Data sovereignty
Your data never leaves your machine.
Not a policy commitment. An architectural fact.
Marigold runs open-weight models on your own hardware. There is no vendor in the inference path to send data to, misconfigure a policy around, or have a breach on your behalf. Once set up, it runs airgapped -- inference makes no external request at all.
Sending data to a commercial AI API is, structurally, the same decision as sending it to any third-party processor. Self-hosting removes that decision rather than managing it.
Three kinds of risk, one architectural fix
01
Regulatory risk
NHS data frameworks, FCA conduct rules, GDPR restrictions on international transfer, legal professional privilege -- binary, externally imposed constraints. If your data can't leave the building, running the model inside the building is the only architecture that satisfies the rule outright.
02
Exposure risk
Not everything sensitive is regulated -- competitive intelligence, unannounced product work, client relationships. A commercial provider's breach becomes your breach the moment their processor held your data. With nothing sent externally, there's no third party left to have the breach.
03
Training reappearance risk
Commercial providers train on submitted data, and demonstrated cases exist of training data reappearing in outputs. Open-weight models are fixed and published before you run them -- there is no feedback loop, because there is no vendor collecting anything to feed back.
What actually leaves the building
| Criterion | Self-hosted (Marigold) | Commercial AI API |
|---|---|---|
| Inference location | Your own hardware | Provider's infrastructure |
| Prompt sent to a third party | Never -- none in the path | Always |
| Training on your inputs | Never -- fixed published weights | Varies, opt-out often required |
| Third-party data processing agreement | Not applicable -- no processor exists | Available, still a third party |
| Works fully offline after setup | Yes | No |
| Open-weight models | Yes | No |
Frequently asked
Does self-hosting remove the need for a data processing agreement?
There's no third-party processor in the inference path, so there's nothing to sign a DPA with for that part of the pipeline. You remain sole controller and sole processor of your own inference. That's an architectural fact, not a legal guarantee -- it doesn't by itself make a given use of data lawful, which still depends on your legal basis and sector.
Does Marigold train on my data?
No. Marigold runs open-weight models as published -- it doesn't train, fine-tune, or send inference inputs anywhere for that purpose. There's no feedback loop between what you run and what any model learns next, because there's no vendor on the other end collecting it.
What data actually leaves my machine when I use Marigold?
None, once set up. Models download once to local disk. After that, inference runs entirely on your own hardware and the stack can be disconnected from the network entirely -- see the setup tutorial for the exact environment variables involved.
Is self-hosting alone enough to be GDPR compliant?
It resolves the cross-border transfer and third-party processor questions, which are usually the hardest part. It doesn't by itself make every use of personal data lawful -- that still depends on your legal basis, retention practice, and sector rules. Marigold does not provide legal advice.
For a fuller treatment of these three risk categories in general -- not specific to Marigold -- Bay Information Systems has written on it at greater length: Why Private Inference?
Nothing to send. Nothing to leak.
Open source. Clone it, run it on your own hardware.