Regulation (EU) 2023/2854 — Article 23

Article total: 1 part · 0 Council drafts · 1 Parliament amendment

Tracked part

Removed wording is struck through; added or replacement wording is highlighted.

Institutional text

European Commission proposal

The wording proposed by the Commission at the start of this legislative file.

No standalone Commission wording is mapped to these tracked parts. A newly proposed provision may have no earlier text of its own.

Institutional text

Council Presidency texts

Successive Presidency compromise texts. Their inclusion does not imply agreement or adoption.

No Council wording is mapped to these tracked parts.

Competing proposals

European Parliament amendments

These are alternative tabled amendments. An amendment affecting several tracked parts appears once here, with each target identified.

Search amendments

Additional proposed wording Amendment 684 ITRE · LIBE

14a. In Article 23, the following paragraph 2b is inserted:

For the switching of services involving the training or operation of artificial intelligence models or other compute-intensive workloads, the obligations under paragraph 2 shall be understood as requiring functional portability. Functional portability is ensured where the provider of data processing services enables the customer to:

(a)

export trained model parameters, including model weights, in commonly used and interoperable formats;

(b)

obtain the configuration necessary to redeploy the workload in an alternative environment, including through infrastructure-as-code templates or equivalent means.

Proprietary compute optimisations that are specific to a given hardware or software environment, and that cannot reasonably be replicated in an alternative environment, shall not be subject to the obligations under paragraph 2, provided that the provider offers equivalent functional performance guarantees for the redeployment of the workload. Such optimisations shall not be used to create obstacles to switching within the meaning of point (a) of paragraph 2.

The technical standards and formats referred to in this paragraph, including any exceptions specific to artificial intelligence and compute-intensive workloads, may be specified in the implementing acts referred to in Articles 30(3) and 33(5).

Justification

Article 23 requires providers to enable data portability, but that requirement is technically unachievable for many AI workloads, in particular large models trained on custom high-performance computing clusters relying on hardware-specific optimisations. The amendment introduces functional portability: export of model parameters in interoperable formats and provision of the configuration necessary to redeploy the workload. Proprietary compute optimisations are exempted only where the provider offers equivalent functional performance guarantees, and may not be used to create obstacles to switching.

Preview
against:

Article 23

Removing obstacles to effective switching

Providers of data processing services shall take the measures provided for in Articles 25, 26, 27, 29 and 30 to enable customers to switch to a data processing service, covering the same service type, which is provided by a different provider of data processing services, or to on-premises ICT infrastructure, or, where relevant, to use several providers of data processing services at the same time. In particular, providers of data processing services shall not impose and shall remove pre-commercial, commercial, technical, contractual and organisational obstacles, which inhibit customers from:
2b.
For the switching of services involving the training or operation of artificial intelligence models or other compute-intensive workloads, the obligations under paragraph 2 shall be understood as requiring functional portability. Functional portability is ensured where the provider of data processing services enables the customer to:
1. (a)
  
  export trained model parameters, including model weights, in commonly used and interoperable formats;
2. (b)
  
  obtain the configuration necessary to redeploy the workload in an alternative environment, including through infrastructure-as-code templates or equivalent means.
3. Proprietary compute optimisations that are specific to a given hardware or software environment, and that cannot reasonably be replicated in an alternative environment, shall not be subject to the obligations under paragraph 2, provided that the provider offers equivalent functional performance guarantees for the redeployment of the workload. Such optimisations shall not be used to create obstacles to switching within the meaning of point (a) of paragraph 2.
4. The technical standards and formats referred to in this paragraph, including any exceptions specific to artificial intelligence and compute-intensive workloads, may be specified in the implementing acts referred to in Articles 30(3) and 33(5).
(a)

terminating, after the maximum notice period and the successful completion of the switching process, in accordance with Article 25, the contract of the data processing service;
(b)

concluding new contracts with a different provider of data processing services covering the same service type;
(c)

porting the customer’s exportable data and digital assets, to a different provider of data processing services or to an on-premises ICT infrastructure, including after having benefited from a free-tier offering;
(d)

in accordance with Article 24, achieving functional equivalence in the use of the new data processing service in the ICT environment of a different provider of data processing services covering the same service type;
(e)

unbundling, where technically feasible, data processing services referred to in Article 30(1) from other data processing services provided by the provider of data processing services.

Article 23

European Commission proposal

Council Presidency texts

European Parliament amendments

Article 23 – paragraph 2b (new)

Removing obstacles to effective switching

Compare wording