AI-native wireless? Telcos accelerate deployment of intelligent radio access networks

Telecom vendors and operators are rapidly transitioning from conceptual AI integration to practical deployment of AI-native RAN, leveraging GPUs and CPUs to optimise performance, cost, and latency in next-generation networks, as industry experiments and partnerships accelerate the push towards intelligent wireless infrastructure.

AI-driven radio access networks are shifting from concept to practical strategy across the telecom ecosystem, with major vendors and operators repositioning roadmaps to embed machine learning throughout the RAN. According to NVIDIA and its partners, efforts unveiled this year to build AI-native wireless stacks and reference platforms are designed to marry radio functions with accelerated AI inferencing, signalling a push to make networks themselves active participants in AI workloads.

Momentum accelerated at industry events where demonstrations and partnerships showcased AI-in-the-RAN use cases and outdoor trials. NVIDIA and collaborators reported outdoor, over-the-air implementations with carriers such as T-Mobile, SoftBank and Indosat, while vendors described new collaborations to bring AI to production-grade RAN deployments. Ericsson outlined a vision for networks that transition from carrying intelligence to being intelligent systems, reinforcing the industry-wide pivot observed at Mobile World Congress and related conferences.

Market dynamics are shifting alongside these technical developments: analyst commentary points to a large and rising allocation of investment toward AI infrastructure as firms prepare for an increase in AI traffic and edge workloads. Industry projections published this year show sizeable growth in AI spending and semiconductor revenue, underpinning expectations that RAN vendors and operators will prioritise AI-ready hardware and software as part of multi‑year refresh cycles.

One prominent strand of the debate concerns the role of GPUs in RAN sites. NVIDIA’s AI Aerial architecture and public messaging position GPUs as a way to deliver high‑speed inference and new revenue-generating services at the edge, and the company has emphasised software-defined approaches that allow RAN functions and AI tasks to share accelerated infrastructure. Early operator pilots and benchmarking touted by partners suggest viable performance at carrier scale for selected workloads.

At the same time, chipmakers other than GPU specialists are actively framing alternatives. Intel has introduced new Xeon processors targeted at edge AI and early 6G infrastructure, arguing that tightly integrated CPUs with matrix and vRAN acceleration features can host AI inference alongside virtualised network functions without always needing discrete accelerators. Vendors and operators are thus weighing heterogeneous compute strategies rather than a one‑size‑fits‑all approach.

Practical constraints remain central to deployment choices. Network operators operate at vast scale under strict power and cost constraints, and comparative efficiency , performance per watt and total cost of ownership , is a decisive variable when considering adding power‑hungry accelerators at distributed cell sites. Recent vendor testing and industry analysis highlight measurable gains from more power‑efficient CPU designs, reinforcing conservative operator preferences for upgrading existing sites where possible.

These tensions influence how intelligence is distributed across radio, midhaul and cloud. Some demonstrations have shown that latency-tolerant services can centralise inference in data centres, while other use cases push processing closer to the radio for real‑time sensing, robotics or localised autonomy. Vendors are experimenting with neural accelerators embedded in radios, centralised GPU pools, and hybrid topologies to balance latency, energy and multi‑tenant economics.

The immediate picture is of an industry moving beyond rhetorical enthusiasm toward differentiated deployment strategies: software‑defined AI frameworks and accelerated platforms will play a growing role, but operators will pick architectures that reflect workload characteristics, site constraints and return on investment. If hyperscale demand for distributed intelligence materialises, the RAN could evolve into an extensible AI infrastructure; for now, the path to that future will be determined by a series of pragmatic trade‑offs between performance, efficiency and cost.

Source Reference Map

Inspired by headline at: ^[1]

Sources by paragraph:

Source: Noah Wire SerFuse Wire Servicesvices