The pilot spreads AI models across central cloud and regional sites, aiming for faster responses, resilience if a site fails and voice data kept local where required.
Orange Business and NVIDIA are piloting distributed voice AI inference, a way to help companies deploy real-time conversational agents with lower latency, greater resilience and more control over where AI services run.
The approach spreads AI models across central cloud and regional points of presence (PoPs), distributing workloads to improve responsiveness, capacity and data location. Lower latency allows a more natural conversational flow, the companies said, while the setup optimizes GPU capacity and bandwidth use and can keep voice data local where required.
Also Read: Are Your AI Agents Ready to Be Trusted?
A proof of concept for real-world AI operations
At GTC 2026, Orange and NVIDIA are piloting distributed AI inference across multiple sites in Orange’s network. Orange Business deployed NVIDIA Blackwell in its central cloud and NVIDIA Aerial RAN Computer at distributed locations.
The demonstration uses a wildfire emergency scenario. AI translates the conversation in both directions between a foreign-language caller and the operator, meeting the required translation accuracy and response time. It also transcribes the call and prepares summaries of key information so emergency operators can act quickly.
Models are accessed through Live Intelligence, Orange’s secure, trusted multi-LLM (large language model) platform. A separate AI-distributed inference platform manages workloads across the central cloud and regional PoPs. Integrated routing balances traffic between sites and can redirect requests if a site is unavailable or overloaded.
To build the service, Orange Business is integrating NVIDIA Nemotron models into Live Intelligence. The open-weight models let customers audit, tailor and deploy them to meet their operational and data-residency requirements. Orange Business also supplies the network, cloud, and regional infrastructure, model management, and request routing needed to run distributed AI services end-to-end. That gives customers more control over data processing, service operations, and AI deployment locations, thereby supporting data residency and sovereignty requirements.
Expanding beyond conversational AI
The service is designed to support other uses, including emergency communications, live multilingual translation, enterprise contact centers, video analytics, drone detection, computer vision and the industrial internet of things.
Separately, Orange is assessing AI radio access network (AI-RAN) technology with Nokia and NVIDIA, initially focusing on potential gains in radio spectral efficiency, network capacity and site economics. The assessment also explores whether shared infrastructure could support both RAN functions and AI inference for future enterprise services, such as sensing and drone detection, without degrading radio network performance.
The distributed voice AI prototype is being tested with customers on real-world use cases. Their feedback will help Orange Business refine the architecture, particularly its performance and resilience in a trusted environment.
Also Read: Autonomous AI Fails Without a Real-Time CDP Memory Layer
“Distributed infrastructure is the next frontier of the AI factory buildout, as demand for inference and edge computing accelerates,” said Ronnie Vasishta, senior vice president of telecom at NVIDIA. “Telcos are uniquely positioned to become distributed infrastructure service providers. With NVIDIA, Orange is expanding from the central to distributed sites, powering a new class of real-time intelligence applications that need deterministic performance, reliability and scale, and can be delivered together with connectivity using the AI-RAN architecture.”
“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,” said Anne-Marie Thiollet, chief product and marketing officer at Orange Business. “Our extensive network and infrastructure footprint enable us to combine these capabilities together and help customers deploy AI services with greater performance, resilience and control.”
