Sovereign by default

"Sovereign AI" has been stretched to cover almost anything: a dedicated instance in someone else's cloud, region pinning, a data processing agreement with a retention clause. We use the word, so we owe a definition precise enough to be falsified. Here it is: every prompt, every completion, every embedding, every log line, and every model weight stays on infrastructure Options Technology operates, and there is no third party anywhere in the data path. Not in inference, not in retrieval, not in telemetry.

A claim you can verify by inspection

The useful property of that definition is that it is checkable without legal review. There is no external inference API in the architecture. Every model call goes to a GPU cluster that Options Technology operates. No prompt ever leaves the environment because there is simply nowhere for it to go. Vector search for retrieval is a SQL query against the same per-organisation PostgreSQL database that holds the chat history, so embeddings never travel to a separate service either. Every model call passes through one gateway that Options operates, which is also where it gets audited.

A sovereignty claim that depends on a configuration file is a marketing claim.

This is the difference between private as a tier and private as a default. Most platforms offer privacy as a configuration: pick the right plan, sign the right paperwork, trust the toggle. Here every organisation runs on its own database, not a shared schema with a tenant_id column, and the system has no external inference endpoint to misconfigure. The strongest control is the one that is architecturally impossible to bypass.

What the definition rules out

Once nothing leaves, a whole category of risk simply isn't yours anymore:

• Pricing you don't set. Per-token rates can change under you. A GPU you provision through Options Technology costs what it costs, and the bill is legible.
• Deprecations you didn't choose. When a vendor retires a model, every prompt you tuned against it retires too. A model you host on your own cluster is deprecated when you decide it is, after you have validated its replacement.
• Someone else's retention policy as your retention policy. If a provider holds your prompts for thirty days, you hold your prompts for thirty days, whatever your own policy says. With nothing leaving, your retention policy is just true.
• Quiet model swaps. The weights you evaluated are the weights you serve, pinned in version control, until you change them.

None of these are hypotheticals. They are the ordinary terms of service of every hosted inference product, and they are fine terms for plenty of workloads. They are not fine when the prompts are client positions, patient records, or anything else where "where does this data go?" must have a one-word answer.

What it costs

This is the part vendors selling sovereignty tend to mumble through, and it is the part that makes the claim credible. Sovereignty is not free. It is a transfer of work from the vendor to you.

You provision hardware through Options Technology, and capacity planning is your problem now: too few and inference queues, too many and expensive silicon idles. You do the upgrades, the driver bumps, the CVE patching, the model rollouts, with no vendor quietly handling it behind an API. And you own the 3am page. When inference is down, there is no status page to point at. The status page is yours.

For most teams, most of the time, a hosted API is the right call, and pretending otherwise would undercut everything above. The trade tips when the data is regulated, when the workload is permanent enough that owning hardware beats renting tokens, or when "we'd have to ask the vendor" is an answer your auditors won't accept.

The test

Strip the word back to something operational and it survives: sovereign means every organisation gets its own database, every model call passes through one auditable gateway, and there is no external inference endpoint in the architecture. You can verify it by reading the service graph in an afternoon. We pay for that property in hardware, upgrades, and pages. That is the whole deal, stated plainly, and it is a deal we'd take again.