Three things you can't
have taken back.

Rented AI lives on someone else's account, behind someone else's API, governed by someone else's policy. It can be suspended, deprecated mid-project, or repriced out from under your business — and you have no recourse, because you were never the customer that mattered.

Sovereign AI runs open-weight models locally, served through a standard runtime on hardware you control. The harness — the orchestration and control layer — abstracts the model away, so you are never bound to a single model. When a better or replacement open-weight model ships, the harness adopts it without you re-architecting anything.

Nobody can suspend your account, because there is no account to suspend. Nobody can deprecate your model, because the next one slots in underneath you.

What this looks like technically

• Open-weight models, served locally — abstracted behind the harness so the choice of model is yours and is swappable by task.
• No external dependency on the critical path — the inference that runs your business runs on your side of the wire.
• Model migration as a service — when the frontier moves, the maintenance relationship moves you with it. That's a feature you're paying for, not a tax.

The stack is built on permissively-licensed software — MIT and Apache 2.0 at the core. That is not a marketing detail; it is the legal reason the stack is yours to keep. Permissive licenses hand over full freedom: run it, modify it, keep it, with no seat to cancel and no rug to pull.

The open core is genuine — a usable harness that runs a real single-node stack, the hardened-base scripts, full documentation, and a starter template set. The high-value layers that decay — current harness releases, security hardening, model migration, the premium template library — are the paid, continuously-maintained part.

A member or the co-op may fork the open core. Doing so transfers the entire maintenance burden onto them — the exact burden they joined to avoid. The fork is legal and self-defeating by design.

The open-core split

On rented inference, your data doesn't sit still. Every record your agents touch can become training material — sharpening the very product your competitor rents next quarter. You pay a subscription, and the price includes handing over your edge.

Sovereign Data runs on your own private VPS or hardware. The integration layer — MCP and connectors — reaches your files, databases, internal services, and tools without that data ever crossing to a third party. It does real work against real systems, on your side of the wire.

It never leaves ground you control, never trains someone else's model, and never becomes the product.

Why this is structural, not a promise

• The compute is fungible and yours — your own hardware, commodity bare-metal arranged through the co-op, or a mix. No provider holds your data hostage.
• Confidentiality and residency are settled by location, not policy — if the data physically stays on your hardware, no terms-of-service change can reach it.
• The "get off rented inference" audit quantifies your current exposure — what breaks if the account is suspended, what a price increase costs, what confidentiality or residency risk you already carry.