Approved partner · Anthropic & OpenAI
European AI pipelines,at every level of control.
EU-resident hosting to fully self-hosted
The tier is chosen per workflow
S.01 · The Answer First
Yes, your AI can run entirely in Europe.
Clients increasingly ask for AI pipelines that run entirely in Europe: European inference, EU-only data flows, European providers where the jurisdiction matters. The short answer is yes. Nugawi builds AI pipelines at the level of control your operation requires, from EU-resident managed hosting to fully self-hosted open-weight models, and the tier is chosen per workflow, not per company. The evaluation harness proves the chosen model on your actual work before anything ships, whatever tier it runs on.
What is a sovereign AI pipeline?
An AI pipeline where the data, the processing and the control stay where your governance requires them: European infrastructure, European jurisdictions, and a model you either route to a provider you have chosen or hold yourself.
The brief usually names one or more of these.
- 01
Data residency: prompts, documents and outputs stay in European regions, documented for the GDPR.
- 02
A European provider for the workflows where the counterparty jurisdiction matters.
- 03
Verifiably EU-only processing for a defined set of regulated workflows.
- 04
EU AI Act risk classification, documentation and audit trails as part of delivery.
- 05
The ability to change model, provider or hosting tier later without rebuilding the pipeline.
- 06
The same evaluation evidence whichever model runs.
S.02 · The Control Ladder
Three tiers of control, chosen per workflow.
Each rung buys a different thing, and the difference matters in front of a regulator. Here is the ladder, tier by tier, and what each one actually delivers.
01EU-resident managed hosting
Your pipeline runs on Nugawi's managed EU platform, or on AWS with Amazon Bedrock data residency in UK and EU regions. Prompts, documents and outputs are stored and processed in-region, with enterprise-grade operations, monitoring and rollback behind them. The residency is real and documented; the underlying provider is US-headquartered, which some buyers accept and others cannot.
02EU-pinned inference, European provider
Inference moves to a European provider: a European company, SLA-backed service tiers, and the inference step pinned to European infrastructure. Precision matters here. This tier keeps inference in-region, which answers most GDPR data-residency questions, while limited safeguarded sub-processor transfers remain possible under the provider's terms. We document exactly what the tier covers and what it does not.
03Verifiably EU-only, self-hosted
Open-weight models, such as Mistral's open models, are deployed on European infrastructure you control: on-premises, in your own cloud region, or fully isolated from the network. No data ever crosses the client boundary, no provider can withdraw the model, and you keep the weights themselves. This is the only tier we describe as verifiably EU-only, because you operate every part of it.
S.03 · Governance and Compliance
Compliance ships with the build.
European requirements are not a blocker on any tier of the ladder; they are part of what Nugawi delivers. GDPR data-residency requirements are met by design on every tier, and EU AI Act obligations are handled inside delivery, not as a separate workstream. The risk classification, the documentation and the audit trails ship with the build, whichever tier it runs on.
- 01
GDPR data residency: documented regions, transfer mappings and processing records on every tier, including the sub-processor position of the tier you choose.
- 02
EU AI Act risk classification: every workflow classified against its obligations before it ships, so the risk tier is known rather than discovered later.
- 03
Documentation: architecture, data flows, model provenance and human oversight written down to survive an audit.
- 04
Audit trails: who requested each action, release history with approvals, and evaluation runs as evidence.
- 05
Human oversight: escalation paths and checkpoints where the stakes demand them.
S.04 · Model Independence
A routing decision, not a re-platform.
Nugawi is model-independent by design and an approved partner of both Anthropic and OpenAI, so adding a European model is a routing decision, not a re-platform. The pipeline is wired so the model behind a workflow can change without rebuilding it, and the evaluation harness proves the chosen model on your actual workflow before anything ships. Mistral's models are one of the options we route to, including its open models for the self-hosted tier. No option is a default, and no vendor decides your architecture.
01Any tier of the ladder is available inside the standard build, and the model routing is a configuration decision, not an architectural one.
02For systems already live, the audit maps where the data actually flows today and names what would have to move, and at what effort, to reach the control tier you need.
03Golden cases drawn from your real decisions grade every candidate model on your actual workflow, so the European option is chosen on evidence, not on claims.
S.05 · Questions Buyers Ask
The questions before the pipeline.
What is the difference between data residency and sovereignty?
Data residency says where your data is stored and processed, and an EU-resident region on a managed platform answers that question. Sovereignty asks a harder one: who can access the data, who operates the service, and which legal powers can reach it, including the provider's own jurisdiction. A US-headquartered provider can offer real EU residency; only infrastructure you control, running open-weight models you hold, answers the sovereignty question in full.
Can you run models inside our own environment?
Yes. Open-weight models, such as Mistral's open models, run on European infrastructure you control: your cloud region in the EU, your own data centre, or an environment fully isolated from the network. No data leaves your boundary and the model weights stay with you, so the deployment survives any decision by any provider.
Does EU-pinned inference satisfy our regulators?
It depends on what your regulator requires, and we help you hold that conversation with evidence rather than reassurance. EU-pinned inference keeps the inference step in-region, which satisfies many GDPR data-residency requirements, but limited, safeguarded sub-processor transfers remain possible under the provider's terms. Where the brief demands verifiable EU-only processing, we move that workflow to the self-hosted tier.
Who owns the model in a self-hosted deployment?
You do. The weights of an open-weight model are files you hold: they are downloaded, versioned and run inside your infrastructure, on your hardware or in your cloud tenancy, and they remain yours to operate, audit and upgrade. No vendor can withdraw access to a model that runs inside your boundary, and that is what makes this tier verifiably EU-only.
Do we have to give up Claude or GPT?
No. Nugawi is model-independent by design and an approved partner of both Anthropic and OpenAI, so the right model is chosen per workflow, not per vendor. Adding a European option such as Mistral's models is a routing decision, not a re-platform: Claude keeps the workflows where it is the right model, and a European model serves the workflows where jurisdiction demands one, inside a single governed pipeline.
What does this change cost?
Structure only, and all figures are quoted on application. The three tiers carry different operating profiles: managed EU hosting runs as a platform subscription, EU-pinned inference is billed by usage under the provider's service tiers, and a self-hosted deployment is an infrastructure and engineering investment that removes per-usage model spend. We scope the ladder against your workflows and quote before anything ships. Engagement structure: https://nugawi.ai/pricing.
S.06 · Start
European control, proven before it ships.
Bring us the brief, the procurement requirement or the regulator letter. We map your workflows onto the control ladder, put in writing what each tier covers and does not cover, prove the chosen model on your data with the evaluation harness, and ship the governance with the build. Figures are quoted on application, and the tier you choose stays a per-workflow decision you can change without rebuilding the pipeline.