ITU 6G framework sensing and AI integration transforms network architecture

The integration of sensing and AI as native network functions shifts the value proposition from simple data throughput to environmental intelligence. For the streaming ecosystem, this enables hyper-precise spatial awareness for augmented reality and industrial robotics, but it also creates a new category of data assets derived from radio propagation rather than user input. As networks begin to 'see' without cameras, the strategic competition will center on who controls the semiconductor stack and the AI models interpreting these physical-world signals. Watch for how the EU AI Act's biometric identification rules are applied to radio-derived behavioral data as these technical trials move toward commercial deployment.

The ITU's IMT-2030 framework has catalyzed concrete operator trials that embed AI directly into radio access networks, foreshadowing the sensing-native architecture the standard envisions. In April 2025, Bell Canada and Ericsson conducted the world's first field test of AI-native link adaptation, demonstrating up to 20 percent higher downlink throughput and 10 percent improved spectral efficiency by using real-time neural networks to adjust transmission rates based on changing radio conditions. By early 2026, AT&T and Ericsson extended that work to a Cloud RAN stack running on Intel Xeon 6 SoC, marking the first demonstration of portable AI-native RAN software on commercial off-the-shelf hardware and signaling that the compute layer for future sensing workloads can run on general-purpose silicon rather than proprietary baseband.

The commercial and regulatory stakes around AI-native network functions are intensifying as operators scale these features across production traffic. In mid-2026, T-Mobile ran the first large-scale trial of Ericsson's AI-native scheduler across approximately 43 live sites in Los Angeles, New York, New Jersey, and Salt Lake City, achieving up to 10 percent spectral efficiency gains and 15 percent higher downlink throughput compared to rule-based methods. Ericsson's head of technology for Americas, Akhil Gokul, noted that the feature runs on purpose-built hardware or x86 platforms without requiring hardware swaps, meaning it could cover roughly half of T-Mobile's US footprint where Ericsson supplies the RAN. For AT&T, the feature could span its entire network once the Nokia replacement is complete. These deployments establish the operational precedent for the ITU's vision of networks that natively process environmental signals, and they will likely inform how regulators under the EU AI Act classify radio-derived behavioral data when sensing functions move from trial to commercial service.

On the technical side, the IMT-2030 sensing specifications demand positioning accuracy between 1 and 10 centimeters, a target that current AI-native RAN features are beginning to approach through improved channel estimation and beamforming. Ericsson and Intel jointly benchmarked AI inference models on Xeon 6 hardware during the AT&T trial, measuring inference speed and scalability as key metrics for determining whether general-purpose processors can handle the computational load of integrated sensing and communications. The results showed measurable improvements in AI inference speed, providing a scalable path for operators to deploy sensing-adjacent AI workloads without discrete accelerators. This hardware trajectory matters because the ITU's framework assumes networks will simultaneously communicate, sense, and compute, requiring the semiconductor stack to support all three functions within existing power and cost envelopes.