AdoptionAbandonment

Is AI available in Europe: GDPR, residency and adoption

Learn how GDPR, the US CLOUD Act and the EU AI Act impact European AI adoption and discover the requirements for legally compliant sovereign infrastructure.

, AI Researcher10 min read
In this article8
  1. Sizing European enterprise AI adoption in 2025
  2. How data protection rules delay advanced model rollouts
  3. The AI Act timeline: Staggered entry into force
  4. GDPR and AI models: The EDPB opinion
  5. The Cloud and AI Development Act assurance levels
  6. Establishing continuous traceability for AI workloads
  7. Frequently asked questions
  8. Sources
The flag of the European Union flying against a clear sky
Availability in Europe is rarely a technical question; it is a question of which law the data sits under. Photo: Håkan Dahlström, CC BY 2.0, via Wikimedia Commons

The sections that follow quantify European adoption, then work through the instruments that shape it: the staggered AI Act calendar, the GDPR as interpreted by the EDPB, and the proposed Cloud and AI Development Act sovereignty tiers.

Alongside those rules, access to frontier models in the EU has lagged the United States, a gap driven mainly by data protection compliance rather than by the AI Act itself, which the following sections quantify.

Sizing European enterprise AI adoption in 2025

In 2025, 20.0% of enterprises with 10 or more employees across the European Union used artificial intelligence technologies to conduct their business, according to data released by Eurostat on 11 December 2025[1]. This represented an increase of 6.5 percentage points compared with the 13.5% recorded in 2024, continuing an upward trajectory from 7.7% in 2021 and 8.1% in 2023[1]. Uptake remains heavily concentrated by sector: 62.52% of enterprises in information and communication and 40.43% in professional, scientific and technical service activities used AI, while in all other economic activities the share stayed below 25%[2]. A fuller breakdown by size class, member state and use case is set out in the enterprise adoption survey data.

This expansion has occurred alongside substantial institutional friction. Among EU enterprises that had ever considered using AI technologies but were not using them in 2025, 52.52% pointed to a lack of clarity about the legal consequences and 48.83% to concerns about violations of data protection and privacy, behind only a lack of relevant expertise (70.89%)[2]. Regulatory compliance requirements under European frameworks form a primary governance hurdle for IT procurement teams.

How data protection rules delay advanced model rollouts

While foundational AI usage is growing, European access to frontier foundation models has faced systematic delays compared to the United States. A comprehensive study published by the Centre for the Governance of AI (GovAI) in June 2026 examined 375 frontier large language model (LLM) releases across eight years (June 2018 to May 2026) by Meta, Google, OpenAI, and Anthropic[3]. The findings reveal that 11% of advanced LLM releases were delayed or withheld from the EU market relative to the US, compared to 7% in the United Kingdom[3].

The divergence between EU and US deployment timelines stems from the operational demands of the General Data Protection Regulation (GDPR). Providers seeking to deploy in Europe must substantiate a lawful basis for training and inference, manage data subject access requests against pre-trained weights, and navigate cross-border data transfer regimes.

The study tentatively attributed regulatory factors as the primary driver behind 56 of the 68 recorded delay or non-release instances [3] . Regulatory barriers were driven predominantly by data protection law rather than by the AI Act, and by uncertainties surrounding the legality of training frontier models on personal data obtained without explicit individual consent [4] . Where the blocker is a stalled deployment rather than a prohibited use case, an inference contract that can be checked narrows the exposure that causes the delay. Lyceum, for example, states the hosting region per model rather than for its catalogue as a whole, retains no prompts, offers a Data Processing Agreement on request and bills per token rather than on committed spend; its GDPR position is self-asserted rather than certified, which is precisely the kind of distinction a data-protection review needs to see in writing.

  • Meta exhibited the highest rate of regional withholding, with approximately 26% of its model releases delayed or blocked in the EU and 15% in the UK[3].
  • The European market experienced a 71-day delay for the consumer web application of Anthropic's Claude 3 Opus while it held top ranking on standard capability benchmarks, even though API access remained simultaneous[3].
  • Regulatory enforcement actions by national data protection authorities (DPAs), such as the Irish Data Protection Commission and Italy's Garante, directly prompted frontier developers to pause regional rollouts pending compliance reviews.

The AI Act timeline: Staggered entry into force

Regulation (EU) 2024/1689, known as the EU Artificial Intelligence Act, was published in the Official Journal of the European Union on 12 July 2024 and entered into force on 1 August 2024. Rather than imposing a single effective date for all obligations, Article 113 establishes a staggered enforcement timetable, since extended by later amendment[5].

  • 2 February 2025: Application of Chapter I (General Provisions) and Chapter II (Prohibited AI Practices), banning unacceptable-risk systems such as cognitive behavioural manipulation and untargeted facial recognition scraping[5].
  • 2 August 2025: Application of Chapter III Section 4 (Notifying Authorities), Chapter V (General-Purpose AI Models), Chapter VII (Governance), Chapter XII (Penalties), and Article 78[5].
  • 2 August 2026: General application date across the Union, activating general transparency obligations for deployers and providers. Regulation (EU) 2026/1744, the Digital Omnibus on AI, subsequently deferred the obligations for standalone high-risk systems listed in Annex III to 2 December 2027[5].
  • 2 August 2028: Application of obligations under Article 6(1) for high-risk AI systems integrated into products already covered by Union harmonisation legislation listed in Annex I, extended from 2 August 2027 by the same regulation.

GDPR and AI models: The EDPB opinion

To establish regulatory consistency across the single market, the European Data Protection Board (EDPB) adopted Opinion 28/2024 on 17-18 December 2024 under Article 64(2) of the GDPR[6]. The opinion clarifies how core data protection obligations apply to the development and operational deployment of AI architectures.

The EDPB concluded that AI models trained on personal data cannot be presumed anonymous by default. For a model to be deemed legally anonymous, a controller must demonstrate that it is very unlikely that individuals can be directly or indirectly identified, or that personal data can be extracted through targeted prompt injection or inference queries[6]. Where personal data remains memorized within weights, GDPR principles govern the system continuously.

  1. Legitimate Interest Assessment: Processing personal data for AI training under Article 6(1)(f) GDPR requires passing a cumulative three-step test: proving a legitimate purpose, demonstrating that the processing is strictly necessary without less intrusive alternatives, and ensuring the fundamental rights of data subjects do not override the controller's interest[6].
  2. Reasonable Expectations: Regulators evaluate whether individuals could reasonably expect their public or private data to be scraped or ingested, accounting for context, relationship, and initial collection terms[6].
  3. Downstream Model Contamination: When a foundation model is trained on unlawfully processed personal data, the EDPB stated that this unlawfulness directly affects the legality of downstream deployment, unless complete anonymisation of the resulting weights can be mathematically or empirically verified[6].

The Cloud and AI Development Act assurance levels

In response to strategic dependencies on non-EU cloud and compute providers, the European Commission structured its tech sovereignty initiatives under the AI Continent Action Plan, aiming to build domestic capacity and define unified sovereignty standards for sensitive workloads[7].

The proposed Cloud and AI Development Act (CADA), a key part of that plan, defines cloud and AI sovereignty as four assurance levels for public sector bodies to apply according to their own risk assessments, with providers recognised under the framework by Member States after an audit[7]. The framework provides public sector buyers and regulated private enterprises with standardised criteria to evaluate vendor independence from third-country jurisdiction.

Assurance LevelPrimary RequirementJurisdictional and Operational Scope
Level 1Union InfrastructureData is processed and stored in infrastructure located within the Union.
Level 2Supply Chain TransparencyDemonstrated independence from third countries and verified transparency over the software supply chain.
Level 3EU Ownership and ControlProvider must be owned and controlled from the EU, with strict personnel citizenship and operational criteria.
Level 4Full AutonomyComplete transparency and control over software supply chain, with guaranteed immunity from third-country interference.

Level 1 reflects conventional data residency, where US hyperscaler regions operating in Europe satisfy geographic storage constraints but remain vulnerable to parent-company subpoenas. Level 2 introduces software auditability. Levels 3 and 4 establish binding legal sovereignty: providers must maintain European ownership, exclude extraterritorial executive control, and ensure that operational personnel holding administrative access are EU citizens operating under Union law[7]. The limits of residency as a legal control, including the reach of the US CLOUD Act over EU-located infrastructure, are examined separately in data residency for European inference.

Under CADA, member state authorities audit and certify providers against these assurance tiers. Public administrations handling critical public-sector data and critical national infrastructure operators are guided to mandate Level 3 or Level 4 compliance in procurement tenders, reducing reliance on third-country infrastructure.

Establishing continuous traceability for AI workloads

To satisfy both the GDPR and the obligations established by Regulation (EU) 2024/1689 (the EU AI Act), organizations must transition from point-in-time documentation to continuous runtime governance. Under Article 12 of the AI Act, high-risk AI systems must incorporate automated logging capabilities that record operations throughout their lifecycle.

Runtime traceability requires capturing a deterministic audit trail for every inference request. This audit trail must systematically document the data trajectory from initial user input to final output generation, ensuring that all processing activities remain verifiable against European legal benchmarks.

  1. Inference-Level Provenance: Record the exact model version, system prompt, temperature, and configuration parameters applied to each transaction.
  2. Dynamic Tool and API Call Logging: Monitor and log all external tool calls, parameter payloads, and API response headers to maintain an accurate register of sub-processors.
  3. Retrieval Corpus Boundary Enforcement: Implement runtime policy checks on vector databases and knowledge bases to restrict retrieval strictly to GDPR-compliant, regionally bound data stores.
  4. Cryptographic Audit Trails: Store immutable execution logs and data flow telemetry to substantiate compliance during supervisory audits by national DPAs or the European AI Office.

By implementing continuous telemetry and verifiable data trajectory controls, European enterprises can systematically deploy advanced AI systems while maintaining demonstrable compliance with EU data protection and infrastructure sovereignty regulations.

Frequently asked questions

What percentage of European companies are using AI?

In 2025, 20.0% of EU enterprises with 10 or more employees used AI technologies, up from 13.5% in 2024. Adoption is uneven across member states and rises sharply with company size.

When does the EU AI Act fully enter into force?

The AI Act officially applies from 2 August 2026. However, its implementation is staggered. Prohibitions on unacceptable risk AI practices took effect earlier on 2 February 2025, while Regulation (EU) 2026/1744 deferred the obligations for standalone high-risk systems listed in Annex III to 2 December 2027, and to 2 August 2028 for high-risk AI embedded in products already regulated under Union harmonisation law.

Can legitimate interest be used as a GDPR legal basis for AI?

Yes, the European Data Protection Board (EDPB) confirmed in its December 2024 opinion that legitimate interest can serve as a legal basis for AI processing. Organizations must pass a strict three-step balancing test to prove the processing is strictly necessary and respects fundamental rights.

Why is data residency not enough for EU data sovereignty?

Physical data residency ensures data is stored on servers located within the EU, but it does not protect against foreign jurisdiction. The US CLOUD Act requires providers subject to US law to disclose data upon request, effectively overriding physical residency and compromising true legal sovereignty.

What are the Cloud and AI Development Act assurance levels?

The Cloud and AI Development Act (CADA) establishes 4 assurance levels for cloud and AI sovereignty. Levels 1 and 2 focus on EU infrastructure and transparency, while Levels 3 and 4 require the provider to be owned and controlled from within the EU, preventing third-country legal interference.

Sources

  1. 20% of EU enterprises use AI technologies ec.europa.eu
  2. <span class="mw-page-title-main">Use of artificial intelligence in enterprises</span> ec.europa.eu
  3. Delays to Frontier AI in the EU and UK | GovAI governance.ai
  4. GDPR — not AI Act — delayed release of frontier AI in Europe, research shows | MLex | Specialist news and… mlex.com
  5. Article 113: Entry into force and application ai-act-service-desk.ec.europa.eu
  6. EDPB opinion on AI models: GDPR principles support responsible AI | European Data Protection Board edpb.europa.eu
  7. Cloud and AI Development Act digital-strategy.ec.europa.eu

Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score

Related reading

All Adoption →

Data residency is not data sovereignty

Discover the limits of data residency for AI inference in Europe. Learn how jurisdiction, operational control, and data retention impact true sovereignty.

A tape library aisle in the CERN computer centre