Digital sovereignty represents a crucial issue for France, particularly in the face of the rapid rise of artificial intelligences. Hosting a model like ChatGPT on sovereign cloud infrastructures raises strategic questions. The initiative to implement a French and secure solution faces both technical and ethical challenges.
The question arises from a desire for data autonomy and security, which becomes fundamental in a globalized context. Actors like Linagora collaborate with various providers to build this innovative infrastructure. The possibility of making a French chat accessible relies on the optimization of local and certified resources, thereby ensuring regulatory compliance and user protection.
Technical challenge: deploying Lucie on a sovereign cloud
As part of the deployment of the open-source linguistic model Lucie, Linagora engages in discussions with several sovereign cloud players, including Exaion, Outscale, OVH, and Scaleway. This model, designed to compete with services like ChatGPT, poses considerable technical challenges, especially in terms of infrastructure. The Lucie model, which includes 7 billion parameters, aims to demonstrate the feasibility of an open-source solution in the context of data sovereignty.
The CEO of Linagora, Michel-Marie Maudet, emphasizes the goal of proving the viability of a local equivalent to ChatGPT at the upcoming Paris Open Source AI Summit, scheduled for January 22nd. This context highlights the stakes of AI technology and the performance expectations of sovereign clouds.
Capacity of sovereign clouds facing generative AI models
Hyperscalers, such as GAFAM, generally offer optimal solutions for large-scale deployments, with infrastructure capable of handling high traffic volumes. However, Linagora focuses on the offerings of sovereign clouds, motivated by the need to ensure data confidentiality. “No French cloud has yet completed such an operation,” asserts Maudet, while acknowledging that Scaleway stands out with its resources, such as the 1000 Nvidia H100 GPUs already deployed.
For Linagora, assessing the capabilities of clouds includes a meticulous analysis of different traffic scenarios and request volumes. The company must determine the best configuration to support its models and process user requests in real-time.
A multicloud approach to optimize performance
Taking into account various needs, Linagora envisions a multicloud architecture as a potential solution. Maudet specifies that “a single sovereign cloud will not be able to cover all our use cases,” emphasizing the need for interaction between different providers. This approach could combine the capabilities of various sovereign clouds, thus offering flexibility and redundancy in AI processing.
This multicloud configuration should enable Linagora to route requests to the most appropriate clouds, depending on whether users wish to store their data on a particular infrastructure. In this regard, Outscale, whose GPUs are labeled SecNumCloud, could play a central role in this optimization.
Responses from sovereign cloud providers
Some companies, like Scaleway, claim to already be ready to handle heavy workloads. The lead product manager AI, Frédéric Bardolle, highlights their ability to manage significant load increases, reaching up to several hundred thousand requests per second. This claim is further supported by the experience gained during the global launch of the voice chat Moshi, demonstrating that Scaleway possesses the necessary robustness.
Furthermore, OVHcloud presents its AI Endpoints service, capable of processing generative AI models while maintaining robust infrastructure layers. Their solutions are adapted to various model sizes, deploying GPUs according to the specific needs of clients, with a focus on efficiency and customization.
Outscale, for its part, emphasizes its ability to establish dedicated infrastructures for organizations requiring strict data protection. David Chassan, Outscale’s strategy director, mentions a portfolio of 350,000 clients, which gives the company a strong position in the cloud services market. This provider is committed to maintaining a secure working environment, meeting its clients’ requirements for data confidentiality.
Towards French autonomy in AI
The debate regarding the possibility of hosting a French version of a model similar to ChatGPT on sovereign infrastructures is generating growing interest. While several players in the cloud sector believe they can rise to this challenge, the issue of the adequacy of technical capabilities and compliance with data protection regulations remains central.
Initiatives by companies like Linagora aim to promote the development of AI ecosystems within the European Union while respecting digital sovereignty. Implementing high-performance linguistic models in a secure framework could encourage broader adoption of generative AI within French companies and institutions.
Frequently asked questions
Can we really host a ChatGPT-like model on a sovereign cloud in France?
Yes, it is possible to host a ChatGPT-like model on a sovereign cloud in France. Companies like Linagora are working with sovereign cloud providers to deploy AI models at scale while ensuring data security and sovereignty.
What are the main advantages of hosting a ChatGPT on a sovereign cloud?
Hosting a ChatGPT on a sovereign cloud ensures the protection of sensitive data, prevents data transfer across borders, and ensures compliance with local privacy regulations.
What cloud infrastructures are available to host a French ChatGPT?
There are several sovereign cloud providers in France, such as OVH, Scaleway, and Outscale, that offer suitable infrastructures for hosting AI models like ChatGPT.
Is it necessary to use a multicloud solution to host a ChatGPT?
While this is not mandatory, a multicloud approach may be necessary to address different use cases and ensure sufficient resources to handle a high volume of simultaneous users.
How can we ensure user security when using a ChatGPT hosted on a sovereign cloud?
To ensure security, it is essential to leverage security certifications, such as SecNumCloud, and implement robust security protocols to protect user data.
Can open-source language models be integrated into a sovereign cloud?
Yes, open-source language models, like Lucie developed by Linagora, can be integrated into sovereign cloud infrastructures, providing an alternative to large American solutions.
What types of machine resources are necessary for deploying a ChatGPT on a sovereign cloud?
The deployment requires powerful GPUs suitable for AI, such as Nvidia H100s, capable of handling heavy processing and high traffic loads.
What is the scaling limit of a ChatGPT hosted on a sovereign cloud?
The scaling limit depends on the infrastructure established by the sovereign cloud provider and can vary based on available machine resources and architecture configuration.
Is the sovereign cloud solution more expensive than the options from hyperscalers?
Generally, sovereign cloud solutions may represent a higher cost due to the resources and security guarantees offered. However, this cost can be justified by data protection and legal compliance.
What steps should be taken to deploy a ChatGPT on a sovereign cloud?
The steps include assessing resource needs, selecting a sovereign cloud provider, choosing appropriate machine models, and finally, integrating and configuring AI solutions on the chosen infrastructure.
