3
Orange is preparing to test Nokia’s AI-RAN platform in its live network, putting claimed radio-efficiency gains to the commercial test while exploring the bigger proposition of shared compute for AI workloads.
Smaller start – Orange will start small with Nokia AI-RAN, testing performance, energy efficiency, operational impact and total cost of ownership before scaling.
Three models – It will consider all of Nokia’s deployment models – AirScale plug-ins, standalone AI-RAN nodes, and cloud-native COTS deployments – to provide flex and variety.
Bigger proposition – Orange sees potential in sensing, positioning, and other AI applications, but the economics of putting spare “fallow compute” to work remain unproven.
Orange is preparing to take Nokia’s AI-RAN platform out of the lab and into its live network, it said today at Intelligent RAN Forum. The firm is putting Nokia’s promise about gains in spectral efficiency and energy consumption to the test in a commercial set-up, while also exploring whether the platform will become an engine for parallel AI workloads, beyond just RAN optimisation. Orange said it expects to deploy all three of Nokia’s AI-RAN set-ups, likely starting with plug-ins for its existing AirScale base, expanding to new AI-RAN nodes, and a cloud-native deployment.
Nokia is experimenting on AI-RAN with other telcos besides, including: T-Mobile in the US; NTT Docomo and Softbank in Japan; Indosat Ooredoo Hutchison in Indonesia; du and e& in the UAE; Mobily, stc, and Zain in Saudi Arabia; A1 Group in Austria; Chunghwa Telecom in Taiwan; and TPG Telecom in Australia.
Speaking during a panel session at the RCR event (available on-demand), Laurent Leboucher, group chief technology officer and executive vice president for networks at Orange, said: “We are ready to test in real life [and] not only in our lab.” Orange will start small, he said, and assess the value of the technology before deciding whether to extend the deployment. “[We are] starting small at first, really to understand the value it can bring us. When we are convinced and when we are fully, let’s say, comfortable with all the operational aspects, we will decide to move forward and extend.”
Nokia, also on the panel, reckons it has achieved improvements of about 20 percent in spectral efficiency in test conditions. It is targeting 50 percent by mid-2027, with its full commercial launch, and more than 100 percent in 2028. Leboucher (pictured below, bottom) commented: “These gains are huge; times-two is very significant. The real test is whether they can be delivered consistently in a live network across different traffic loads and radio conditions – with a compelling total cost of ownership. That is what we intend to [find out] – to understand this broader picture.”
He added: “We’d also like to [see how] it can also help prepare for future traffic, typically for AI driven usage, which we expect to be more uplink-intensive [and] latency sensitive.” The AI-RAN concept splits roughly three ways: AI for RAN to optimise radio performance, AI on RAN for running AI workloads at the edge, and, nominally, AI and RAN as a shared compute fabric between the two. With $1 billion from GPU-linchpin Nvidia, Nokia is looking to embed accelerated compute as the latter into carrier 5G/6G infrastructure to prop-up RAN management and edge AI cases.
Speaking on the panel, Pallavi Mahajan, chief technology officer and AI officer of Nokia (pictured above, left), remarked: “We are making [the RAN] a programmable platform for both connectivity and compute.” Nokia has three deployment models for its AI-RAN platform – an AI-RAN capacity plug-in for existing RAN sites, a standalone AI-RAN node, and a cloud-native deployment using off-the-shelf (COTS) servers – to fit with operators’ “installed bases and transformation strategies”. All three share the same underlying software stack – rather than creating (new) separate technology silos.
The flex between is important, says Orange,. Leboucher said upgrading the existing AirScale base could be the most pragmatic approach in many locations, while dedicated AI-RAN nodes could address very high-capacity requirements, and cloud-based AI-RAN fits with Orange’s wider cloud-native evolution. “What is key is that those different models are consistent together,” said Leboucher. “They must coexist, without creating operational fragmentation. We need a common software foundation.”
But choice of architecture could determine how much capacity is available for AI workloads beyond the RAN itself. Mahajan said operators would need to consider traffic load, whether a deployment is rural or urban, how much AI compute is required for spectral-efficiency gains and how much additional capacity is available for other AI applications “Every physical AI case will dictate what you can run where,” she said. Some workloads will run at the cell site; others will run in a central office or in the cloud – depending on latency and compute requirements.
Better radio efficiency is most relevant in heavily loaded cell sites, where spectrum is constrained; it might, potentially, defer some network expansion for operators, whether by new sites or new spectrum. But the grand pitch for the sector is that accelerated compute in RAN infrastructure will host non-telco workloads also – as part of this so-called ‘AI grid’ concept, which ropes in cell sites as distributed inference points closer to the edge (and the user, ultimately). Nokia presents its AI-RAN solution as an open and programmable platform for both workloads.
“It has to be open, it has to be programmable, it has to be extensible, and of course it has to be secure,” said Mahajan. But the first applications will remain firmly within the telco domain. Leboucher pointed to spectrum and interference management and more automated radio configuration as near-term candidates, particularly applications that require direct access to RAN data and real-time local processing. He also sees a longer-term opportunity with sensing (ISAC), positioning, and monitoring – for infrastructure upkeep, industrial safety, drone detection, other apps.
“Those are the use cases where we see quite a clear business case, even if there is a lot to be worked out with Nokia and with others,” Leboucher said. The precise location of the compute will be more important as the apps emerge. Not every workload needs to run at the cell site, clearly; AI-RAN architecture will depend on latency requirements, traffic loads, and spare AI compute. Nokia’s different AI-RAN deployment models will serve-up different amounts of “fallow compute”, said Mahajan – to be used for additional AI workloads.
As such, the investment case (still somewhat stretched-out) is not just about whether AI improves the RAN; the likes of Orange, and the rest of Nokia’s AI-RAN operator club, will presumably be making initial calculations about where to place additional compute – and for what, and for what kind of returns. Nokia is mindful of the maths, speculative as it is. Mahajan said AI-RAN deployments should deliver the right total cost of ownership for operators while also enabling “newer monetization models” beyond connectivity – ultimately to generate new sources of revenue.
But operational hurdles remain. Orange wants to introduce probabilistic AI into a deterministic mission-critical environment. Faster software innovation cannot come at the expense of network stability, said Leboucher. Every new capability needs to match requirements for performance, resilience, security, predictability. The ability to introduce functions selectively, and measure their impact and roll them back, will be critical. Mahajan said the discipline is to move from ‘black box’ to “glass box” principles – for traceability and explainability, and the ability to revert if AI models under-perform, mis-behave, or fail.
Which makes AI-RAN as much about changing the operating-model, as it is about changing tech. “It’s not just technology here,” said Leboucher. “We are talking about changing the way we work.” Operators must change how they supervise autonomous infrastructure and manage safeguards, he said. The live-network testing will provide the first practical evidence of whether that model works. For now, Orange is not committing to one AI-RAN architecture or a particular commercial rollout. It is starting with a more urgent question: does the tech deliver, and not mess up?
If it can, and it makes their network operations more efficient, then the bigger question is what operators do with all the compute when the RAN itself becomes just another (ultra-reliable) workload.
