Public sector AI cloud

European Government AI Cloud: how public bodies decide what runs where

European governments are sorting their AI workloads by how much foreign control each one can tolerate, and writing the answer into procurement rules. The European government AI cloud that results is a split estate, with everyday services on any compliant cloud and the systems that touch citizens’ rights or state security on infrastructure the state can control.

Why does AI push public bodies towards sovereign infrastructure?

Much of public-sector IT can run in a foreign-operated cloud region with manageable risk. AI is different because of what governments use it for. Annex III of the EU AI Act, the Act’s list of high-risk uses, reads in large part like a list of state functions, covering eligibility for public assistance and healthcare services, law enforcement, migration and border control, and support to judicial authorities.

A public body using AI in those areas is the deployer of a high-risk system, which brings duties to keep the system’s logs and assign human oversight. Under Article 27, bodies governed by public law also have to carry out a fundamental rights impact assessment, a review of how the system could affect the people it is used on, before they first use it.

Those duties assume the deployer can reach into the running system. A ministry that cannot vouch for its own audit logs, because the platform is administered under another jurisdiction, has a compliance problem to solve, as our post on why the EU AI Act demands localised compute infrastructure sets out.

How will procurement rules decide where government AI runs?

The Commission’s proposed Cloud and AI Development Act (CADA) would give public bodies a single sovereignty framework with four assurance levels, which they apply based on their own risk assessments. The levels run from EU-located data processing at level 1 up to full control of the software supply chain at level 4, and our post on how EU privacy laws turn into an advantage walks through each one.

What matters for government AI is how the levels map onto public work. Wilson Sonsini’s analysis of the proposal points out several details:

  • AI systems are in scope: The proposal covers remotely hosted and operated AI systems as well as conventional cloud services.
  • Routine work can stay at level 1: Activities such as event management, tourism campaigns and general administration can use the lowest level.
  • Sensitive work needs level 2 or above: Bodies contributing to public order in sectors covered by NIS2, the EU’s cybersecurity directive, or in national security, defence, justice or law enforcement may only use providers at levels 2, 3 or 4, depending on how sensitive the activity is.
  • Non-EU providers are not shut out: Public bodies may use non-EU providers that meet at least level 3 and are controlled from a jurisdiction the Commission has approved.

CADA is a proposal, and its text is likely to change before it becomes law. The Commission’s own summary stresses keeping “the vast majority of the market” open to partners, so the stricter levels are aimed at the sensitive end of public work.

What does sovereign AI adoption look like in government?

Public buyers are putting these principles into contracts, and their purchases fall into a few distinct patterns:

  • The state hosts the model itself: France’s Ministry of the Armed Forces awarded Mistral AI a framework agreement, overseen by its defence AI agency AMIAD and open to the armed forces and affiliated public bodies, with Mistral’s solutions deployed on France’s own infrastructure. That keeps the models on hardware the state controls, and Mistral also plans to fine-tune them on defence data.
  • Governments build a platform with industry: France and Germany announced plans for a partnership with SAP and Mistral to deliver sovereign AI for public administration, including decision-support agents for civil servants and citizen-facing agents for eligibility checks. Once those checks concern public benefits, they can fall within Annex III.
  • Institutions buy against a sovereignty score: The Commission awarded its sovereign cloud tender, worth up to €180 million, to four European-led providers scored against its own Cloud Sovereignty Framework and its Sovereignty Effectiveness Assurance Levels (SEAL). Three reached SEAL-3, while a Proximus-led consortium using Google Cloud technology through its partner S3NS, operated by EU companies, reached the minimum SEAL-2. The Commission noted that “even non-European technologies, when operated within a strict and appropriate framework, can meet the minimum level of sovereignty required”.

What should a public buyer ask an AI cloud provider?

For a public body choosing where its AI runs, the useful questions concern who can act on the infrastructure:

  • Log access: The deployer should be able to retain and inspect the logs the AI Act requires without a foreign operator in the path.
  • Administration and legal reach: The buyer needs to know who administers the platform and under which law they can be compelled to act.
  • Data and model weights: Fine-tuning data and the weights that result from it should stay inside the agreed perimeter.
  • Exit: The buyer’s data and any fine-tuned weights should be able to move to another provider when the contract ends.

Each answer turns on who controls the platform and which laws apply to them, which is why jurisdiction decides whether a cloud counts as sovereign more than the location of the servers.

From government tenders to regulated industry

The criteria public buyers are writing are set to reach beyond government tenders. Under the CADA proposal, private entities in high-criticality NIS2 sectors such as energy, transport, banking and healthcare would be allowed to apply similar selection criteria, so the standard set by public procurement is likely to become the one their boards and regulators measure them against. The wider case for why Europe needs compute it controls is set out in our guide to data sovereignty in AI.

Neurotechnology Cloud runs dedicated and isolated private AI cloud environments from its data centre in Vilnius, Lithuania, for organisations whose most sensitive AI workloads need to stay under European control.

If you are deciding which AI workloads belong inside a sovereign perimeter, get in touch.

Share: 

Contact us

Interested in our products, custom solutions, or partnership opportunities? Have questions about our technologies or need more information before purchasing? Fill out the form, and our team will get back to you as soon as possible.