← R&D Log
2026-11-03HPC6 min read

Sovereign compute: what a single-tenant GPU cluster guarantees

Sovereign AI is about where data lives, who reaches the hardware, and who owns the weights. What a dedicated cluster changes, and what it does not.


"Sovereign AI" gets used to mean almost anything, so it's worth being specific. For an organisation training models on proprietary data, the question is rarely whether a GPU is fast enough. It's whether the data, the training runs, and the resulting weights stay under the organisation's control. Shared cloud capacity makes that hard to prove, even when the provider is careful.

What "single-tenant" changes

A single-tenant cluster is dedicated hardware for one client. Nothing else runs on those nodes, and the storage, network segment, and credentials belong to that client's environment. Three things become true that are harder to establish on shared infrastructure:

  • Isolation you can audit. The question "who else could read this disk" has a short answer: nobody outside the client's own access list.
  • Weights you own. Checkpoints are written to the client's storage and versioned there. Nothing is pushed to a shared registry by default.
  • Air-gapping as an option. A cluster can run with no route to the public internet, which matters for some defence and public-sector workloads.

Our featured engagement used this pattern: an 8-GPU cluster with a custom REST API, fine-tuning vision-language and image-generation models on the client's internal digital assets, inside an air-gapped environment.

What it does not change

Dedicated hardware does not make a project secure by itself. Three things still need work:

  1. Access control. Who can submit jobs, read datasets, and export weights is a policy decision. It has to be written down and enforced in the pipeline, not assumed.
  2. Supply chain. Container images, Python packages, and model weights from outside the boundary are a common route for problems. Pin versions and scan them before they enter the cluster.
  3. Operations. A sovereign cluster that nobody patches is a liability. Security patching and drift monitoring belong in the contract from day one.

The procurement question to ask

When a buyer asks for sovereign compute, we suggest three concrete questions:

  • Where does each class of data physically sit, and who holds the keys?
  • What leaves the boundary, and how is that logged?
  • If the vendor disappears, can the client keep running the cluster and the models?

The third question is the one most often skipped. A sovereign setup should survive the loss of its builder. That means documented infrastructure-as-code, the client holding the credentials, and a model registry the client can export at any time.

Where the line sits

Single-tenant hardware is the right answer when the data is sensitive and the training is ongoing. For a one-off experiment on public data, a managed cloud service is usually cheaper and faster to start. The point of sovereignty is control, and control is only worth its cost when the data needs it.


Clean-room write-up — public / synthetic data only.