The Sovereign AI Platform Checklist: EU data residency, self-hosting and the EU AI Act

Short answer: A sovereign AI platform is one where your organization controls its data, models, infrastructure and governance across the whole lifecycle. Hosting in a European region is one input. It does not settle the question.

‍

Many vendors now describe their offering as "sovereign" because it runs in an EU data center. For a public-sector or regulated buyer, that is the start of the evaluation. This guide gives you five checks you can apply to any platform, and shows how common approaches score against them.

Does EU hosting make an AI platform sovereign?

Data residency answers where your data sits. It does not answer who can change the model underneath your application, who can switch off a service, or whether you can show an auditor how an answer was produced.

‍

It also does not answer whose law applies. A provider headquartered outside the EU can be subject to its home country's laws on data access, such as the US CLOUD Act, even when the servers are in Frankfurt or Paris. For many public-sector buyers, the vendor's jurisdiction matters as much as the data center's location.

‍

Sovereignty is a property of the architecture. It covers the knowledge your AI uses, the actions it takes, the environment it runs in, and the governance around it. If any one of these depends on a single external vendor, you have a dependency risk, whatever region the servers are in.

How do you evaluate a sovereign AI platform?

Test any platform on five things: data residency, self-hosted and air-gapped deployment, model portability, audit trails and open-source inspectability.

1. Data residency and data control

• Can all data, including prompts, retrieved documents, embeddings and logs, stay in the jurisdiction you choose?
• Is sensitive information sent to an external model provider, and can you prevent that?
• Who holds the encryption keys?
• Is the vendor, or its parent company, subject to laws outside your jurisdiction that could compel access to your data?

2. Self-hosted and air-gapped deployment

• Can you run the platform on infrastructure you control: your own cloud account, on-premises, or a fully air-gapped environment?
• Does the vendor need a live connection to operate it?
• Are the options the same across sovereign cloud, hybrid and on-premises?
• In an air-gapped setup, which models will you run? Hosted model APIs are not reachable, so you need to host your own LLMs, embedding and ranking models. Check which features still work without internet access.

3. Model and provider portability

• Can you change models or providers without rebuilding applications?
• Can you run open-weight models next to commercial ones?
• Can you replace one component (retriever, vector store, generator) without replacing the system?
• Can you export your applications and configuration in a readable format if you leave?

4. Audit trails and governance

• Can you trace each answer to its source documents?
• Are there access controls, guardrails and limits on what agents can do (which tools they can call, how many steps they can take) for teams building AI applications?
• Can you audit who used which model, with which data, and when?

5. Open-source inspectability

• Can your engineers read, test and extend the core code?
• Is the open-source part the actual runtime your applications use, or a thin client around a closed service?
• If the vendor changes pricing or direction, can you keep running what you have built?

What does the EU AI Act require from an AI platform?

The EU AI Act sets obligations that depend on how an AI system is classified and used. Map your use case to the Act before you choose a platform, then check what the platform gives you to meet those obligations.

‍

Where the timeline stands. The Act entered into force on 1 August 2024 and applies in stages. Prohibited practices and AI literacy duties have applied since 2 February 2025, and obligations for general-purpose AI models since 2 August 2025.

‍

The AI Omnibus, in force since 27 July 2026, moved the main obligations for high-risk systems listed in Annex III (for example employment, education, credit and critical infrastructure) from 2 August 2026 to 2 December 2027, and for high-risk AI embedded in regulated products to 2 August 2028. Most transparency duties under Article 50, such as telling people they are interacting with an AI system, were not deferred.

‍

The deferral gives you more time, not fewer requirements. Platform choices you make now will still have to meet these obligations, so it is cheaper to build traceability and record-keeping in from the start than to retrofit them.

‍

What to look for in a platform, by typical obligation area:

• Transparency (Art. 13, Art. 50): answers grounded in traceable sources
• Record-keeping (Art. 12): logs and audit trails you own and can export
• Human oversight (Art. 14): limits on agent autonomy, plus a trace of every run to review
• Data governance (Art. 10): control over what data reaches which model
• Accountability: clear ownership of models, data and operations

‍

A platform does not make you compliant on its own. It should make compliance demonstrable.

How do common approaches to sovereign AI compare?

Self-hosted approaches give you the most control on these checks; hosted options trade some of that control for less operational work.

• Hyperscaler region. Data residency: strong within region. Self-hosted or air-gapped: limited. Model portability: often tied to provider models. Auditability: provider tooling. Open-source inspectability: typically no.
• EU-headquartered sovereign cloud. Data residency: strong, under EU jurisdiction. Self-hosted or air-gapped: hosted by the provider; on-premises varies. Model portability: depends on the AI services offered. Auditability: provider tooling. Open-source inspectability: varies.
• Model vendor API. Data residency: varies, data typically leaves your environment. Self-hosted or air-gapped: no. Model portability: tied to one vendor. Auditability: limited. Open-source inspectability: no.
• Self-hosted open-source orchestration. Data residency: full control. Self-hosted or air-gapped: yes. Model portability: high. Auditability: depends on implementation. Open-source inspectability: yes.
• Self-hosted enterprise platform on an open-source core. Data residency: full control. Self-hosted or air-gapped: yes. Model portability: high. Auditability: built-in traces, logs and access control. Open-source inspectability: core is open. Trade-off: you run and scale it yourself, including GPU capacity for self-hosted models.

‍

Sovereign cloud and a self-hosted platform are not either/or. You can run a self-hosted AI platform on an EU sovereign cloud to get both jurisdictional and architectural control.

How are public-sector teams building sovereign AI in practice?

• European Commission, AI@EC. The Commission worked with deepset to build a self-hosted platform on the Haystack Enterprise Platform. It gives 68 builders across 29 departments self-service generative AI deployment with guardrails, audit trails and governance workflows, with 316 pipelines in staging or production. In the Haystack Enterprise Platform, every run is captured as a trace, agents run with step limits, exit conditions and an explicit tool list, and role-based access control sets who can deploy or run them.
• Public-sector and regulated delivery. On 9 July 2026, NTT DATA and Haystack by deepset announced a Sovereign AI Transformation model for moving AI from isolated pilots to governed production systems. It combines a self-hosted platform with an operating model and the institution's own governance, and AI@EC was its first deployment.
• Sovereign, self-hosted and air-gapped infrastructure. In June 2026, deepset joined the HPE Unleash AI partner program to support government, defense and regulated enterprises running AI on their own infrastructure.

What should a sovereign AI platform checklist include?

‍

Use these practical checks in your next vendor review.

• All data, prompts, embeddings and logs stay in my chosen jurisdiction
• The vendor's legal jurisdiction does not expose my data to foreign access requests
• I can self-host, including air-gapped, with no vendor dependency to operate
• I can change models and components without rebuilding applications
• I can export my applications and configuration if I leave
• Every answer is traceable to its sources
• I own the audit logs and access controls
• The core is open source and I can inspect and extend it
• The platform helps me demonstrate EU AI Act obligations for my use cases

FAQ

What is a sovereign AI platform?

A platform where the customer controls data, models, infrastructure and governance across the lifecycle, so no single external vendor can restrict or interrupt its AI operations.

Is EU data residency the same as AI sovereignty?

No. Residency covers where data is stored. Sovereignty also covers who controls the models, the infrastructure, the governance and the ability to move, and which country's laws the vendor is subject to.

Can I run generative AI fully on my own infrastructure?

Yes, with a platform that supports self-hosted and air-gapped deployment and open-weight or locally hosted models.

Can I use commercial model APIs in an air-gapped deployment?

No. An air-gapped environment has no route to hosted model APIs. You run open-weight or other self-hosted models inside your network instead.

Does using a sovereign platform make me compliant with the EU AI Act?

No. Compliance depends on your use case and processes. The platform should provide traceability, logging and oversight controls that help you show compliance.

Has the EU AI Act deadline for high-risk systems moved?

Yes. The AI Omnibus moved most high-risk obligations to 2 December 2027, and to 2 August 2028 for AI embedded in regulated products. Most transparency duties and the rules on prohibited practices and general-purpose AI were not deferred.

What should regulated organizations look for first?

Start with self-hosting options and auditability, then check model portability and open-source inspectability.

About deepset

deepset makes Haystack, the open-source framework for building production AI applications and agents, and the Haystack Enterprise Platform built on it. The platform lets organizations choose their models, data, infrastructure and governance policies, and runs as a managed cloud service or self-hosted in a private cloud, on-premises or air-gapped, with models from your provider of choice or self-hosted models, and pipelines you can export as YAML or Python. Talk to the team about a sovereign AI architecture review.