IBM Bob Goes Self-Hosted: Enterprise AI Rollout Moves Into the Customer Data Center

The fact: Bob moves into the customer data center

IBM is bringing Bob, its AI assistant for legacy code, inside the customer's data center. An enterprise AI product launch of this kind moves a procurement bid before it moves a benchmark.

The official release from IBM Germany^[1] describes the new self-hosted option, and the trade publication IT Jungle^[2] covered it on 5 October 2026.

The move closes a gap opened in early July 2026. Back then IBM had presented the premium Bob packages for IBM Z, for IBM i and for Java modernisation, as heise^[3] recounts. Those packages relied on external model services.

The flaw sat right there: with external services, code and application context can end up with the model provider.

Now the perimeter changes. Companies develop and modernise with AI while code, data and workflows stay inside their own controlled infrastructure. IBM names four environments: on premise, private cloud, sovereign cloud and air gapped systems.

What this option really is

The release carries one precise condition, and it matters more than the headline: the guarantee holds when the models also run inside that same infrastructure. The perimeter closes with two pieces, the assistant and the model.

Anyone who installs Bob in house and points the model at a public API is back at the starting line.

A distinction follows that marketing happily blurs: announcement and production availability remain two different things. The sources report a product development. Commissioning inside an air gapped environment remains a months-long project, with accelerators, licences and operational capacity to line up.

The second piece of IBM's message concerns compliance. Organisations working on sensitive source code, regulated data or critical infrastructure live under tight security requirements, and those requirements limit the use of public AI platforms. The self-hosted option turns that limit into a purchasing criterion.

From the model to the perimeter: the competitive shift

The code assistant market has competed for two years on suggestion quality. Whoever wrote the better patch won the bid.

This move shifts the axis. For a bank running COBOL under audit, the real question concerns location: where the code sits while the model reads it. Model quality slides to the second line of the evaluation grid.

It confirms a thesis this desk has repeated for months: the competitive moat in enterprise AI sits in deployment, before the model. The vendor that controls the place of execution holds the contract for years. The vendor with the highest score holds the demo.

IBM is playing at the table where it already has a seat: Z, i and Java migrations inside companies that buy hardware and support from the same supplier. Legacy modernisation becomes an open contest inside closed environments too. That home advantage weighs more than ten points on a public leaderboard.

Who feels the pressure

The pressure arrives on three fronts, and it touches three different categories of supplier.

• Code assistants sold as pure cloud services: they lose ground when the tender asks for a closed perimeter.
• Integrators selling legacy modernisation by the day rate.
• Generalist model providers: value slides towards the package that runs inside the customer's own house.

Integrators face the most urgent reaction. An assistant that reads COBOL and proposes the rewrite cuts billable hours, and the margin moves towards whoever sells project governance instead of day rates.

The game stays open for everyone. A self-hosted option also opens the door to open models installed in house, and that door counts for IBM's competitors as much as it counts for IBM.

The Armonk company's advantage sits in specialisation. COBOL, RPG and Java inside Z and i environments form narrow ground, and generalist competition covers it thinly. The enterprise competitive map reads like this: few vendors, long contracts, rare replacements.

For a Chief Digital Officer the quarter's work is a review of the supplier portfolio. The question to put in the tender concerns the perimeter of execution, with a written and verifiable answer.

Sovereignty as a line item

Two phrases appear in the release that a year ago lived in conference halls: sovereign cloud and air gapped. Today they sit on a price list.

The European market buys jurisdiction. A board signs more readily when data residency sits in the contract, and this confirms a position the desk has defended for a long time: sovereignty is a product, with a price and a datasheet.

The supplier map reads by flag, as well as by score. IBM arrives at that table with a good card: the European mainframe customer already owns the hardware, the support contract and a compliance office that knows the environment.

For an investor the signal counts twice. Demand for controlled environments supports the revenue of whoever sells infrastructure and managed services, and compresses the value of whoever sells generic access to a model.

The market thesis to verify next quarter concerns the mix: how much of new AI revenue comes from closed environments.

The CFO's arithmetic

Self-hosting moves the spend, instead of cutting it.

The cloud service licence leaves the books. Three items come in: accelerators, infrastructure operations and the staff keeping the internal model running. The price list for frontier models falls with every release, and compute remains the line that grows.

The calculation gets concrete: cost per developer of the premium package against the full cost of an internal environment with dedicated accelerators. The second item wins when the compliance constraint is real, because the alternative is giving up AI on legacy code.

One negotiating lever comes from the July price list. IBM keeps the premium packages for Z, for i and for Java modernisation in its catalogue, and the new installation option adds to that structure. Anyone negotiating today has three levers: the package, the perimeter and the support.

The budget to protect is the operations one, where spend accumulates month by month.

What to decide in the next 90 days

The quarter brings four decisions, one per role.

• Chief Strategy Officer: choose between a single perimeter supplier and an in-house assembly of assistant and model.
• CFO: review the spend line for code assistants and open a line for accelerators.
• Chief Digital Officer: ask every supplier for a written answer on where the model executes.
• Technology Investor: weight the portfolio towards whoever sells controlled environments.

The first move is a question to the current supplier. Ask where the model runs, who keeps the logs, which contract covers data residency and what it costs to bring everything in house.

The second move concerns time. An air gapped environment requires internal capability, so the useful window opens now, with a commissioning budget and a small team.

A feasibility test is enough to shift the bid. A pilot on a single COBOL application portfolio, with the internal model and the code staying in house, delivers the board the figure that counts: hours saved per modernised function.

The market has moved. For anyone with legacy under audit, AI-assisted modernisation passes from the wish list to the tender list, and the selection criterion speaks of location before score.