The EU Data Act's Cloud-Switching Rules and AI Infrastructure

From 12 January 2027 a cloud provider may not charge a customer for leaving. That is a real change and it is not the most consequential part of Chapter VI for anyone running inference, because the strongest guarantee in the chapter does not apply to the services AI workloads mostly use.

What Chapter VI covers

Chapter VI of Regulation (EU) 2023/2854, Articles 23 to 31, addresses switching between data processing services. The definition is broad: a digital service enabling ubiquitous and on-demand network access to a shared pool of configurable computing resources, provided as a service. That covers infrastructure, platform and software services alike, which means a managed inference API, a vector database service and a GPU cluster are all in scope even though the chapter treats them differently.

The Regulation became applicable on 12 September 2025, so the contractual and technical obligations discussed below are in force now. The charge prohibition is separately timed, which is the part most often misreported.

The charge phase-out and its date

Article 29 sets out a two-stage regime. From 12 January 2027, providers of data processing services shall not impose any switching charges on the customer for the switching process. In the transitional window that runs until that date, providers may impose reduced switching charges, and those charges must not exceed the costs incurred by the provider and directly linked to the switching process concerned.

Two clarifications that matter commercially. First, egress fees charged in the ordinary course of using a service — routine data transfer that is not part of a switch — are not the same thing as switching charges, and the Regulation targets the latter. In practice several major providers have announced free egress for customers leaving, which is a commercial response in the same direction, but a voluntary programme with conditions is not the statutory right and you should read which one you are being offered. Second, the transitional cap is a cost cap: a provider charging in that window must be able to relate the charge to costs actually incurred for that switch, which is a materially different posture from a published egress rate card.

Article 29 also prohibits, from the same date, charges associated with the switching process for the parallel use of services during the transition. That detail is what makes a real migration feasible: running old and new in parallel while you cut over is the normal shape of a migration, and being charged twice for it was a live deterrent.

The contractual duties in Article 25

Article 25 requires the rights of the customer and the obligations of the provider in relation to switching to be set out clearly in a written contract, and specifies several terms it must contain. The structural ones:

• A maximum notice period for initiating the switch, which shall not exceed two months.
• A mandatory maximum transitional period of 30 calendar days, during which the provider assists the switch and the service continues. It may be extended where technically unfeasible within 30 days, on notice, with the extension bounded by the Regulation.
• An exhaustive specification of the categories of data and digital assets that can be ported during the switch, and separately of those that cannot — including any that are specific to the provider’s internal functioning.
• A minimum retrieval period after termination of at least 30 days, during which the customer can still retrieve its exportable data.
• Erasure after the switch. The provider must erase all exportable data and digital assets once the retrieval period ends, which is the mirror image of the retrieval right and the clause to check against your own retention schedule.

The exhaustive specification is the sleeper obligation. It forces a provider to write down, in the contract, what you cannot take with you. For AI infrastructure that list is where the interesting content is: index structures, tuned service configurations, cached embeddings, managed-service state. Reading it before signing tells you more about your real lock-in than any architecture diagram.

Functional equivalence, and who gets it

Article 30 contains the chapter’s strongest technical duty, and it is deliberately narrow. Where the service concerns infrastructure — the IaaS layer — the provider must take reasonable measures to enable the customer, after switching, to achieve functional equivalence in the use of the destination service. Functional equivalence is defined in Article 2 as re-establishing, on the basis of the customer’s exportable data and digital assets, a minimum level of functionality in the environment of the new service such that the service delivers a materially comparable outcome for the same input features.

For services other than infrastructure — which is where platform and software services sit — the duty is different and weaker: providers must make open interfaces available, free of charge, and where relevant make available the necessary information for the customer to implement compatible services, with reference to open interoperability specifications or European standards where they exist. That is an interface-and-export obligation, not an outcome obligation. Nobody has to make your managed service work the same way somewhere else.

Article 31 then carves out further: certain services provided as a non-production test version, and custom-built services not offered at broad commercial scale, are subject to a reduced regime, and providers must inform prospective customers before contract of any such exemption. Bespoke arrangements — a negotiated capacity agreement for accelerators, for instance — are the shape most likely to fall near that line.

What this changes for an AI workload

Map your stack against the two duties and the picture becomes concrete.

Raw compute is well covered. GPU instances, block storage, object storage and networking are infrastructure, so functional equivalence applies and the charge prohibition will bite hard on the data-transfer costs of moving a multi-terabyte corpus or a set of checkpoints. For training workloads with large datasets, the January 2027 change is a genuine reduction in the cost of leaving.

Managed model services are not. A hosted inference endpoint, a proprietary foundation model available only from one provider, a managed fine-tuning service or an integrated agent platform sit above infrastructure, so the open-interface duty applies rather than functional equivalence. And there is a harder limit underneath the legal one: no obligation to provide open interfaces can give you the same model somewhere else. If your application depends on the behaviour of a specific proprietary model, the Data Act moves your data and cannot move your model, and no future amendment will change that.

The exportable-data definition is where to focus. Exportable data broadly means data the customer imported or generated through its use of the service, and excludes data that is specific to the provider’s internal functioning and protected by intellectual property rights or constituting a trade secret. For an AI stack, that boundary decides whether you can export a vector index or only the source documents, whether tuned weights produced by a managed fine-tuning service come with you, and whether evaluation history and prompt versions are yours. Read the Article 25 specification for the answer rather than inferring it.

Two practical moves follow. First, treat the Article 25 clause list as a procurement checklist at renewal: ask for the exhaustive specification in writing, and the answers you get are a better lock-in assessment than any internal estimate. Second, where you need genuine portability at the model layer, the Data Act will not supply it and the architecture has to — keep prompts, evaluation sets and orchestration in your own repository rather than in a provider’s console, and keep the interface to the model narrow enough that a second implementation is a week of work rather than a quarter. The relevant contract terms to negotiate alongside this are covered in AI contract clauses, and the sibling Chapter II regime is in connected-device data and AI training.