At OUTSCALE EXPERIENCES 2026, Mireille Mege, Public Sector Customer Success Manager at OUTSCALE (Dassault Systèmes), and Clara Pascolo, Chief Operating Officer at SMLB Next, explained how Nuage Public simplifies public-sector access to sovereign cloud, drawing on its partnership with OUTSCALE and the generative AI bundles from Mistral available on the OUTSCALE Marketplace.
For public administrations and parapublic organizations, public cloud provides a practical entry point to secure, compliant and scalable environments capable of adapting to business needs.
The success of these projects relies on a structured approach: identifying a relevant initial use case, ensuring the protection of sensitive data, selecting the appropriate level of support and anticipating industrialization from the earliest stages.
Nuage Public as a starting point
The UGAP Nuage Public marketplace provides access to cloud services for more than 26,000 public- and parapublic-sector beneficiaries. It enables organizations to define their needs, specify their technical and regulatory requirements, and then be directed to cloud providers capable of meeting them.
This approach facilitates the launch of artificial-intelligence projects, often in the form of a proof of concept. It responds to two main trends. Some organizations want to implement the Cloud au Centre doctrine promoted by DINUM. Others want to use AI in a more secure and controlled environment.
The Nuage Public marketplace combines several factors that are important for public procurement: negotiated prices, a simplified process, guidance toward relevant offerings and shorter deployment times. In some cases, a project can start in less than a month, while the purchase order can be processed very quickly when the requirements are clearly defined.
Public-sector priorities
Organizations launching a sovereign-AI project generally have three main expectations.
Controlling data location
The first question concerns where data is located and how it is processed. Users want to prevent sensitive data from being transferred to environments they do not control. This concern applies in particular to administrative data, health information, identity documents, financial data and defense-related information.
Sovereignty therefore becomes a central decision criterion. Data must be processed and stored in an environment that meets the organization’s requirements, with particular attention paid to access, permissions and the platform’s operating conditions.
Reducing shadow-AI risks
Employees may be tempted to use AI services directly available online, particularly to query language models with professional information. This practice creates a risk of losing control over the data entered and where it is sent.
A sovereign-AI environment provides users with a controlled framework. The organization can define authorized services, manage access and encourage the use of models hosted in an environment that meets its security requirements.
Preparing for industrial deployment
A project often begins with a limited use case. Once the initial results are positive, other needs emerge. The organization may then want to analyze more documents, automate additional steps or use several models.
The selected architecture must therefore be able to evolve. A scalable offering facilitates the transition from proof of concept to production and then to multiple use cases. This industrialization capability must be considered from the outset to avoid having to redesign the entire technical environment later.
Three AI project offerings
Public-sector organizations can choose between several approaches. The right option depends on the desired level of control, the number of models required, the available resources and the expected level of support.
Consuming compute resources
The first use case consists of providing the resources required to run language models. The organization consumes compute including CPU, RAM and GPU and then uses the models at its disposal.
This approach offers significant technical freedom. It is suitable for teams capable of deploying and administering their own environment. However, they must manage several components themselves, including resource configuration, Nvidia drivers and the components required to run the models.
Deploying a Mistral bundle
The second use case is based on Mistral bundles. This offering provides a ready-to-use solution, with a Mistral model deployed in a tenant dedicated to the customer. Deployment takes place through the Marketplace and can be completed quickly after the order has been placed.
The customer receives a directly accessible endpoint to begin its work. Mistral models are supplied with their licenses, and the environment is dedicated to the organization. Data therefore remains within the designated environment in France, in accordance with the characteristics of the offering.
The bundles offer unlimited token consumption. However, capacity remains determined by the available hardware resources, particularly the number of GPUs associated with the model and the use case. Depending on requirements, one or two GPUs may be offered.
This option is especially suitable for organizations that want to use a language model without managing the entire technical stack themselves. It reduces initial complexity and accelerates access to an operational solution.
Building a Mistral AI Studio platform
The Mistral AI Studio platform is designed for projects requiring multiple models and several usage scenarios. It may be appropriate when an administration wants to analyze documents, perform optical character recognition, verify information or cross-check a form against an identity document.
In this configuration, the offering is built together with the customer. Use cases are analyzed in advance to select the appropriate models. It is generally advisable to begin with two or three clearly defined use cases and then design a coherent solution around those needs.
This approach creates a customized and scalable environment. Mistral also supports the customer in tuning the solution and adapting it to the identified use cases. Additional support may be provided for Kubernetes-related operations, including model updates and platform upgrades.
Toward hybrid model consumption
Organizations may need several model families depending on their use cases. A Mistral model may be suitable for one scenario, while a Llama model or another open-source model may be more effective for another task.
This is leading toward hybrid model consumption. The stated objective for the first half of 2027 is to offer an AI Factory capable of adapting GPU consumption to the intensity of the use cases and the models being used.
Such an approach should make it possible to allocate resources according to actual needs. It opens the way to more flexible use of AI infrastructure, with greater adaptability across models, workloads and expected performance levels.
Moving from proof of concept to production
Choose a manageable first use case
The choice of the first use case influences the rest of the project. An organization benefits from starting with a clearly defined need that is useful enough to produce measurable results and simple enough to be mastered quickly.
This initial stage helps identify technical constraints, security requirements, support needs and monitoring arrangements. It also provides valuable feedback before AI is extended to more complex processes.
The approach can follow three priorities:
- select a concrete and relatively simple use case;
- measure results and resource consumption;
- prepare for expansion to other use cases.
Organize security workshops
Security workshops play an important role in projects involving several models or sensitive data. They help define authorized users, roles, access rights and the conditions governing use of the platform.
This preparation should take place before operational deployment. It helps establish a trusted framework and align the technical environment with the organization’s internal security requirements.
Security also concerns project governance. Teams must know what data may be processed, by which users and under what conditions. These rules facilitate adoption and limit uncontrolled use.
Plan for operational support
The required level of support depends on the selected offering and the skills available internally. An organization consuming compute resources directly must have the capabilities needed to administer its environment. A turnkey solution reduces this burden. A multi-model platform requires closer collaboration around use cases, models and their evolution.
Support may also cover Kubernetes operations, model updates, version upgrades and platform optimization. This becomes particularly important when the solution moves from experimentation to production.
Managing costs and environmental impact
Controlling consumption is a major issue in public procurement. The available budget must be considered alongside the resources used, including GPUs, tokens and other cloud components.
Monthly monitoring makes it possible to identify resources that have been ordered but are seldom used, and then adjust the environment. This consumption review also helps prepare for project expansion using actual data rather than initial estimates.
FinOps provides a framework for financial management. It makes it possible to compare deployed resources with observed usage and arbitrate between performance, sovereignty, cost, support and scalability.
Green IT complements this analysis. A carbon calculator can be used to track changes in the footprint associated with resource consumption. It may reveal overconsumption caused by virtual machines left running during inactive periods or by resources being deployed without a corresponding operational need.
Financial and environmental consumption monitoring should begin in the early stages of the project. It promotes more precise resource usage and supports decisions related to scaling.
Balancing sovereignty, performance and scalability
The choice of an AI environment cannot be based on a single criterion. Organizations must consider data location, security, expected compliance, performance, cost, support quality and scalability at the same time.
Sovereignty is particularly important when data is sensitive. In this context, an environment located in France and designed to meet security requirements provides an appropriate framework for public- and parapublic-sector projects.
Performance must then be assessed according to the use case. A model intended for document analysis does not necessarily require the same resources as an environment running multiple models and several simultaneous processes.
Finally, scalability must be incorporated into the decision. A solution suitable for a proof of concept may become insufficient as the number of users, documents or use cases increases. The initial choice should therefore preserve the possibility of gradual industrialization.
Conclusion
Sovereign AI on public cloud provides public-sector organizations with a practical framework for experimenting with, securing and industrializing their use of artificial intelligence. The Nuage Public marketplace simplifies access to cloud providers, cloud offerings and conditions suited to public procurement.
The most effective approach is to start with a manageable use case, secure access and data, monitor consumption and then prepare for expansion to multiple models and scenarios. Compute-resource offerings, Mistral bundles and the Mistral AI Studio platform make it possible to adapt the level of freedom and support to the project’s objectives.
Discover the solutions available on the OUTSCALE Marketplace
