EU Data Act: Cloud Switching, Interoperability, and What It Means for Your AI Infrastructure

The EU Data Act (Regulation 2023/2854) became applicable in September 2025, and its cloud switching and interoperability provisions are already reshaping how enterprises think about cloud strategy. For AI infrastructure — where vendor lock-in is deepest — the implications are significant.

What the Data Act Requires

Cloud Switching Rights (Articles 23-31)

The Data Act gives customers the right to switch between cloud providers and port their data:

• Switching period: Cloud providers must support customer switching within a maximum transition period — and cannot charge switching fees beyond the actual cost of switching
• Data portability: All customer data (including metadata, configurations, and derived data) must be exportable in a structured, commonly used format
• Functional equivalence: Providers must offer sufficient information to enable the customer to achieve functional equivalence in the destination environment
• No contractual barriers: Contract clauses that obstruct or penalize switching are void

Interoperability Requirements (Articles 28-31)

• Open interfaces: Cloud providers must provide open interfaces for interoperability with other services
• European standards: Where harmonized standards exist, providers must comply
• API accessibility: APIs must be documented and accessible to facilitate data portability and interoperability

Impact on AI Infrastructure

AI infrastructure has the highest vendor lock-in in enterprise IT. The Data Act forces a rethink:

Training Data and Datasets

• Data lake portability: Training datasets stored in provider-specific formats (BigQuery, Redshift, Synapse) must be exportable. Plan for open formats (Parquet, Delta Lake, Iceberg) from the start.
• Feature stores: Features computed and stored in provider-specific feature stores create lock-in. Consider open-source alternatives (Feast) or ensure export capabilities.
• Data labeling: Annotations and labels created using provider tools must be exportable in standard formats.

Model Artifacts

• Model format: Models trained using provider-specific frameworks may not be portable. ONNX and standard model formats improve portability.
• Model registry data: Metadata, lineage, and governance data in provider-specific model registries must be exportable.
• Training configurations: Hyperparameters, training scripts, and pipeline definitions should be provider-agnostic where possible.

Inference Infrastructure

• Serving endpoints: Models deployed on provider-specific serving infrastructure (SageMaker, Vertex AI) create operational dependency. Kubernetes-based serving (KServe, Seldon) provides portability.
• Auto-scaling configurations: Provider-specific scaling policies don't port. Abstract scaling logic from provider implementation.

Architecture for Portability

Compliance Checklist

• Inventory all cloud services used for AI workloads with provider lock-in assessment
• Document data formats used for training data, feature stores, and model artifacts
• Verify cloud contracts comply with Data Act switching provisions
• Create switching plans for critical AI services with estimated timeline and cost
• Test data export from each cloud AI service at least annually
• Evaluate provider-agnostic alternatives for highest lock-in services
• Include Data Act compliance in vendor selection criteria for new AI infrastructure

Strategic Implications

The Data Act doesn't require multi-cloud. It requires the ability to switch. This is an important distinction:

• Single cloud is fine — as long as you can demonstrate the ability to move if needed
• Managed services are fine — as long as data is exportable and you have a documented alternative path
• Deep integration is fine — as long as you understand the switching cost and have it documented

The risk isn't using a single cloud provider. The risk is being unable to leave when business, regulatory, or geopolitical circumstances require it.