Signal Atlas

Research library

AI-native networks

AI built into the network and computing pushed to the edge, close to people and machines. What 6G adds: AI is a core design goal of 6G, not an add-on.

Of 26 findings in this theme, 14 are real today, 6 designed and 6 forecast.

Real today 14

  • 3GPP adds AI predictions to how base stations cooperate. 3GPP finished AI/ML data collection and signalling between 5G base stations in late 2023 and, in Release 19, added predicted slice and coverage information. Release 20 candidates include AI for energy saving and mobility.
  • 5G-Advanced Release 19 puts AI into the radio link. Release 19 turned a Release 18 study into specifications for AI-based beam management and positioning, plus phone-side prediction of channel conditions. A harder two-sided AI method for compressing channel feedback is still being studied in Release 20.
  • AI controller cut Verizon cell-site energy 15% on average. Verizon, Samsung and Qualcomm put an AI energy-saving app on a multi-vendor RAN controller in Verizon's commercial network. Field tests showed 15% average energy savings and up to 35% per sector in quiet hours, by switching parts of cells off when traffic is low.
  • AI-RAN Alliance grew from 11 to 132 members. The AI-RAN Alliance launched at MWC in February 2024 with 11 founders including NVIDIA, SoftBank, T-Mobile, Ericsson, Nokia, Samsung, Microsoft and AWS. It had 132 members by March 2026, when it showed 33 demos of AI across network layers.
  • Bell rolls out 5G standalone as step to 6G. Bell says its 5G+ Advanced service, built on a standalone 5G core, is rolling out across the Greater Toronto and Hamilton Area and describes it as paving the way towards 6G.
  • Edge and fog computing keep urgent decisions local. Health IoT spreads processing across the device, phone, edge or fog servers and cloud, balancing delay, bandwidth, privacy, resilience and energy. The nanonetwork design likewise puts a fog layer near the body for faster local control.
  • Edge-computing standards now reach small devices and developers. ETSI's Multi-access Edge Computing group published its first Phase 4 (2024-2026) specifications in September 2025. They simplify network APIs for app developers, align with the open-source CAMARA APIs and extend edge computing to constrained and far-edge devices.
  • Low latency needs computing at the network edge. Trials cited by the IEEE paper found that 5G only reaches low delay if data and processing move to the cell tower. A remote robot link over an ordinary network suffers hundreds of milliseconds; 4G radio adds about 50-100 ms; a well-designed 5G network can go below 50 ms; and light itself needs about 5 ms for a London-Turin round trip.
  • Open RAN specs now cover AI model management. The O-RAN Alliance's catalogue reached 157 unique specification titles by July 2026, including APIs for training, deploying and storing AI/ML models in its RAN Intelligent Controllers and energy-saving policies such as deep hibernation. It completed Specification Release 5 in June 2026.
  • Personal AI agents that act, not just chat. AI assistants are moving from chat to taking actions with permission, keeping memory across devices, and processing personal data on-device or in verifiable private clouds.
  • Robot AI is also built to run offline. In June 2025 Google DeepMind released a robotics AI model that runs on the robot itself, including on Apptronik's Apollo humanoid. The aim is quick reactions and keeping working where connectivity is poor.
  • SoftBank ran base-station radio processing on GPUs. In an October 2025 outdoor trial at NVIDIA's California campus, SoftBank's AITRAS ran massive-MIMO radio processing fully in software on an NVIDIA GPU. Sending 16 data layers at once roughly tripled throughput versus a 4-layer setup.
  • Software sensors estimate, they do not measure. A virtual sensor is software that estimates a quantity such as stress, sleep or risk from other readings. Apps often show only the final label.
  • Virtual worlds train robots and self-driving cars. VR and simulation spaces are being used to train the AI inside robots, drones and autonomous vehicles before they meet the real world. Persistent simulated worlds could also test whole city designs.

Designed, not deployed 6

  • 6G architecture study covers AI agents and computing. 3GPP's 6G architecture study, approved June 2025, has tasks beyond connectivity: AI, sensing, data frameworks and computing. It treats AI agents, software that acts toward goals for someone, as network users.
  • 6G network would train and run AI itself. ITU's AI-and-communication scenario has the network collect data, train AI models across many sites, share them and run them close to users, for uses from medical-assistance devices to digital-twin predictions.
  • Adaptive systems and models drift after release. Learning systems can change after they ship, and people change too through illness, age or recovery. Models then drift and need recalibration, retraining or a pause on automated action.
  • ITU group is defining what AI-native networks mean. ITU-T Study Group 13 set up a focus group on AI-native telecom networks in July 2024. It works on definitions, use cases, architecture and proof-of-concept 'build-a-thons', and was still meeting in July 2026.
  • Robots and phones become decision-making network nodes. The IEEE paper expects telecom control to shift to massively distributed systems in which phones, cars and robots negotiate their own network resources. Software-defined networks and virtualised functions already make networks largely software-run.
  • US plan fast-tracks permits for AI data centres. The July 2025 US AI Action Plan recommends faster environmental permitting for data centres and opening federal land for them and their power plants, to speed up the computing that AI needs.

Forecast 6

  • Analysts expect AI-RAN to replace spending, not grow it. Dell'Oro forecasts about $35 billion of AI-RAN revenue over 2026-2030 and expects AI-RAN to be roughly a third of the RAN market by 2029. It says total RAN spending stays flat (about $160 billion over 2025-2029) and near-term AI-RAN is mostly non-GPU.
  • Billions of linked devices may show swarm-like intelligence. The paper predicts that once enough IoT devices are connected, AI could steer clusters of them so they behave like a flock or swarm, with abilities none has alone. It expected this to emerge in the 2020s with 5G as the communication fabric.
  • Body and IoT devices may form distributed decisions. The IEEE white paper projects that body-area and IoT devices may talk to each other, share decisions and eventually show emergent intelligence. Today this mostly means a chain: a sensor detects, a controller decides, a cloud model scores, a clinician acts.
  • Forecasters expect big AI capability jumps before 2030. Forethought research summarised in the research expects major AI capability gains before 2030, with automated AI research possibly compressing scientific progress, limited by chips, energy and supply chains.
  • NVIDIA invests $1 billion in Nokia for AI-RAN. In October 2025 NVIDIA agreed to invest $1 billion in Nokia, which will add NVIDIA-based AI-RAN products; T-Mobile US planned field trials starting in 2026. The companies claim operators could move from 5G-Advanced to 6G by software upgrade.
  • TELUS plans Canada's first commercial AI RAN controller. TELUS and Samsung announced in September 2025 that they would deploy Canada's first commercial RAN Intelligent Controller, including an AI energy-saving manager that predicts traffic and switches power-saving features on automatically.

Other themes