Generative AI has long been envisioned as a service simply plugged into standard Cloud infrastructure. That assumption no longer holds once AI is expected to deliver reliable results on industrial topics: sizing a mechanical part, anticipating a supply chain disruption, validating a regulatory configuration. A general-purpose language model, however powerful, lacks an understanding of physical reality and can therefore produce results that are not physically correct. This observation, made by Florence Hu-Aubigny, Executive Vice-President Research & Development at Dassault Systèmes and Chief Architect of the 3DEXPERIENCE platform, during OUTSCALE EXPERIENCES 2026, reflects the same rationale behind the design of AI Factories by OUTSCALE: an infrastructure built from the ground up for training and running industrial AI models, within a sovereign and secure framework.
An architecture designed to retain nothing
The first guarantee offered by AI Factories rests on a principle that is simple to state but demanding to implement: no data is stored there on a lasting basis. The AI Factory operates as a compute space, not a storage space: data passes through, computation runs, and everything is then returned to OUTSCALE’s sovereign IaaS infrastructure, where certifications reside and where data persistence is fully guaranteed. Arnaud Bertrand, CTO of OUTSCALE and Senior Vice President of OUTSCALE, Dassault Systèmes, sums up this architectural choice plainly: the goal is to ensure that, at no point, any trace of customer data remains in the compute space.
This non-persistence is paired with physical and logical segregation by customer: each organization has its own dedicated AI endpoint, guaranteeing the isolation of its workloads and its intellectual property. This architecture was built in partnership with Nvidia (based on a generation of GB200 and GB300 components, paired with H200 GPUs) and with Bull for hardware integration, providing state-of-the-art compute capacity while retaining full control over its implementation.
Elastic access to multiple model families
Through a dedicated control plane, AI Factories provide access to a pool of models that includes Nvidia’s NIM models, such as Nemotron, as well as the models OUTSCALE already works with, such as Mistral. OUTSCALE will also add Aleph Alpha and LightOn to this pool to round out its sovereignty offering, and will provide access to LLaMA and OpenAI models as well. This pool powers a multi-model agentic platform that can orchestrate, depending on the use case, whichever model is best suited, whether open, sovereign, or international.
This elasticity comes with a shift in how GPU resources are allocated. Rather than static allocation (where a customer pays for reserved capacity that isn’t always fully used), OUTSCALE is building dynamic allocation, adjusted to actual usage.
One standard, two deployment scales
AI Factories are deployed according to two complementary approaches. On one side, international zones in Europe, the United States, and Asia offer maximum elasticity and allow workloads to be distributed across several continents. On the other hand, a strictly SecNumCloud zone, isolated and self-contained, is reserved for the most sensitive workloads, with no possibility of spillover to international zones.
It is this combination: non-persistence, per-customer segregation, model diversity, dynamic allocation, and dual-speed deployment – that allows OUTSCALE to offer a credible response to organizations operating in critical, regulated sectors: healthcare, energy, aerospace, defense, and the public sector. These AI Factories, initially designed to meet Dassault Systèmes’ internal needs, are now being gradually opened up to all companies and institutions that share the same requirement: industrializing AI without ever compromising on the sovereignty of their data.
To find out more, watch the replays of Arnaud Bertrand’s and Florence Hu-Aubigny’s talks.
