Elia Studio

← All articles

Sovereign AI: avoid premature infrastructure spending

Alexandre Gon · Head of AI & CTO6 estimated min read

Size sovereign AI for a mid-sized organisation: requirements, usage-based inference, dedicated capacity and review conditions.

A mid-sized organisation should connect permanent AI capacity to a real need before buying it. Elia Studio proposes defining data-control requirements first, then comparing eligible inference options. The aim is to avoid overinvestment without discarding sovereignty requirements—not to promise that an arbitrary model API meets them. The right starting architecture should be proportionate and its boundaries explicit.

Sovereign AI: proportionate investment, from requirements to capacity sizing.

Sovereignty is not a server size

Infrastructure decisions begin with the information the system will use and the assurances it must provide. Selecting dedicated hardware first can turn an intention to retain control into an unjustified technical commitment.

Definition: inference means running a model to produce a response. Billing, isolation and processing location are separate properties that need verification.

Usage-based billing does not establish sovereignty. Dedicated capacity does not establish it on its own either. Examine the specific offer, data flows, access and responsibilities.

CNIL, the French data protection authority, explains that deployment choices depend on use cases and data. For external infrastructure, it highlights contracts, authorised access and processing conditions. CNIL FAQ, published 18 July 2024, reviewed 4 November 2026, in French.

Requirements first, price second

Our method separates two decisions. The first determines which options are eligible. The second compares how those options operate and what they cost for the intended work.

An offer failing a mandatory constraint does not become acceptable because it is cheaper. If no usage-based offer meets the retained requirements, examine dedicated capacity, reduce the scope or defer the affected use.

Think of a workshop: reserving a machine can make sense, but its performance alone does not justify buying it. For AI, the analogy stops at data and security requirements. Those cannot be traded away through a financial comparison disguised as a technical decision.

Establish the control boundary

Before discussing a graphics processing unit, or GPU, we propose documenting inputs, processing, copies and administrative access. Include logs and supporting services rather than limit the analysis to the model.

QuestionEvidence or decision
What data enters the system?Categories, permitted uses and exclusions
Where does it go?Processing and destination map
Who can access it?Responsibilities and support access
What does the service retain?Retention and reuse conditions
Which assurances apply?Contract and exact qualification scope
How can we exit?Exports, dependencies and reconstruction

This is an Elia Studio working method, not certification. It complements our sovereign architecture and control guide without replacing project-specific technical and legal validation.

Do not borrow a provider’s qualification

ANSSI states that SecNumCloud recognises a specific cloud offering. Hosting an application on a qualified offering does not automatically qualify the application. Clients retain responsibility for assessing and managing their risks. ANSSI FAQ, reviewed 4 November 2026, in French.

That rules out shortcuts such as assuming a French provider makes an entire solution sovereign. Procurement needs the reference and scope of the actual offer, not a general logo.

Elia Studio claims no organisational certification in this article. Its role is to translate requirements into choices and request appropriate evidence from prospective providers.

Compare capacity models without assuming a winner

Once requirements are defined, economic comparison can examine different supply models. These categories support scoping; they do not guarantee that each option is available within the required assurance boundary.

Model to examineEconomic questionEssential check
Usage-billed inferenceVariable spending and contract conditionsProcessing, isolation, retention and availability
Capacity reserved for a periodCommitment and expected utilisationExit, access and change conditions
Permanent dedicated capacityCommitment and operational workloadActual need, maintenance and continuity

Usage billing does not automatically mean there is no minimum commitment. Procurement must examine the actual terms. Capacity should not be compared using a provider’s headline rate while ignoring obligations attached to it.

An illustrative decision, not a client estimate

Consider an assistant used irregularly to prepare documents, with human review before finalisation. Potential account count does not establish simultaneous demand.

The team can compare eligible options on the same authorised documents, required output quality, acceptable delays and several volume assumptions. It then examines peak demand and operating conditions.

An eligible usage-based service may be a starting candidate. If it fails the requirements, it is excluded for the affected data. No price or break-even threshold is invented in this example.

Model cost is not complete solution cost

Integration, data preparation, evaluation, operations and exit remain relevant. This article concentrates on when to commit capacity rather than providing another general cost catalogue.

Include human correction in the comparison. Changing infrastructure while degrading the final document can move cost rather than reduce it. Compare outcomes the business accepts, not compute prices alone.

A first measurement plan can record requests, accepted quality, correction effort and attributable costs over the same period. These are proposed measures, not operating results claimed here. Our automation ROI guide provides a complementary perspective.

Prepare the next decision

A proportionate starting point need not remain unchanged indefinitely. Define reassessment triggers: new data, steadier demand, latency requirements, provider changes or stronger isolation needs.

Reversibility requires understanding what can be exported and rebuilt. A contractual clause and a tested migration represent different levels of evidence and should remain distinct in the decision record.

Our nemrut-core customisation case shows how a business workflow can be implemented alongside cloud planning. It should not be interpreted as evidence that a target cloud deployment is complete.

Sovereign requirements and a proportionate budget

Elia Studio connects agent development, architecture and provider selection. The aim is to make the use case tangible and the investment explainable: what is purchased now, what requirements it meets and what would justify an evolution.

The objective is not the cheapest cloud at any price. It is a proportionate path that retains the agreed requirements and supports an informed decision.

Explore our AI roadmap and cost-control services or contact the studio to examine a use case, its data and operating conditions. This article is neither legal advice nor a compliance attestation.

Frequently asked questions

Does sovereign AI require a dedicated GPU?
There is no universal rule. Establish data, access, processing and operating requirements, then examine eligible offers. Dedicated capacity may be necessary in some contexts. Its dedicated nature alone does not establish sovereignty or compliance, and an economic comparison cannot compensate for an offer failing a mandatory requirement for the intended data.
Is usage-based inference automatically sovereign?
No. Billing does not establish processing location, control of access or data handling. The specific offer must meet the project’s requirements. If it fails a blocking criterion, a lower price does not make it acceptable for the data involved. Evaluate contractual conditions, configuration and supporting services before selecting a starting option.
Can savings be claimed before deployment?
A comparison or recommendation can be described as such, but not as realised savings. A financial outcome requires explicit assumptions and measurements. This article supplies no client prices, percentage saved or universal break-even point for dedicated infrastructure. It distinguishes a proposed decision method from a delivered and measured operating result.
Does an application inherit SecNumCloud qualification?
ANSSI states that hosting an application on a qualified offer does not automatically qualify that application. Examine the actual service scope and project responsibilities. Elia Studio claims no qualification of its own here, and the framework is not a compliance attestation. Qualified infrastructure does not remove the client’s risk-management responsibilities.

↑ Back to top