Between US restrictions on exporting GPUs to China, uncertainties around access to AI models, and the extraterritorial reach of the Cloud Act, dependence on hyperscalers has become a strategic issue. When MLOps pipelines rely on infrastructure subject to external political decisions, the question is no longer technical: it becomes geopolitical.
The cloud battle is fought not only on performance but on sovereignty. To free oneself from the models provided by hyperscalers, it is essential to develop MLOps by leaning on sovereign clouds.
La dépendance aux clouds extra-territoriaux : un risque systémique
Entrusting your data and your AI pipelines to a provider under foreign jurisdiction implies embracing several risks: that a regulatory change could block access to a critical resource (GPU, framework, API) overnight; that a player could unilaterally terminate on-site support, as Atlassian did by forcing a migration to the cloud for its customers; that reversibility becomes a distant dream, as proprietary services trap organizations in locked architectures; or that data could be exploited by third parties.
For MLOps pipelines, this dependence translates into operational fragility, manifested by deployments and monitoring that rely on services beyond one’s control, compliance constraints that are impossible to guarantee on clouds governed by laws other than the GDPR, or a scalability capacity that hinges on the goodwill of external providers, themselves subject to geopolitical tensions. In short: building AI pipelines without control of the underlying stack is building on sand.
Vers une alternative souveraine, avec l’open source et un cloud européen
The good news is that another path exists. If the industry has long claimed that this alternative would be more difficult to implement and maintain, the era of LLMs and autonomous agents has brought open source and sovereignty to the forefront—more accessible and more relevant than ever. Especially in the AI landscape, where open-source tools constitute the vast majority of production-grade solutions, far ahead of proprietary offerings.
Constructing a sovereign MLOps stack rests on two pillars: first, open-source tools such as Kubeflow, MLflow, Feast, Kubernetes, and PostgreSQL, each specialized for a specific mission (orchestrating and industrializing ML workflows, managing model features, building the infrastructure, etc.). These building blocks enable a modular, scalable, interoperable, and durable base stack. They also come ready-to-use from most operators.
The second pillar is embodied by sovereign operators, offering GDPR-compliant environments, SecNumCloud-compliant, located in Europe, and committing to transparency and reversibility of infrastructure services. More than a toolkit, they provide a trusted space upon which to base services. It therefore becomes possible to design overflow flows between private infrastructures and cloud services. Hybridization provides disaster recovery or business continuity in case of incidents, but also resilience to shortages of compute resources like GPUs, and access to cutting-edge hardware without capital expenditure.
These two pillars converge into an MLOps platform that is transparent, with auditable code and identified dependencies, maintainable thanks to the absence of opaque proprietary dependencies, predictable because costs and scalability rest on clear contracts, and sovereign, preserving end-to-end control over data and models.
Un enjeu collectif
Digital sovereignty is not a defensive posture. It is a sustainability strategy. Building sovereign MLOps pipelines guarantees that AI innovations are not weakened by decisions taken outside the company. It also gives European ecosystems the capacity to compete—not through sheer scale, but through resilience and mastery.
To ensure that their AI models do not depend on decisions made in Washington, in Beijing, or even in Brussels, the companies that consume or supply these solutions must recognize that the stakes go beyond technique and touch the collective ability to innovate freely and sustainably.
Alexis Gendronneau is the Director of Data & AI at NumSpot
