Research Radar/5G & 6G/France · United Arab Emirates
Big AI models are too slow to run a radio network
A survey of AI agents for 5G and 6G finds a clock gap. Fast network control loops close in 10 milliseconds to 1 second, while a large model needs 200 milliseconds to several seconds to answer.
A tutorial and survey from EURECOM and partners maps AI agents onto 5G and 6G networks and finds a timing gap. Near real time control loops run between 10 milliseconds and 1 second, while a 30 to 70 billion setting model on a high end GPU takes 200 milliseconds to several seconds per query. The authors argue models should plan while lightweight rules react.
What happened
A cell tower gets busy at half past five, and signal quality drops. Therefore, somewhere a system has to shift capacity around, and it has to do it before your call breaks up. That decision happens in a fraction of a second.
Phone companies want AI agents to make those decisions. An agent is a model that can plan, use tools and act on its own. The dream is a network that fixes itself while the engineers sleep.
A new survey looks at that dream and finds a clock problem. The fast control loop in a modern radio network runs between 10 milliseconds and 1 second. A large model on a top graphics card needs 200 milliseconds to several seconds to answer one question. Therefore, the model is often slower than the loop it is meant to steer.
The test
There was no test, because this is a tutorial and survey. It is called “LLM-Powered Agentic AI for 5G/6G Networks: A Tutorial and Survey on Architectures, Protocols, and Standardization”. It was written by Mazene Ameur, Abdelkader Mekrache, Bouziane Brik and Adlen Ksentini. They work at EURECOM in France, Simula Metropolitan in Norway and the University of Sharjah in the United Arab Emirates. It went up on arXiv on 17 July 2026.
The paper does two jobs. First it teaches. It explains how a modern network is built, and how AI agents plan and use tools. Then it surveys. It maps agent ideas onto the parts of a 5G and 6G network, reviews the standards bodies, and lists six open problems.
It also covers the new ways agents talk. The paper notes that Anthropic introduced the Model Context Protocol in November 2024, and that Google announced Agent-to-Agent communication in April 2025. Both are now being pointed at telecom systems.
The result
The main finding is the clock gap above. Put the two numbers side by side. A loop that must close in 10 milliseconds cannot wait 200 milliseconds for an answer. And 200 milliseconds is the good case, for a model of 30 to 70 billion settings running one query at a time on a high end card.
Memory is the second wall. The survey says models in the 7 to 70 billion range need 14 to 140 gigabytes of graphics memory just to run. Squeezing them can cut that by four to eight times. However, the authors add a warning, because the accuracy cost of that squeezing has not been measured for telecom work.
The third problem is the scariest, because the survey cites a study where 82.4 percent of tested models were broken through messages from another agent. The same models fell to 52.9 percent through poisoned reference documents, and 41.2 percent through a direct attempt by a person. In other words, an agent trusts another agent far more than it trusts you.
The authors suggest a fix in shape rather than speed. Split the job in two. Let the model plan over seconds and minutes. Let small, plain rules react in milliseconds. The model sets the goal. The rules do the twitching.
What it means
THE MODEL SHOULD PLAN AND THE RULES SHOULD REACT. That is the line worth keeping, and it reaches far past phone networks.
Think about any control job you know. A greenhouse. A delivery fleet. A trading desk. A factory line. In each one there is a slow question and a fast question. What should we aim for this hour? And what do I do in the next tenth of a second? A large model is good at the first and bad at the second.
The security point matters just as much for small teams. If you run two agents that pass work to each other, the second one may act on anything the first one says. Most safety training was built to stop a person tricking a model. However, it was not built to stop a model tricking a model.
Business ideas from this paper
- A timing map that shows a client which of their decisions can wait for a model and which cannot. Who buys it: teams adding AI to control systems in factories, logistics or energy. A price to test: 1,200 dollars for a one page map of ten decision points. A one-week test: do it free for two firms, publish one map with the names removed, and count how many firms ask for theirs.
- A message checker that sits between two agents and blocks instructions passed from one to the other. Who buys it: small teams running two or more agents that hand work to each other. A price to test: 40 dollars a month per project. A one-week test: release it free for one agent framework, then count how many teams turn it on and leave it on.
- A plain language test set for AI network assistants, built on free network software. Who buys it: vendors selling AI tools to phone companies, who currently have nothing to prove their claims with. A price to test: 2,500 dollars for a scored report. A one-week test: score two open assistants, publish the scores, and count how many vendors ask to be scored next.
How sure can you be?
Take the numbers as pointers, not as measurements, because this team ran no experiment. Every figure comes from work by other people.
The latency figures are typical ranges, not one clean test. The survey says the answer time depends on model size, batch size, decoding method and hardware. So 200 milliseconds is a rough floor for one class of setup, and not a fixed law.
The 82.4 percent attack figure comes from a study the survey cites. The survey does not say how many models were tested or what tasks they were doing. Treat it as a warning flag, not a settled rate.
The authors name the deepest gap themselves. There is no telecom test set for agents. Without one, nobody can compare two products or check any vendor claim. Until that exists, every number in this field rests on one team testing one thing.
What would settle it. A shared, open test bed on free network software, with several teams running the same jobs and reporting the same units.
Do this today
Write down the fastest decision your system has to make, in milliseconds. Then time your model on that same decision. If the model is slower, it belongs in the plan and not in the loop.
Source: LLM-Powered Agentic AI for 5G/6G Networks: A Tutorial and Survey on Architectures, Protocols, and Standardization, July 2026. arXiv:2607.16066 · arxiv.org (preprint · not yet peer reviewed).
Just Out Tech explains new research in plain language. This article was drafted with AI assistance and checked by a human against the original source.
- The survey reports that fast radio control loops run between 10 milliseconds and 1 second, while a 30 to 70 billion setting model on a high end GPU takes 200 milliseconds to several seconds per query.
- Models in the 7 to 70 billion range need 14 to 140 gigabytes of GPU memory to run, and compression cuts that by four to eight times at an accuracy cost the survey says is unmeasured for telecom work.
- The survey cites a study in which 82.4 percent of tested models were compromised through messages from another agent, against 52.9 percent by poisoned documents and 41.2 percent by direct injection.
It is a two-part tutorial-and-survey with no experiment of its own. It states that agentic AI inside Near-RT and Non-RT RIC components at production scale has not yet been demonstrated.
- When it reaches you
- Conformance and certification for AI-native network functions have to exist first. Our estimate is not before the 2030s, because the paper reports no O-RAN conformance procedure for LLM xApps and no demonstrated accuracy at the sub-10 ms E2 deadline.
- Who is building on it
- No author holds a company affiliation. The authors are at EURECOM France, Simula Metropolitan Norway and the University of Sharjah UAE. No code, data or model release is stated.
- Who paid for the research
- The European Union Horizon Programme paid, through the 6G-INTENSE project (Grant No. 101139266) and the FLECON-6G project (Grant No. 101192462). The survey reviews both of its funders as case studies.
Questions people ask
what is an agentic AI network?
It is a network run by AI agents that can plan, call tools and act without a person in the loop. The survey maps this idea onto the control, management and AI parts of 5G and 6G. It is a research direction, not a deployed product.
why can a model not run a radio control loop?
Speed. The fast loop must close in 10 milliseconds to 1 second. A large model on a top graphics card needs 200 milliseconds to several seconds to answer, plus the time to build the prompt and call tools. The survey proposes splitting slow planning from fast reaction.
did the authors test anything?
No. It is a tutorial and survey. The team explains the background, maps the field, reviews standards work at 3GPP, ETSI and the TM Forum, and lists six open problems. Every number it quotes comes from other studies.
why are agents more dangerous in pairs?
Because safety training was built to stop a person tricking a model, not a model tricking a model. The survey cites a study where 82.4 percent of models executed a harmful payload when it arrived from a peer agent, well above the rate for a direct attempt.
Mark Alex
Mark Alex is the founder and Managing Director of Real Biz Digital, a technology company operating out of Nairobi since 2018. He works in agentic AI and the Model Context Protocol, AI governance, enterprise software architecture and cybersecurity. He holds an MSc in Mechatronical Engineering from Obuda University in Budapest and a BSc in IT, Forensic Technology and Cybercrime, from USIU-Africa in Nairobi, and has published IEEE conference research on an AI-powered digital twin for greenhouse systems. He is the author of seven books. Between 2020 and 2024 he mentored more than 200 university students and interns in Nairobi. He writes every Just Out Tech article from the original research paper.