From Standard to Simulator: How GSMA's Open Telco AI Challenge Traces the Long Road to Autonomous RAN

From Standard to Simulator: How GSMA’s Open Telco AI Challenge Traces the Long Road to Autonomous RAN

On 10 November 2026, in Doha, the GSMA’s Open Telco AI initiative announced two competitions — TelcoRetrieve and the AI Agents for RAN Challenge — developed with ETSI, IEEE, ITU and Huawei, carrying a combined prize pool of EUR 55,000. On its face, this is a modest piece of industry housekeeping: a pair of contests inviting operators, researchers, developers and startups to build AI systems that answer telecoms questions accurately and solve radio engineering problems inside Giulia, an open-source radio network simulator from Universitat Politècnica de València. A webinar is scheduled for 18 November; judging will weigh accuracy, efficiency, reasoning capability and security.

But read against the last decade of RAN history, the announcement is more interesting than a prize draw. It is the moment the industry’s most influential trade body formally concedes that the hard part of AI-native networks is not the radio, and not even the compute — it is the reasoning layer that sits above both, and the trustworthiness of the data and standards that layer reasons over. That is a very different problem from the one the industry set out to solve when Open RAN first entered the vocabulary.

The road from interfaces to agents

The original Open RAN promise was architectural. Disaggregate the base station, standardize the interfaces between radio unit, distributed unit and centralized unit, open the management layer, and let a market of interoperable components emerge. The intellectual bet was that openness at the interface level would, over time, produce competition, lower costs and faster innovation. For years the industry measured progress in interoperability events, multi-vendor proofs and fronthaul compliance.

What that framing underweighted was software. Once the RAN became disaggregated and cloud-native, the interesting questions migrated upward — to the RAN Intelligent Controller, to xApps and rApps, to the orchestration and automation layer that would eventually be called the SMO. The industry’s vocabulary shifted with it: from “open interfaces” to “open platforms,” from “virtualized” to “cloud-native,” and finally, in the last two years, to “AI-native.”

That migration is visible in the news flow of just the past week. Verizon is deploying the Ericsson Intelligent Controller nationwide to run AI-native rApps — including a Cell Anomaly Detector, a Root Cause Explainer and an Anomaly General Optimizer — after using an early deployment to automate fixed wireless access inter-RAT handover. AT&T’s network CTO Yigal Elbaz described moving “pretty much in every layer” of the network to open, disaggregated, programmable architecture, shifting cell-site management to its SMO platform as the foundation for what he called “the agentic era,” with an internal goal of Level 4 autonomy by 2030. Vodafone Egypt trialled Ericsson’s AI-native scheduler for link adaptation, reporting up to a 14% gain in radio spectrum efficiency. Ethio Telecom, working with ZTE, reported an 11% increase in 4G capacity and a 40% gain in peak-hour user throughput from AI-driven beam scheduling on a pre-commercial AIR RAN deployment.

Each of these is a data point in the same trajectory: intelligence moving from the lab into live networks, and from optimization of individual parameters toward something closer to autonomous decision-making. And each of them, implicitly, depends on a capability the GSMA challenge is now trying to formalize — an AI system that can reason about a radio network, use tools, test interventions and understand the consequences.

Why a simulator, and why now

The AI Agents for RAN Challenge asks participants to build agents that investigate network conditions, test possible interventions and account for how changes propagate across the network. Tasks span fault diagnosis, outage impact analysis, coverage and radio configuration, site placement and energy management. Crucially, the environment is Giulia, an open-source simulator, not a live commercial network.

That choice is the most revealing detail in the announcement. For years, the industry’s AI-RAN narrative has been anchored in demonstrations on live networks — T-Mobile’s first AI-RAN call on a commercial network, field trials with Nokia’s CUDA-accelerated anyRAN software and NVIDIA’s Aerial RAN Computer, and nearly 20 operators publicly announcing AI-RAN collaborations. Those demonstrations matter, but they are poor environments for building and evaluating autonomous agents. Live networks are expensive to perturb, hard to reproduce and impossible to benchmark fairly across vendors.

A simulator solves the evaluation problem. It gives the industry a shared, reproducible substrate on which to compare how well different agents diagnose faults, allocate spectrum or plan sites — and, just as importantly, to test whether an agent’s reasoning holds up when the network pushes back. The judging criteria make the intent explicit: accuracy, efficiency, reasoning capability and security. This is not a contest to find the best optimizer. It is a contest to find out whether agents can be trusted to operate infrastructure.

That framing connects directly to the second competition. TelcoRetrieve targets the accuracy of AI systems working with technical standards — the problem of grounding model outputs in authoritative telecoms specifications rather than plausible-sounding invention. Anyone who has watched a large language model confidently misstate a 3GPP clause understands why this matters. If operators are going to delegate network engineering tasks to agents, the agents must be able to retrieve and apply standards correctly, and to show their work.

The economics underneath the experiment

The timing is not accidental. The commercial case for AI in the RAN has been steadily sharpening, and it now rests on two distinct propositions that the industry has learned to separate: AI for RAN, and AI and RAN.

The first — using AI to optimize radio performance — is producing measurable, if modest, gains. A 14% spectral efficiency improvement in a high-interference trial, an 11% capacity lift in a pre-commercial deployment: these are real numbers, but they are incremental, and they must be weighed against computing costs, power consumption and the operational complexity of running AI workloads alongside radio processing. As one recent analysis of the semiconductor landscape noted, a successful AI-RAN demonstration does not establish that large-scale deployments will deliver lower total cost of ownership.

The second proposition — hosting AI workloads on RAN infrastructure — is where the larger strategic bet lies. An Omdia AI RAN Operator Survey published this week found that 77% of respondents expect AI RAN and end-user AI services to co-evolve, with 31% anticipating a tightly integrated model in which RAN functions and AI workloads run on mostly unified hardware and software. The logic is straightforward: a common platform eliminates proprietary hardware silos, reduces operational complexity and enables low-latency edge AI services that a siloed architecture cannot support.

Both propositions depend on automation that goes well beyond parameter tuning. They depend on agents that can reason across domains — radio, transport, core, energy — and on standards knowledge that can be retrieved reliably. That is precisely the capability the GSMA challenge is designed to cultivate.

The standards question hiding in plain sight

There is a deeper historical resonance here. The telecom industry has spent three decades building standards bodies whose output is, by design, precise, exhaustive and often impenetrable. 3GPP specifications run to thousands of pages. ETSI, ITU and IEEE each maintain their own bodies of work. The value of that precision is interoperability; the cost is that only trained engineers can navigate it.

AI agents change the calculus. If an agent can reliably retrieve and apply standards, the industry gains a way to make its accumulated technical knowledge machine-actionable — a genuine unlock for automation. If it cannot, the industry risks deploying systems that sound authoritative while being subtly wrong, which in a radio network is worse than no automation at all.

That is why the involvement of ETSI, IEEE and ITU alongside the GSMA and Huawei matters more than the prize money. The challenge is not just a competition; it is an attempt to build a shared evaluation culture around telecom AI — one in which claims about agent capability can be tested against common tasks, common simulators and common standards. The industry has learned, through years of AI-RAN hype cycles, that demonstrations are cheap and benchmarks are hard.

The gap between a simulator and a network

None of this means the path to autonomy is short. The gap between an agent that performs well in Giulia and an agent trusted to reconfigure a live cell site is wide, and the industry’s own roadmaps acknowledge it. AT&T’s Level 4 target is set for 2030, not next year. Verizon’s rApp deployments are being staged carefully, with self-developed rApps tested alongside vendor ones. Ericsson’s AI-native scheduler was trialled, not deployed at scale.

The historical pattern here is familiar. Open RAN itself followed a decade-long arc from concept to commercial reality, and its proponents repeatedly had to recalibrate expectations as the gap between architectural promise and operational economics became clear. AI-native RAN is on a similar arc, and the GSMA challenge is best understood as an attempt to shorten the painful middle of that arc — the years in which the industry argues about what “autonomy” means and how to measure it.

There is also a competitive dimension. The AI-RAN landscape is crowded with vendor platforms, GPU-accelerated compute stacks and operator-specific strategies. A neutral, open simulator and a shared set of agent tasks give the ecosystem something it currently lacks: a vendor-agnostic way to compare approaches. That is the same logic that made Open RAN’s interoperability events valuable in the early years, applied to a new layer of the stack.

What to watch

The 18 November webinar will introduce the challenge mechanics and explain how entries will be evaluated. The results will be worth watching not for who wins the EUR 30,000, but for what the winning agents reveal about the state of the art: whether reasoning agents can outperform rule-based automation on fault diagnosis, whether retrieval-augmented systems can be made reliable enough for standards work, and whether security considerations can be baked in from the start rather than bolted on.

Those answers will shape the next phase of the AI-RAN conversation. The industry has spent several years proving that AI can improve the RAN. The harder question — the one the GSMA is now putting to a broader community — is whether AI can be trusted to run it. The answer will not come from a single demonstration on a live network. It will come from the unglamorous work of simulation, benchmarking and standards grounding that this challenge, in its small way, begins.

Sources

Related Posts

Vodafone's Málaga Lab and the Silicon Fork Behind AI-RAN

Vodafone opens a multi-vendor AI-RAN hub in Málaga while VodafoneThree trials Nokia and NVIDIA — and Orange weighs GPU versus ASIC.

VodafoneThree's AI-RAN Lab Goes Live: Inside the Málaga Hub Where AI-for-RAN, AI-and-RAN, and AI-on-RAN Get Measured

VodafoneThree opens a Málaga AI-RAN hub with Nokia and NVIDIA, testing AI-for-RAN, AI-and-RAN and AI-on-RAN on shared compute ahead of 2027 field trials.