In this interview with Mobile Europe, Orange’s Laurent Leboucher, explains why he is watching AI RAN’s progress with interest, and in particular, the business case for it
AI was a dominant theme at Mobile World Congress 2025 and one of the strands was the emerging concept of AI RAN. Vendors and operators are exploring how it could enhance network performance and potentially open new sources of revenue. But not everyone is ready to jump on the AI RAN bandwagon just yet, as the RAN market drops to its lowest level in two decades.
Laurent Leboucher, CTO and Executive VP for Networks at Orange Group (pictured above), said AI RAN is “still a bit futuristic”. But he is watching with interest and talking to other operators about what they are doing, notably Softbank, which has been an early proponent of the technology.
“I would like to understand the business case behind [it] because putting GPUs on every radio site is quite expensive. Is there really a business case for that? I don’t think so. Maybe in the future, but not now,” he said.
Far edge compute nodes
The AI RAN idea, which is led in large part by NVIDIA, encompasses leveraging AI to make radio network operations more efficient as well as a more radical change that could see RAN and AI workloads hosted on the same GPU-based platform. This essentially turns base stations into far edge compute nodes.
In the latter scenario, traditional RAN vendors like Ericsson and Nokia would need to overhaul their virtual RAN products so that their software can run on GPUs rather than CPU-based platforms.
“It’s quite disruptive … That sounds interesting on paper, and also the idea that we could leverage the same hardware to do at the same time the RAN [workloads] and the B2B [AI offerings]. Honestly, I’m not sure that we have today a clear business case … it’s not something that we would implement in the network right now,” he said.
Cost is crucial
But Leboucher, who has just been appointed Chair of the Next Generation Mobile Networks (NGMN) Alliance, has not written the idea off as a non-starter for Orange. He said he was “extremely interested” in it, but his main concern is that it appears to be too expensive to be viable at this point.
“I’m open to looking at it… but there will be a big question on the price points. Are we able to do it at the right price point to introduce it in the network? I don’t think it is the case today,” he said.
Orange is not alone in its wait-and-see approach to AI RAN as operators have not exactly flocked to the cause. The AI RAN Alliance was launched at MWC in 2024 by founding members that include Arm, Ericsson, Microsoft, Nokia, NVIDIA, Samsung, Softbank, and T-Mobile US. The group has grown to 75 members in its first year, but it has attracted just seven operators, none of which are European, and used MWC as an opportunity to invite more organisations to join its endeavours.
AI for network
While it is early days for AI RAN, Orange has taken a two-pronged approach to AI that focuses on how the technology can make networks more efficient (AI for network) and how networks need to adapt to deliver AI-based services to customers (network for AI).
“AI for network and network for AI are two sides of the same coin… If we start to transform our network to make it a platform for AI workloads, we will need a lot of very smart capabilities to automate the network and to address the requirements for the new AI workloads,” explained Leboucher.
AI for the network use cases are more mature as Orange has been working on these for many years. Examples include streamlining network planning processes, reducing the time it takes to do root cause analysis for network issues, minimising energy consumption at mobile sites, and reducing the number of trips field engineers need to make to inspect sites.
The operator is now focused on scaling these use cases across its OpCos in Europe.
“Today, our programme is more like a collection of different bottom-up use cases. We try to expand and replicate each one in the different countries, but it is not easy. Part of the challenge is that the foundations are not completely homogeneous, especially in Europe,” he said.
To address this fragmentation, the operator is working to “harmonise” tools and “OSS landscape”, while leveraging its shared network operations centre (NOC ) in Europe to build a scalable platform, he explained.
Network for AI
On the flipside of the strategy, Orange is addressing the impacts that new AI applications will have on the network. The first of which is that the uplink will become “very important” because AI workloads are “multimodal” with text, voice and video.
“That means latency will matter and you will need to have stable connectivity”, according to Leboucher.
Another AI-induced network impact is deciding from what point to deliver the “inference” – that is, when trained models make decisions to deliver the AI app. “We need to do the inference somewhere. The first place where it makes sense is on the device,” he said.
But between the device and the cloud, Orange needs to be able to deliver AI-based B2B services to enterprises that meet their data protection and sovereignty requirements. For these applications, Leboucher said public cloud is not an option for many businesses and they need a private or on-premises alternative.
“The easy case is on prem. What is less easy is edge in the network. For edge in the network, more for sovereignty reasons than latency, we consider that we could use private cloud as another way to do inference … compared to doing it in the public cloud,” he said.
