Orange Business and NVIDIA pilot voice AI across cloud and regional network sites

Orange Business is putting voice AI across central-cloud and regional network sites, promising quicker conversations and services that can withstand disruption. Its October 21, 2026, collaboration announcement with NVIDIA describes a prototype undergoing customer testing—not a finished deployment with published performance results. A wildfire emergency-call demonstration illustrates what the companies want it to handle.

The proposal combines computing capacity with Orange’s network footprint. Instead of placing all AI processing in one central location, the architecture distributes models and workloads across central-cloud resources and regional network sites, known as points of presence. Orange says that arrangement can reduce response delays, use GPU capacity and bandwidth more efficiently, and help keep voice data local where required.

A conversation routed across multiple sites

For the pilot, Orange has deployed NVIDIA Blackwell hardware in its central cloud and NVIDIA Aerial RAN Computer at distributed locations. The companies describe the work as distributed AI inference: running trained models to process requests across multiple sites in Orange’s network. The announcement places the pilot at GTC 2026, with central and regional computing resources working within the same architecture.

Routing is a central part of the design, not simply the presence of hardware in several places. Orange says integrated routing balances traffic between sites and can redirect requests when a location is unavailable or overloaded. That is the stated mechanism behind the resilience promise: requests can move to another site rather than remain directed at one that cannot handle them.

The software also separates access to models from management of the computing work. Users access AI models through Live Intelligence, Orange’s platform for working with multiple large language models. An AI Distributed Inference platform manages workloads across the central cloud and regional sites.

Distributed inference will be a critical part of the next phase of enterprise AI. It requires compute, connectivity and orchestration across multiple sites and regions.

Anne-Marie Thiollet, Chief Product and Marketing Officer, Orange Business

An emergency call as the demonstration

The demonstration uses a wildfire emergency scenario involving a caller and an operator who speak different languages. It brings together three jobs during the call, rather than presenting a conversational agent that only answers questions:

• Translate the conversation in both directions, allowing the caller and emergency operator to communicate in their respective languages.
• Transcribe the call, producing a written record alongside the spoken conversation.
• Prepare summaries that capture key information to help emergency-service operators act quickly.

Orange presents translation accuracy and total response time as requirements for that experience. But the announcement supplies no measured translation accuracy, response times or resilience results. The scenario therefore describes the intended workflow, not evidence that the system has met a published emergency-service performance standard.

Model choice and processing location

Orange is also integrating NVIDIA Nemotron models into Live Intelligence. These are open-weight models, whose learned parameters are available for customers to inspect and adapt. Orange says that flexibility lets customers audit, tailor and deploy the models according to operational needs and data-residency requirements—the rules governing where data is processed or kept.

The proposed offering joins those model controls with network infrastructure, cloud resources, regional computing and request routing. Orange is positioning the combination as a way for enterprises to control both where AI runs and how the service operates. Customer feedback will now help refine the prototype’s architecture, particularly its performance and resilience in real-world use cases.