which is the 16 vCPU and 34 GiB difference in the table above. Probably is not good enough for a purchase. Get it in writing.

Which situation are you in

You run IBM i alongside a larger x86 estate that already has OpenShift. This is who IBM built it for. Point a single-node proof of concept at a model endpoint you already pay for, connect one IBM i developer, and test it on a program everyone knows. Your security team gets identity through your own Active Directory, audit logs on your cluster, and source going only to a cloud account under your contract.

Your policy says source code never leaves the network, and you have no GPUs. Self-hosted Bob can satisfy the policy, but only with local hardware. Eight H200s for a team of six RPG developers is not a serious proposal. Laguna on a quantized build is the plausible path, and it is the one you would need to test hardest on your own code before anyone signs a GPU purchase order.

You are a Power-only shop with a handful of developers. Self-hosted is not your product. You would be standing up an x86 container platform to run a coding assistant, and the platform would outlive the reason you bought it. Your realistic options are to get IBM’s SaaS data terms in front of whoever owns the policy, or to evaluate SaaS Bob on code you are allowed to share. The $20 route into the IBM i package is still the cheapest honest test. If the answer is still no, documenting programs and capturing what your senior developers know does not have to wait for Bob. The IBM i AI Field Guide covers exactly that work with 35 IBM i-specific prompts and templates, and a prompt runs in whichever assistant your policy already allows.

Questions to put to IBM before a proof of concept

Send these in writing. The answers are not on any public page today.

What is the self-hosted license price per developer, and is the IBM i package priced the same as the SaaS $40 a month?
Does the IBM i documentation search work on self-hosted, and does it require rag.enabled: true at install time?
Which model has IBM tested the Premium Package for i against on self-hosted, and can we see the results on RPG and CL tasks?
What GPU sizing does IBM recommend per 10 concurrent developers for Laguna S 2.1?
If a later release adds a stronger local model, is it covered by our license?

The September release also removed HTTP+SSE transport from Bob’s MCP client. If your team connects Bob to an IBM i MCP server, as in our Copilot computer use piece, confirm the server speaks Streamable HTTP or stdio before you upgrade.

Bottom line

Self-hosted Bob is a real answer for IBM i teams that live inside a larger company with an x86 OpenShift platform and an approved cloud model account. For them, start the proof of concept this quarter. For the much larger group of IBM i shops that run a Power server and not much else, it is an x86 infrastructure project dressed up as a coding assistant. Do not buy a cluster to run an assistant. Test SaaS Bob on code you can share, and revisit self-hosted when IBM publishes a price and an IBM i result on a local model.

Frequently Asked Questions

Can IBM Bob run on-premises for an IBM i shop?

Yes, since September 24, 2026, but the backend runs on a Red Hat OpenShift cluster with amd64 (x86) worker nodes, not on Power. IBM’s minimum reference for a dedicated cluster is nine nodes, about 84 vCPU and 168 GiB, before you add anything to serve the model.

Does self-hosted IBM Bob include the IBM i premium package?

IBM’s release post says the IBM i package works in a self-hosted deployment and is assigned in the Admin UI. Its entitlements page only documents Z add-ons, so confirm in writing whether the IBM i documentation search (RAG) is available on your install before you sign.

How much does self-hosted IBM Bob cost?

IBM has not published a price. Self-hosted is part of the Enterprise plan and is sold through IBM sales or a Business Partner. The public SaaS prices are $20, $60 and $200 a month for Pro, Pro+ and Ultra, plus $40 a month for the IBM i package.

Which AI models can self-hosted Bob use?

One core model at a time. On an air-gapped network that is NVIDIA Nemotron 3 Ultra or Poolside Laguna S 2.1 on your own GPUs. With outbound access, it can be Claude Sonnet 5 or Opus 4.8 on AWS Bedrock, Gemini 3.7 Flash on Vertex AI, or GPT 5.6 Sol on Azure, billed to your own cloud account.