Running AI on European Cloud: What Changes, and What Still Needs Checking

Imagine a manufacturing company preparing to use AI to summarise supplier quality reports. The demonstration works, the operations team wants to proceed, and the managing director asks whether the documents can stay within an approved European environment. Before signing off, the company needs to trace where those reports go, who can access them and what happens when a service fails.

The recent Dataiku-Scaleway partnership offers a useful starting point for that assessment. This article examines what the announced hosting option changes and how to evaluate the remaining dependencies before putting sensitive business information into an AI application.

What the partnership changes

Scaleway published its Dataiku partnership announcement on 6 October, with the release itself dated 29 September 2026. It describes an option for customers to run Dataiku's AI platform on Scaleway infrastructure. Dataiku supplies software for preparing data, building AI applications and governing their use; Scaleway supplies the underlying hosting infrastructure.

Our reading is that this gives existing Dataiku users another infrastructure choice to investigate while retaining their application platform. For a company already using that software, it creates a concrete question for its next architecture review: which projects could run there, with which supporting services and operating responsibilities?

The announcement describes joint work on use cases and deployment options. It does not specify a universal migration procedure, a complete list of available configurations or a customer-specific support arrangement. Ask the companies to confirm those details for the version and features you intend to use before placing a migration date in your business plan.

European hosting can be one requirement in that plan. Define the others separately: permitted processing locations, access to confidential information, continuity when a supplier is unavailable and the ability to change arrangements later. This gives the technical team specific conditions to demonstrate and management a clear basis for approval.

Follow a document through the whole application

For generative AI, the software that prepares a request and the model that produces the answer can involve different services. Dataiku's LLM Mesh, its system for managing connections to large language models, supports both external model providers and self-hosted models. A large language model is the software that interprets and generates text; choosing where the application runs therefore leaves a separate decision about where that model runs.

Return to our hypothetical quality-report assistant. Suppose it retrieves a report from company storage, selects relevant passages and sends them to a model to draft a summary. Draw that route on one page, naming the organisation and location responsible for each step, including where the final summary is saved.

We recommend recording the actual service address and approved region for every model connection. Ask which document passages, instructions and identifying details the application sends, whether the recipient retains them, and whether another service can take over when the chosen model is unavailable. Request contractual answers alongside a demonstration of the configured behaviour.

Make the information boundary concrete. In this example, supplier names might be permitted while drawings and personal contact details are excluded; those are hypothetical business rules for the team to agree. Test that excluded content stays out of requests and that an unavailable approved model causes the application to stop or follow an explicitly approved alternative.

Treat the resulting map as part of the application's operating instructions. Assign someone responsibility for reviewing it whenever a model, connection or data source changes. That makes the approval useful beyond the first demonstration.

Check the assistants used to build the application

The finished application is only one part of the review. Dataiku's own AI assistant documentation, showing version 15 when checked, distinguishes services operated by Dataiku from a mode using a customer-selected model connection. In the Dataiku-operated mode, inputs are processed by Dataiku and third-party AI providers; in the customer-selected mode, requests go directly to the chosen provider and Dataiku says it does not process those requests.

The same documentation says both modes require internet connectivity to a Dataiku gateway. It also describes third-party assistants whose connections may use the LLM Mesh or go directly to outside services. These are documented product capabilities, not evidence of which features any particular Scaleway deployment enables.

Our recommendation is to include development assistants in the same approval process as the business application. Confirm which features are enabled, what each receives and which terms apply to its use. A team should be able to explain both how the supplier report is summarised and what information a developer can send while asking an assistant to help build that workflow.

For the first trial, enable only the features needed to prove the use case. Give administrators a written rule for approving additional connections, with the information required for that decision. Review the documentation for the exact installed version rather than assuming every deployment behaves alike.

Include support in the data-flow review

Troubleshooting deserves its own test. Dataiku documents an instance diagnosis archive for investigating problems, containing logs, system information and, optionally, configuration details. The documentation says it excludes dataset contents and managed-folder data, and that customers can inspect the archive before sending it to support.

That distinction is useful. It avoids assuming that support automatically receives the underlying business documents, while making clear that other information can be shared. Our advice is to prepare and inspect an example archive from the trial environment, then agree who may approve its transfer and which channel they should use.

Request written answers about the organisations handling infrastructure and software support, their access arrangements and the escalation process. Include the destination and retention of submitted diagnostic material. Do not infer those answers from the location of the machine running the application.

For the hypothetical manufacturer, rehearse a failed overnight summary job. The exercise should show who investigates first, what evidence they collect, what can be shared externally and who tells operations that the summaries will be delayed. Give the business owner a usable fallback, such as reviewing the original reports manually, while the technical team works on the fault.

Make recovery cover the data as well as the software

Dataiku's backup documentation draws an important boundary: its application data directory contains configuration and projects, but excludes datasets held in external databases or cloud storage. The page also distinguishes deployment types, explicitly excluding Dataiku Cloud from those instructions. A backup procedure must therefore match the actual setup.

For our example, we would ask the team to demonstrate recovery of the complete supplier-report workflow. That means identifying the application configuration, original reports, stored summaries, access arrangements and model service needed to produce another usable result. Assign a backup or reconstruction method to each part, and name the person responsible for keeping it current.

Make the recovery test answer a business question: after an interruption, can the quality team find yesterday's reports and safely resume today's work? Record how long recovery takes and which information, if any, must be re-entered. Agree acceptable limits with the business owner before the system becomes part of a daily process.

Also test a planned change of model against a fixed set of representative reports. We would assess whether the replacement preserves the details that matter to the quality team, such as defect descriptions and unresolved actions. Treat an acceptable result as something to demonstrate again when the model changes.

Give the first deployment a clear acceptance test

We recommend starting with one bounded use case whose value and information requirements are easy to explain. For the manufacturer, that might be drafting summaries for a quality manager to review, with no automatic supplier messages or changes to production records. Use synthetic or otherwise approved test documents until the proposed data flows and access rules have passed review.

Agree the acceptance conditions before running the trial. The business owner should define what a useful summary contains and which mistakes make it unacceptable. The technical owner should demonstrate approved model connections, access restrictions, support procedures and recovery, keeping evidence that another member of the team can understand.

Include awkward cases in that exercise: an incomplete report, conflicting statements, a failed model request and an attempted connection to an unapproved service. Check that the application identifies uncertainty, preserves the source material for review and follows the agreed failure behaviour. Record unresolved limitations plainly rather than treating a successful demonstration as permission for every future use.

The decision at the end can be modest: approve this application, for this information, with these services and named owners. If it requires a model or support arrangement outside the company's approved boundaries, management can assess that specific trade-off or narrow the use case. The team then has an actionable decision instead of a general instruction to make AI sovereign.

For us, the value of a European hosting option is the opportunity to build and verify that concrete arrangement. A useful handover gives your team the deployment record, data-flow map, operating instructions and evidence from the trial. Those are the materials that let you explain what the application does today and judge whether a proposed change remains acceptable tomorrow.

Thinking about migration? Book a free consultation to discuss your situation.