Tech • AI • Robotics • Game

ENFR

From AI Infrastructure to the 6G Future

7/10
NVIDIANVIDIANovember 3, 2026 at 01:00 PM37:56
Audio player

TL;DR

Telecom operators are rapidly turning their networks, sites and edge infrastructure into AI platforms, using distributed power, fiber and software-defined RAN to open new revenue streams, improve performance and prepare for 6G.

KEY POINTS

Telecom’s distributed infrastructure is becoming AI capacity

Roughly 100 gigawatts of power embedded across telecom infrastructure and about 100,000 sites including mobile switching offices, exchanges and points of presence are emerging as strategic assets for the AI economy. As hyperscale development runs into land and power constraints, carriers’ distributed real estate, power availability and dense connectivity are becoming valuable for edge AI and distributed inference.

Connectivity is shifting from utility to core AI enabler

Fiber, backplane links and over-the-air connectivity are increasingly central to linking AI factories, moving models and serving low-latency applications. Telecom networks are no longer positioned only as pipes for consumer and enterprise traffic; intelligent connectivity is becoming a foundational layer for AI deployment across manufacturing, robotics, emergency response and cloud services.

Operators are building AI factories and monetizing them

Carriers are moving from GPU as a service toward models as a service, inference as a service and application delivery. In North America, Telus reported that demand for its first AI factory was fully sold out, a signal of how fast carrier-hosted AI capacity is being commercialized. Across Asia and Europe, operators are increasingly discussing monetization, demand growth and returns on investment in terms rarely heard in telecom a few years ago.

AI-RAN is moving into mainstream deployment

One of the biggest shifts is the move toward AI-RAN, where radio access networks run on the same accelerated computing stack as AI workloads. The change depends on software-defined RAN reaching a performance-per-watt threshold comparable to proprietary systems. Momentum has accelerated with 19 leading telcos involved and 147 members in the AI-RAN Alliance, up from an initial group of 11 companies in 2024.

Deutsche Telekom and Nokia showed uplink gains

Deutsche Telekom is expanding its industrial cloud into a distributed cloud model and piloting AI-RAN with Nokia. In lab tests, the companies reported 10% to 15% uplink performance improvement using Nokia software, reinforcing the case that AI and RAN can share one common infrastructure and software stack.

Edge AI demonstrated faster robotics and emergency translation

A robotics demonstration from Agile Robots compared a machine running its world model locally against one connected to an edge RTX Pro 6000 system. The edge-connected setup completed pick-and-place tasks more smoothly and with faster cycle times, illustrating why AI changes the economics of edge computing. Orange also piloted a real-time emergency translation service using a full-duplex speech model with tone and semantic recognition, showing how low-latency AI at the edge can support public safety use cases.

New RAN hardware is aimed at base stations and centralized sites

New systems were introduced for the Aerial RAN Computer family, including a 4U MGX platform from Supermicro and a 2U MGX platform from Dell, alongside earlier systems from Quanta and DGX Spark. The push is to show that GPUs can run at base stations and centralized RAN sites while supporting both radio workloads and AI applications. DGX Spark was described as capable of serving six cells, 100 megahertz and 4T4R.

Commercial deployments are starting to appear

Elisa demonstrated radio access software and an AI application running simultaneously on compact MGX systems, including a video-to-text workload. BT plans live trials in Birmingham in 2027 using Nokia software on this architecture. In the United States, T-Mobile completed a first commercial call on the new system design, an early sign that software-defined AI-RAN is moving beyond pilots.

Open-source tools are reshaping wireless research

Wireless research is also being opened up through software and simulation. A differentiable link-layer simulator called Sionna and the Aerial 5G stack have together drawn more than 1 million downloads, mostly from researchers and industry labs. A 6G developer program now has more than 8,000 members, and recent demonstrations included around 50 over-the-air demos on real hardware at Mobile World Congress.

Agentic AI is beginning to automate network operations

Operators and vendors are starting to use post-trained open models for troubleshooting, site inspection and predictive maintenance. Cisco is launching Crosswork AI powered by NeMo Retriever, aimed at network troubleshooting and hypothesis generation. TCS and BT developed a process in which site photos are analyzed immediately so installers can verify work before leaving, reducing evaluation cycles that previously took more than three weeks. NTT Data is also building predictive degradation tools to detect faults before they happen.

Nokia and Nvidia are pushing AI-native wireless economics

The Nokia partnership is centered on making wireless networks software-defined, programmable and AI-native for 5G and 6G. Nokia said it expects 50% spectral efficiency improvement by next year and 100% the year after, highlighting how AI-based optimization could materially change network economics while enabling services such as industrial robotics, healthcare access, city surveillance and voice AI for emergency response.

CONCLUSION

Telecom is moving from a connectivity business toward a distributed AI infrastructure business. The scale of existing sites, power and network reach gives operators a chance to become central players in the next phase of AI deployment.

Ask a question

More from NVIDIA