Background image: 123rf

The alliance is expanding its focus on AI RAN services while tackling workload placement, infrastructure sharing, and the commercial case for turning mobile networks into distributed AI platforms.

In sum – what to know:

Implementation push – The AI-RAN Alliance wants to move beyond individual demonstrations towards tested deployment patterns, measurable performance, and clearer routes to commercial adoption.

Four new priorities – Its work spans edge inference, physical AI, semantic communications, and multimodal RF sensing and ISAC, with security and AI workload placement emerging as strategic concerns.

Monetization challenge – The Alliance is exploring edge compute, inference and sensing as services, alongside latency-guaranteed connectivity – to give operators revenue beyond conventional network access.

The AI-RAN Alliance is moving beyond tech demos to develop implementation blueprints, test frameworks, and commercial models for AI in mobile networks – as the industry looks to turn AI-RAN research into deployable systems and revenue. “We are not a standards development organization,” said Alex Jinsung Choi, principal fellow in Softbank’s technology unit and CTO office, and chair of the AI-RAN Alliance. His point, at Intelligent RAN Forum last month (available on demand), was to remind everyone that the telecoms industry is turning AI-RAN theory into action.

The alliance’s latest work focuses on how operators might turn distributed radio infrastructure into a platform for AI services. Its expanded scope – part of its “working group three”, focused on AI on RAN – now covers four service-led themes: edge inference, physical AI, semantic comms, and multimodal spectrum sensing. It wants to make the commercial requirements a part of the technical discussion, said Choi. “Group three is organized around services, not only technologies. That helps us ask the right commercial questions.” 

Those questions are about the latency, privacy, and computing required for the service being delivered. “The main shift is toward implementation,” said Choi. 

From services to implementation

The alliance has more than 142 member organizations across more than 17 countries, he said. Its work is divided into three areas, as commonly understood: AI for RAN, AI and RAN, and AI on RAN. “Each major work area should aim [at] practical evidence, clear architecture, and a realistic path to commercial deployment.” Choi outlined their definitions and remits, as commonly understood.

Choi – three working groups, four task groups

AI for RAN covers the application of AI to improve RAN performance, including channel estimation, interference mitigation, mobility management, scheduling and link adaptation. AI and RAN focuses on sharing infrastructure between RAN workloads and AI computing, including dynamic allocation of compute resources. AI on RAN uses networks and edge infrastructure to deliver AI services. The disciplines are connected by two practical questions, he said: how to secure AI workloads alongside critical network functions, and where AI computing should take place.

Third-party AI applications must be isolated from critical RAN functions, he said; decisions about where to run AI workloads need to account for latency, model size, data locality, privacy, transport costs and available computing resources. These questions become more pressing as operators consider running AI applications at the network edge, he said – where the attraction is proximity to users and devices but the economics depend on demand, computing capacity, and the ability to deliver services reliably.

So: why run AI at the edge? And: what does the operator bring? The group’s expanded AI-on-RAN portfolio covers applications that could use localized computing – including video, glasses, visual positioning, autonomous driving, robots, and assistive systems. Other areas include semantic communications, which considers meaning and context in the transmission of information, and RF sensing for applications such as intrusion detection and real-time environmental awareness. The logic, as always, is that mobile infrastructure could do more than provide connectivity. 

“The RAN is no longer only a connectivity platform,” said Choi. “It can become [an] edge AI inference platform… for robots, vehicles, drones, and physical AI.” But the availability of compute at the network edge does not automatically make it commercially useful. The central question for telco-based edge inference is to determine “when it provides real advantages to have the compute close to the user”. The alliance identifies four potential service building blocks: latency-guaranteed connectivity, edge compute as a service, inference as a service, and sensing as a service.

What can operators sell?

Operators could offer connectivity with defined latency commitments to a particular edge site, rent out nearby computing capacity, host AI models as managed inference endpoints, or use integrated sensing and communications (ISAC) to provide environmental awareness. ISAC is a big topic for the AI RAN Alliance, evidently. Choi also pointed to potential applications across automotive, manufacturing, defense, healthcare, and logistics. “The opportunity is to combine connectivity, local computing, AI inferencing, and sensing around [a] clear business problem,” he said.

That last qualification is important. The alliance’s commercialisation task group is developing business blueprints, market assessments and potential paths to edge-inference revenue. Its stated purpose is to connect technical progress with customer value and a credible route to adoption – rather than just to assume new network capabilities will automatically generate demand. It is a mistake this industry has made before, and one SK Telecom alluded to in a separate presentation. 

As above, the alliance has identified two strategic priorities: RAN security and AI placement. The security question arises as operators consider hosting third-party AI models, containers, and data pipelines on infrastructure that also supports critical RAN functions. This requires “a very strong isolation and clear operational boundaries”, said Choi. He added: “AI innovation cannot come at the expense of RAN reliability.” The second issue is where workloads should run and what computing resources they actually need. 

“Not every AI model should run at the cell site, and not every workload needs GPUs,” he said. Instead, placement should depend on latency, model size, data locality, privacy, transport cost, available compute, and service requirements. As such, the alliance’s work on shared infrastructure is being coordinated with its edge-services work – with the application helping to define demand and the infrastructure determining the most appropriate place to serve it. This is a practical constraint on the idea of turning the RAN into a distributed inference platform. 

From blueprints to validation

The alliance also has four cross-cutting task groups – around data for AI, test and methodologies, agentic AI, and commercialization. Their remits range from structured data collection, synthetic data and privacy protection to common testing conditions, performance indicators and lab certification. The agentic-AI group examines how systems can act on operator intent and network conditions, while the commercialization group focuses on business models and market opportunities.

Choi described the agentic-AI process as a control loop in which the “operator sets intent such as specific coverage or lower power budget, and the system observes conditions, selects an action, checks the result, [and] improves future decisions”. The objective is to help manage complex networks while retaining human oversight. Data quality and repeatable testing are also intended to make claims about AI performance more credible. New approaches should be compared with a clear baseline under consistent conditions to assess improvements.

The AI-RAN Alliance also has four endorsed labs providing vendor-neutral environments for integration, benchmarking and validation. Their capabilities include synthetic data generation using digital twins and ray tracing, testing model portability across CPUs, GPUs and NPUs, and collaborative research and development. “A model that only works in one environment has a limited value,” said Choi, explaining the importance of portability testing. The labs are intended to help operators, vendors and research teams assess solutions before wider deployment.

“The key change is that AI is moving from simple prediction problems to making decisions and taking actions,” Choi said. The alliance presented 33 demos at MWC in Barcelona in 2026 – across all three working groups, for all manner of use cases. It also presented digital-twin environments for AI agents to test possible actions and effects before making changes in live networks. The demonstrations are evidence of the industry’s progress with AI RAN across radio optimisation, shared infrastructure, edge computing, and network management. 

The next challenge, he said, is to bring these capabilities together into platforms that can be tested against measurable performance requirements and deployed commercially. Its stated objective for 2026/27 is to turn the RAN into what Choi calls “a distributed fabric for AI inference” – with computing placed where it creates the most value.