Digital transformation practice

Private & sovereign AI: systems inside your boundary

Some data cannot leave the building. We run open-weight models on hardware you control, with the audit trail intact. Week 1 ends with a plan and a price, in writing.

A walled paper courtyard with a padlocked vermilion gate, a stack of paper server cabinets and an open ledger inside.

What the service covers

Sovereignty belongs to the whole system: the model, the prompt logs and the embeddings all stay inside your boundary, and every part can be audited.

  • An architecture document

    The cluster layout, the model choice, the throughput sums and the failure modes, each with its reasons.

  • A reference deployment

    The model served inside your network, with monitoring, autoscaling and logs from the first deploy.

  • An audit log and a model card

    Every prompt, answer and model decision logged in order, and a model card ready for your regulator and your customers.

  • An operations runbook

    Patching, rollback and capacity planning written down, so your team can run the cluster without us.

3 ways to engage

Pick the one closest to where you are. Week 1 settles the scope, in writing.

  1. 1What you get

    Feasibility review

    A written go or no-go on private deployment, read against your data, your rules and your needs, with the rough cost if you go.

    How it runs

    One week of discovery that ends with a plan and a price, in writing.

  2. 2What you get

    The private build

    Architecture, deployment, monitoring and a runbook, built with your platform team inside your network.

    How it runs

    Fixed scope and fixed cost, agreed before work begins, with a named person on every gate.

  3. 3What you get

    The embedded team

    A senior team on model updates, capacity, patching and escalations once the cluster is live.

    How it runs

    In your accounts and on your direction. Documented handover, or we keep operating it with you.

Who the service is for

Private & sovereign AI / Fit check

It fits

  • Banks, insurers and public bodies whose regulators expect every AI decision to be logged and explained.
  • Hospitals and health firms holding patient records that may not leave their own systems.
  • Firms with European customers whose contracts fix where the data must live.

It doesn’t fit

  • Teams whose only reason is a vague fear of foreign clouds. A hybrid setup costs less.
  • Anyone without a realistic budget for the hardware and for running it.
  • Light use where a hosted model costs less and the data is free to leave.

Questions about this service

Every engagement runs the same method: an agent proposes, a person approves.

Which models do you recommend?
Open-weight models, chosen on capability and on what the licence allows. We test a shortlist on your real work rather than public benchmarks, and the hardware budget settles the rest. The choice and its reasons go in the decision log.
How much hardware do we need?
We work it out for your workload: how many tokens, how many people at once, and how fast each answer must come back. The plan also says what happens at peak load and during a model swap.
How do you handle model updates?
Like a database migration. A new model takes a small share of traffic first, the evaluation suite runs against it, a regression stops the rollout, and the rollback plan is written before the deploy. With open weights, you choose the timing.
Is it cheaper than a hosted model?
It depends on volume. At low volume a hosted model usually costs less. At high volume your own hardware costs less per answer, and the data stays where it must. The feasibility review models it on your workload.
Air-gapped or hybrid?
Both work. Air-gapped, the cluster and the model registry sit inside your own network and updates arrive as signed packages. Hybrid, the control plane runs in your cloud and the data stays in a region that meets your residency rules.