Telecommunications
How network operators use NVIDIA AI for GPU-based radio access networks, agent-assisted network operations, operator-hosted AI capacity and edge services on 5G, with network reliability, subscriber data protection and lawful data handling at the center of every design.
The problem
Mobile operators run radio, transport and core networks assembled from many single-purpose systems, each with its own management tools and alarm streams. Engineers spend much of their time triaging faults and tuning cell parameters by hand, and a change that looks safe on paper can degrade service for thousands of subscribers.
Traffic is shifting as well. Phones, cameras, robots, drones and AI agents now send AI workloads over the network, while radio spectrum stays scarce and costly. Operators also want revenue beyond connectivity and are looking at what they can sell from assets they already own: sites, power, fiber and long customer relationships with businesses and government.
Every change has to respect very high availability targets and strict rules on subscriber records, location data and lawful interception.12
The approach
NVIDIA's work with operators has four strands. In the radio network, NVIDIA AI Aerial runs RAN software on GPUs so that 5G processing and AI workloads can share one infrastructure, which NVIDIA presents as a route to 6G through software upgrades rather than hardware swaps. Its Aerial RAN Computer family is built from GPUs, BlueField-3 and Spectrum-X networking, and researchers model radio systems with the open source Sionna library and the Aerial Omniverse Digital Twin.
In operations, NVIDIA describes agents built on NVIDIA AI Enterprise that reason over network and business data and simulate a change before acting on it. For new revenue, operators put their sites and power to work as AI factories: Deutsche Telekom now provides computing capacity to companies, research institutions and the public sector from its Munich Industrial AI Cloud, whose hardware includes DGX B200 systems. NVIDIA also promotes Metropolis vision agents running at the edge of 5G networks, and its AI-RAN page highlights an announcement with T-Mobile on physical AI applications over AI-RAN ready infrastructure.
Not every operator needs GPUs in the RAN. Virtualized RAN on general-purpose servers or purpose-built baseband hardware serves current 5G needs for many networks, and plenty of operations work starts well with conventional analytics and rule-based automation from the existing OSS vendor. Hosting AI capacity for others only makes sense where there is local demand, available power and a sales channel to reach buyers.123
Conceptual architecture
Diagram as a list
Applications & solutions
- Edge AI services on shared RAN GPUs (Metropolis agents, NIM)Offers video search, monitoring and model APIs close to customers
- Operations agents on NVIDIA AI EnterpriseDiagnose faults and propose configuration changes for engineers to approveConnects to Cell sites and fronthaul (Spectrum switches, Aerial RAN Computer)
- AI services for enterprises and the public sectorSells compute, models and applications under local data rules
Models & frameworks
- Radio network digital twin (Aerial Omniverse Digital Twin)Simulates cells and city-scale coverage before changes go liveConnects to Operations agents on NVIDIA AI Enterprise
Inference & runtime software
- GPU-accelerated RAN software (NVIDIA AI Aerial)Runs Layer 1 and Layer 2 processing and shares spare GPU capacity with AI jobsConnects to Edge AI services on shared RAN GPUs (Metropolis agents, NIM)
Operations & orchestration
- Network and business data (OSS, BSS, telemetry)Supplies alarms, KPIs, configuration history and customer care records under access rulesConnects to Operations agents on NVIDIA AI Enterprise
Accelerated computing
- Operator-hosted AI factory (DGX systems)Provides in-country GPU capacity for training and inferenceConnects to AI services for enterprises and the public sector
Networking, power & facilities
- Cell sites and fronthaul (Spectrum switches, Aerial RAN Computer)Carries radio traffic to distributed units with the timing the RAN needsConnects to GPU-accelerated RAN software (NVIDIA AI Aerial)
Technologies and their roles
NVIDIA AI Enterprise2
Software base for operations agents and AI services
NVIDIA says operators build AI-powered operations on AI Enterprise and can use it to create new AI offerings on AI-RAN infrastructure.
NVIDIA NIM1
Packaged models for operator services
NVIDIA's AI-RAN page names NIM for operators building generative AI offerings delivered over shared AI-RAN infrastructure.
NVIDIA Spectrum-X Ethernet1
Fronthaul and AI-RAN networking
NVIDIA lists Spectrum-X in its Aerial RAN Computer and uses Spectrum switches to route fronthaul traffic to virtual distributed units.
NVIDIA Omniverse1
Radio network digital twins
NVIDIA says the Aerial Omniverse Digital Twin simulates wireless systems from a single cell to city scale.
NVIDIA Metropolis2
Vision agents at the network edge
NVIDIA presents Metropolis agents running at the edge of 5G networks for video search, monitoring and infrastructure inspection.
NVIDIA DGX3
Operator-hosted AI factories
Telekom's Munich AI factory for industrial, research and public sector users includes NVIDIA DGX B200 systems.
What you need first
- Alarm, KPI and configuration history collected in one governed data platform with clear ownership
- A lab or digital twin where changes are tested before they reach live cells
- RAN engineering skills in Layer 1 and Layer 2 processing, timing and fronthaul, if GPU-based RAN is in scope
- Power, cooling and space assessments for GPU servers at cell sites, aggregation points or central offices
- A classification of subscriber, location and lawful interception data that states which records may be used for AI and which may not
- A change process that keeps a human approver for any network change an agent proposes
- Identified anchor customers and a sales channel before building hosted AI capacity
Risks and how to reduce them
- Outages caused by automated network changes
- Start agents in advisory mode, test every change in a twin or lab, roll out in stages with automatic rollback, and keep emergency call paths away from experimental systems.
- Misuse of subscriber and location data
- Limit training data to what the purpose needs, pseudonymize identifiers, keep data in-country where law requires, and keep lawful interception data out of AI pipelines entirely.
- AI jobs starving RAN processing on shared GPUs
- Give RAN functions guaranteed priority and resources, and test peak-hour behavior before mixing AI workloads on the same servers.
- Security exposure from tenant AI services on network infrastructure
- Isolate tenant workloads from network functions, apply zero-trust access controls and patch edge nodes on a fixed schedule.
- Underused hosted AI capacity3
- Sign anchor customers before building and grow in phases; Deutsche Telekom reported more than a third of its Munich capacity in use at opening, with named industrial and research users.
Documented examples
Deutsche Telekom (T-Systems) · Telecommunications and cloud services
Deutsche Telekom Industrial AI Cloud: an NVIDIA-based AI factory in Munich
Deutsche Telekom opened its Industrial AI Cloud in Munich on 4 February 2026, with nearly 10,000 NVIDIA Blackwell GPUs in DGX B200 systems and RTX PRO Servers. Telekom says it was over a third utilized at opening, with Agile Robots and PhysicsX among early users.
In production
Related
Sources
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.