Guides › What is data sovereignty

What Is Data Sovereignty? A Definition for the AI Era

Data sovereignty is the principle that data is subject to the laws of the jurisdiction in which it is collected, stored or processed, and, in its operational sense, that the organization holding the data decides who can access it, where it is processed, how long copies persist and who may learn from it. It differs from data residency, which is only where data is stored, and from data localization, which is a legal requirement to keep certain data in-country. AI systems reopened the question because a model query moves data at the moment of inference, in ways a storage region was never designed to govern. This page gives the definition, the laws behind it, the distinctions that get confused, and what sovereignty requires once an AI system is in the loop.

The definition, and the two senses it carries

In law, data sovereignty means that data is subject to the legal authority of the jurisdiction where it sits or is processed, and that a foreign government may also be able to reach it through the entity that controls it. In operations, it means control: the holder of the data decides who can read a record, which systems process it, how long copies exist, and whether anyone outside the organization is permitted to learn from it. Most disputes about whether a service is “sovereign” come from one side using the first sense and the other the second. A cloud region inside the EU satisfies a residency requirement and may satisfy part of the jurisdictional test, while leaving the operational questions, and the question of foreign legal reach, unanswered.

The concept predates AI by decades. It became a procurement question in the 2010s as cloud services concentrated in a handful of providers headquartered in one country, and a compliance question after the EU's GDPR made cross-border transfers conditional. It became an engineering question in 2023, when the most capable models started running outside every customer's boundary and every useful workflow began with showing the model some data. The data sovereignty for enterprise AI guide takes the engineering question from here; this page stays with the definition.

Sovereignty, residency and localization

Three terms, three questions. Residency asks where data is physically stored and is settled by a region or data center selection. Localization asks whether a law requires certain data to stay inside a country's borders and is settled by statute. Sovereignty asks whose laws govern the data, who can compel its production, and who controls its processing and retention, and is settled by the provider's corporate domicile, the contract and the architecture together. A service can be resident in Frankfurt, satisfy no localization rule because none applies, and still fail a sovereignty review because the operator is subject to a foreign disclosure order. The sovereignty vs residency page works through the comparison in detail.

The laws that make it concrete

No single statute defines data sovereignty. It is the combined effect of regimes that each answer one of the three questions, and the table lists the ones that come up in enterprise reviews. Two features of the list matter for AI. None of these regimes mentions prompts, retrieval context or model weights by name, and every one of them applies to them, because each is a copy of regulated data crossing a boundary.

What changes when an AI system processes the data

Residency controls were designed for data at rest. An AI system touches data in three motions that a storage region does not cover. Inference processes every prompt and every retrieved document wherever the model is served, and regional storage of a tenant says nothing about regional inference unless the provider commits to it separately. Retrieval pulls documents out of a store in one jurisdiction and hands them to a model as context, so the documents never changed residency while copies of them crossed in the query path. Training and fine-tuning encode data into weights that are served from wherever the model runs and cannot be deleted from the model afterward. A fourth motion sits with the provider: abuse monitoring, support tooling and telemetry hold prompts and outputs under the provider's retention terms, visible to its subprocessors.

This is why a sovereign AI deployment is defined by what crosses rather than by where the tenant lives. The working definition the rest of this library uses is that an AI deployment is sovereign when the organization decides, and can demonstrate, what sensitive data reaches any third party, in what form, for which user, and for how long it persists. In practice that comes down to five properties: data stays in the systems that already govern it, each retrieval is scoped to the person asking at query time, real identifiers are replaced before anything reaches a third-party model, every retrieval is recorded, and the deployment model, whether cloud, private cloud, on-premises or disconnected, follows the obligations attached to the data. The sovereign AI deployment page compares those deployment options, and the zero data retention page explains why a retention commitment is one of the five and not a substitute for them.

How the term is used in practice

Three usages are common and worth telling apart. Cloud providers use “sovereign cloud” for regions operated under local legal structures, sometimes by a local partner, to reduce exposure to foreign law; the proposed EU Cloud and AI Development Act would grade such services by tier, and it is not yet law. Regulators use “data sovereignty” when they mean transfer rules, as in GDPR Chapter V, where a transfer outside the EEA needs an adequacy decision, standard contractual clauses or another mechanism plus a transfer impact assessment. Security teams use it to mean operational control, which is the sense an architecture can deliver. When a vendor says its AI product is sovereign, ask which of the three it means, and then ask what crosses at inference time.

Where Hardshell sits

Hardshell is built for the operational sense of the term. It connects enterprise data to ChatGPT, Claude, Microsoft 365 Copilot, Gemini and internal assistants without migrating it, enforces each user's existing permissions at query time, substitutes stand-ins for real identifiers before content reaches a third-party model, and records every retrieval under opaque IDs. It tests what an AI system can leak before rollout and maps each finding to MITRE ATLAS technique IDs, and it deploys as a cloud service, in a private cloud, or as an offline bundle for environments where the data cannot leave at all.

Frequently asked questions

What is data sovereignty in simple terms?

Data is governed by the laws of the place where it is collected, stored or processed, and the organization holding it decides who can access it, where it is processed and how long it is kept. For AI, it means controlling what data reaches a model provider and being able to prove it.

What is the difference between data sovereignty and data residency?

Residency is where data is stored, settled by a region selection. Sovereignty is whose laws govern the data, who can compel access, and who controls processing and retention. A Frankfurt region settles residency; it does not settle sovereignty if the provider is subject to a foreign disclosure law.

Is data sovereignty a legal requirement?

Not as a single rule. It is produced by several regimes together: transfer rules such as GDPR Chapter V, localization laws such as China's PIPL and Russia's 242-FZ, access laws such as the US CLOUD Act, and protection regimes such as NIST SP 800-171 and CMMC for controlled unclassified information. Which ones bind you depends on the data and the jurisdictions involved.

Does using an EU region make an AI service sovereign?

It makes the storage resident. Inference may be served elsewhere unless the provider commits to in-region processing, retrieval moves copies of documents in the query path, and a US-headquartered provider remains reachable under the CLOUD Act wherever the data sits. Check the inference commitment, the subprocessor list and the contract, not only the region.

What does data sovereignty mean for generative AI specifically?

Every prompt, retrieved document and fine-tuning upload is a transfer to the model provider, and provider retention holds a copy afterward. Sovereignty for generative AI means deciding what crosses, in what form, for which user and for how long, and being able to show it.

Can an organization be data sovereign while using ChatGPT, Claude or Copilot?

Yes, if the data that reaches the provider is scoped to the asking user's entitlements, identifiers are substituted before the boundary, retrievals are recorded, and the contract covers training, retention and residency for the tier in use. The enterprise tiers provide the contractual half; the architectural half has to be built or bought.