Reference framework

Sovereign AI Levels: a shared way to say how sovereign you actually are.

"Sovereign AI" has become a slogan that means five different things in the same meeting. This is our attempt to fix that — five levels and six dimensions that turn "how sovereign are we?" into a question you can actually answer, defend to a regulator, and price. Grounded in the EU Cloud Sovereignty Framework and the plain reality of the CLOUD Act.

The starting point

Sovereignty is a posture, not a product.

You don't "buy sovereignty." You choose, workload by workload, where you need control, where dependence is acceptable, and where a partner is worth more than ownership. Three principles run through everything below.

It's multi-dimensional

Data, model, compute, operations, jurisdiction and assurance can each sit at a different level. Your real level is the floor across the dimensions that matter for that workload — not the best one you can point to.

Higher isn't better

Every step up costs speed, money, and often model quality. The right level is the one your regulator and your board can defend — and no higher. Over-engineering sovereignty is a real, common mistake.

It's a business decision

Sovereignty belongs on the business side — a question of risk, regulation, and trust — not an IT checkbox. The people who own the P&L and the compliance exposure should own the level.

The framework

Five levels, from global default to fully autonomous.

Each level raises the guarantee — and the cost. Read them as a ladder you climb only as far as a given workload requires. The meter on each shows relative sovereignty; the coloured rail marks the line where foreign law stops reaching your data.

L0

Global Default

Public global cloud and foreign proprietary AI APIs. No sovereignty posture.

Data
May leave the EU; residency not guaranteed.
Model
Foreign proprietary API; no control over weights or hosting.
Compute / infra
Global regions, foreign-owned and operated.
Operations
Administered from anywhere, including third countries.
Jurisdiction
Full exposure to foreign law (e.g. US CLOUD Act).

Fit: public or non-sensitive data, experimentation, speed-first use where sovereignty simply isn't a requirement.

L1

Regional Residency

Data stays in the EU, contractually — but on a foreign-owned platform.

Data
EU data boundary; residency guaranteed by contract. Customer-managed keys optional.
Model
Foreign or EU-hosted API within the region.
Compute / infra
EU regions of a global hyperscaler.
Operations
Mostly EU, but remote support paths may exist.
Jurisdiction
Still exposed — a foreign-owned provider can be compelled under foreign law even for EU-resident data.

Fit: moderate-sensitivity workloads where residency and GDPR posture are enough, and full jurisdictional independence isn't required.

The CLOUD Act line. Below it, your data is in the EU. From L2 up, it moves progressively beyond the reach of foreign law. This is the single sharpest decision in the ladder.
L2

Operational Sovereignty

An EU-operated sovereign environment with a separate control plane.

Data
EU-resident, customer-held keys, transparency over who can access it.
Model
Models run inside the sovereign environment; inference stays in-boundary.
Compute / infra
Dedicated EU sovereign cloud — e.g. an EU-operated hyperscaler offering or partner-operated region.
Operations
Operated and administered by vetted EU personnel; no third-country remote admin.
Jurisdiction
Reduced, but bounded by the provider's ultimate ownership.

Fit: regulated workloads (finance, health, critical infrastructure) that need operational control and auditable access without going fully self-hosted.

L3

Jurisdictional Sovereignty

EU-owned, outside foreign law — with models and compute you control.

Data
Under EU jurisdiction end to end; no foreign compulsion path.
Model
Open or EU-hosted weights you can run, inspect, and move; no lock-in to a foreign API.
Compute / infra
Sovereign compute from an EU-owned provider or your own stack.
Operations
Full operational and technical independence; documented exit rights.
Jurisdiction
Immune to extraterritorial law — the provider entity is EU-owned and controlled.

Fit: highly regulated, critical-infrastructure, or strategically sensitive workloads where foreign legal reach is unacceptable.

L4

Autonomous Sovereignty

Self-hosted, isolated, no runtime dependency on anyone.

Data
Never leaves your controlled environment.
Model
Weights held and run entirely in-house; full custody.
Compute / infra
On-premise or air-gapped; owned outright.
Operations
Operated solely by your own cleared personnel.
Jurisdiction
No external dependency to compel — sovereignty by isolation.

Fit: national-security-grade, defence-adjacent, or classified workloads. Highest cost and friction, with a real trade-off against access to frontier model quality.

The other axis

Six dimensions — because sovereignty is never a single number.

A level is a useful shorthand, but the honest assessment runs dimension by dimension. A system can be L3 on data and L1 on operations at the same time. These are the six axes to score.

01

Data

Where it lives and, more importantly, who has the legal authority to compel access to it.

02

Model & weights

A foreign proprietary API you can't see inside, versus weights you can host, inspect, and move.

03

Compute & infrastructure

Where the GPUs and the control plane physically sit — and who ultimately owns and operates them.

04

Operations

Who administers the system day to day, and whether anyone in a third country can reach into it.

05

Jurisdiction & law

Exposure to extraterritorial law such as the CLOUD Act, driven by the provider entity's ownership.

06

Assurance & certification

Provable, not asserted: EUCS level, CSF assurance level, and independent audit.

Applying it

Which level does your workload actually need?

Start from the regulation and the risk your board carries — not from the technology. As a starting point, this is where typical enterprise segments land. The floor is a minimum, not a ceiling.

Segment
Typical floor
Main driver
General enterprise (retail, industrials)
L1
Data residency and optics; cost-sensitive, low regulatory pressure.
Healthcare & life sciences
L2–L3
GDPR special-category and patient data; strong access control.
Financial services
L2–L3
DORA, outsourcing rules, exit and concentration risk.
Telecom & critical infrastructure
L2–L4
NIS2, national-security relevance; some now building their own.
Public sector & defence-adjacent
L3–L4
Classification, procurement rules, foreign-reach intolerance.
The goal isn't the highest level. It's the defensible one — matched to the risk your regulator and your board actually carry, and no higher.
Where WenableAI fits

We help you set the level — and stand up the capacity to reach it.

A framework is only useful if it leads to a decision and a working result. That's the part we own.

From "how sovereign?" to a running capability.

EnableBrokerSet up

We work with the business side to decide the right level per workload, own the requirements and architecture, and govern the build — while vetted sovereign platform partners deliver the physical layer. We don't build the data centre; we make sure the AI opportunity lands on infrastructure that holds the level your board can defend, and we execute until your organization can take it over.

How this maps to the standards

This is WenableAI's synthesis, built to be interoperable with the emerging EU reference points — the EU Cloud Sovereignty Framework (CSF) and its assurance levels, and the EUCS certification scheme (Basic / Substantial / High). Our L1–L3 broadly track the CSF's rising assurance; L4 sits above it as full self-hosting. It is a reference for aligning a conversation, not legal advice — sovereignty obligations should always be confirmed with qualified legal and compliance counsel for your specific case.