Self-hosted AI sovereignty depends on the complete control stack

IBM has made the Bob software-development agent generally available as a self-hosted backend on Red Hat OpenShift. An organisation can operate that backend on premises or in its own cloud account, preserve identity, audit logging and usage metering inside the cluster, and connect either to open-weight models on its own GPUs or to an approved external model service.[1][2][3]

This is a meaningful expansion for codebases that cannot be sent through a vendor-operated SaaS product. It is also a useful reminder that “self-hosted” describes a deployment pattern, not the complete answer to sovereignty. A system can run in a customer cluster while still depending on external models, vendor updates, licensed components or administrators outside the organisation’s legal and operational control.

Data location is one layer of control

IBM says a supported self-hosted model configuration can keep code, development context and build artifacts within the customer-managed environment. In the hybrid route, however, model requests—including the code context they carry—go to the organisation’s own cloud-model account. That can satisfy a policy when the account and region are already approved, but it is not the same as an air-gapped deployment.[1][2]

Running the system creates new obligations

The fully self-hosted route transfers more responsibility to the customer. IBM’s documentation points to model-specific hardware sizing because requirements depend on quantisation, context length and concurrent developers. The customer must also operate OpenShift, GPUs, identity, storage, monitoring, upgrades, vulnerability response and capacity planning. Control and operational burden arrive together.[2]

A cloud service can still be the safer choice for a low-risk workload if it is updated quickly, monitored well and covered by an appropriate agreement. Conversely, an air-gapped deployment can remain insecure if its packages are stale, privileged accounts are poorly controlled or outputs enter production without review. The architecture should follow the workload’s threat model rather than a slogan.

Measure outcomes as well as control

IBM’s initial Bob launch reported adoption by more than 80,000 employees and a 45% average self-reported productivity gain among surveyed users. Those figures are vendor-reported and do not establish what a self-hosted customer will achieve. A production evaluation should track accepted changes, review effort, defects, policy violations, infrastructure cost and time to recover from a failed agent action.[4]

The stronger case for self-hosting is therefore specific: it enables AI-assisted development for workloads that were previously excluded by data-residency or network requirements. The larger claim—AI sovereignty—must be earned across the complete control stack. Keeping the code inside is necessary for some organisations; knowing everything else the system depends on is what makes that control real.

Report an error or suggest a correction