AI-RAN for 6G: The Architecture Question Is Where AI Runs

AI-RAN for 6G is moving from concept demonstrations toward commercial trials, but the most important design decision is still unresolved: where should AI run? Nokia, Nvidia, and T-Mobile have shown that a 5G radio and an Nvidia server can handle connectivity and AI workloads such as video streaming and captioning together. The Seattle demonstration was controlled, using one radio at one site, so it proves feasibility rather than large-scale economics. The trial and the wider AI-RAN debate matter to AI builders because network placement will shape latency, energy use, operating cost, and how much control applications have over real-time decisions.

AI-RAN is a placement problem, not a GPU-at-every-tower strategy

AI-RAN refers broadly to integrating AI workloads into or alongside the radio access network. Nokia and Nvidia are pursuing AI-capable RAN hardware, while Ericsson has introduced software that adds AI to existing radios and base stations without requiring new hardware. Huawei has also announced tools for automated network diagnosis and repair. These approaches are related, but they do not represent one agreed architecture. The companies are taking different paths into AI-enabled RAN.

The distinction between AI hardware and AI algorithms is important. Some compact models can make network decisions in microseconds or milliseconds and may run on modest local chips. Other workloads can tolerate more delay and run at an edge site or in a regional data center. That makes a hierarchical design more credible than placing a large foundation model at every cell tower.

What the hierarchy could look like

The likely pattern is heterogeneous. Small models inside radios and basebands could handle strict real-time control. More capable models at edge sites could process local conditions with less latency than a distant cloud. Regional or central systems could use larger models for planning, reasoning, and coordination. This proposed division of responsibility follows the latency requirements of each network function.

For builders, this resembles a distributed inference system with different model sizes and authority levels. A local model might adjust a radio parameter, while a higher-level service allocates capacity across sites. The interface between those layers may matter more than any individual accelerator: teams will need clear policies for fallback behavior, observability, and which system can override another.

The business case is still unproven

AI-RAN trials show that AI compute can share space and power with RAN workloads. They have not yet established that deeply integrating AI into the RAN produces better economics or capabilities than upgrading existing networks with smarter software. Operators also face the possibility that additional hardware increases energy consumption and capital costs.

Coordination is another operational risk. Separate systems optimizing speed, coverage, and energy can make locally sensible decisions that conflict globally. AI-RAN therefore needs evaluation across sustained network conditions, not only a successful demonstration. The useful benchmark for operators is the improvement over a conventional RAN with software optimization, measured alongside energy, reliability, latency, and human review requirements.

What builders should watch next

3GPP discussions in Madrid are expected to influence how 5G migrates toward 6G, with broader 6G rollout discussed around 2029. Commercial trials are expected before then, but standards and deployment plans may change as the industry tests the architecture. The 6G timeline and standards debate remain part of an evolving process.

The practical decision rule is straightforward: keep hard real-time control close to the radio, move flexible workloads to shared edge infrastructure, and reserve large models for coordination tasks that justify their cost. Until wider trials prove otherwise, AI-RAN should be treated as an architecture to validate, not a reason to assume every tower needs a data center.