Orange wants Nokia’s AI RAN gains proven in a live network
In sum – what we know:
- A software path to 6G – Nokia said an operator deploying its AI RAN today could reach 6G through a software upgrade alone, without a hardware refresh.
- Three ways in – The vendor is offering an AirScale capacity card, cloud AI-RAN on off-the-shelf servers, and a standalone accelerated node, all running the same anyRAN software.
- Start small in production – Orange said it will validate performance, energy consumption and TCO on a limited basis first, and won’t commit to one architecture across the whole network.
Nokia brought the spectral efficiency headline to the Intelligent RAN Forum — 20% demonstrated, 50% by mid-2027, more than 100% by 2028 — but a panel at the forum spent most of its time on everything that has to work underneath those numbers. Pallavi Mahajan, Chief Technology and AI Officer at Nokia, and Laurent Leboucher, EVP Networks and Group CTO at Orange, sat down with RCR’s James Blackman to unpack what AI RAN actually looks like when an operator has to deploy it.
The session was billed around that 2x promise. Nokia argues the RAN has become an open, programmable platform, with hardware-software separation following the same path as cloud computing and servers. Orange accepts the premise but reframes the challenge. For Leboucher, the question isn’t quite whether AI models can improve spectral efficiency, but rather whether the gains hold consistently in a live network, across different traffic loads and radio conditions, at a total cost of ownership that makes sense.
“Capacity, customer experience and economics are part of the same value chain,” Leboucher said. “The real test is whether these gains can be delivered consistently in a live network…”
RAN stops being a fixed stack
Mahajan’s argument starts with architecture. Nokia has decoupled its anyRAN software from proprietary hardware and paired it with Nvidia’s accelerated compute platform, running AI and machine-learning algorithms directly in the baseband to continuously optimize how radio resources are used. The biggest gains come in dense, heavily loaded cells where spectrum is the constraint.
The broader pitch is that innovation no longer rides on generational shifts. Every cycle from 2G through 5G meant custom silicon and long hardware timelines. Under a subscription model with cumulative software releases, operators get spectral efficiency gains, optimization capabilities, security upgrades and AI-native functions without a hardware refresh. Mahajan’s claim was that an operator deploying Nokia’s AI RAN today could reach 6G through a software upgrade alone.
Orange runs deterministic, five-nines infrastructure that Nokia supplies, and now they are putting probabilistic AI models on top of it. But how fast does that actually happen?
Leboucher was measured. “It’s a path,” he said. “It’s a learning journey.” Operators need to observe user experience, get the right data, roll back releases that don’t work and introduce capability selectively where it matters. Faster innovation “cannot come at the expense of the network stability,” he said. Every new capability has to meet demanding requirements on performance, resilience, security and operational predictability.
Mahajan acknowledged the weight of that. He said arriving at Nokia taught him that if a new feature costs even 0.001% of performance, “nobody gets to leave their seats.” The RAN as a platform, he said, is built on glass-box principles — traceability, explainability and rollback to conventional algorithms if a model misbehaves.
Three ways in, one software roadmap
Nokia offers three deployment models on one software stack. The first is a new capacity card that plugs into the existing AirScale chassis — a brownfield upgrade that protects cabling and rack investments. The second is cloud AI-RAN on commercial off-the-shelf servers for operators building cloud-native architectures. The third is a standalone accelerated AI-RAN node for high capacity and flexible deployment, which Nokia said it will detail by end of year. All three run the same anyRAN software and support 4G, 5G and a path to 6G.
Leboucher said the most important thing is not picking one architecture for the entire network. Orange doesn’t have enough experience with the new options yet, Leboucher said, but he believes upgrading existing AirScale infrastructure may be the most pragmatic option in many locations, with dedicated AI-RAN nodes for very high capacity needs and cloud AI-RAN consistent with Orange’s broader cloud-native evolution. The models have to coexist without operational fragmentation, on a common software foundation with consistent lifecycle management.
Orange is ready to test in its live network, not just in the lab. Leboucher said he has agreed with Mahajan to start small, validate performance, energy consumption and TCO, and then extend once Orange is comfortable with the operational aspects.
Placement is the piece nobody has settled. Mahajan used the term “fallow compute” for the extra AI headroom in the network, and how much you have depends on which deployment model you chose. Every physical-AI use case dictates what can run on a cell tower, what needs the central office and what belongs in the cloud. Traffic load, geography, headroom requirements and latency all factor in, and there is no fixed answer. “Not every workload has to be running on a cell tower,” Mahajan said.
The apps are where the value is supposed to be
Nokia frames the RAN as a programmable platform for both connectivity and compute. Distributed applications from Nokia or third parties connect through the E3 interface, and the argument is ecosystem rather than lock-in. An operator can bring in a sensing algorithm today and swap in a better one later without being tied to a single vendor.
Leboucher’s first candidates are deliberately telco. The most relevant DApps, he said, are those that need direct access to RAN data and real-time local processing, where a centralized cloud would be too slow or too inefficient. He pointed to spectrum and interference management and radio configuration across thousands of sites and cells, moving off Orange’s existing SON platform to something more software-oriented with a shorter lifecycle.
Beyond that, the roadmap opens to non-telco use cases. Leboucher named integrated sensing and communication, precise positioning, industrial safety, infrastructure monitoring and drone detection. Drone detection is handled as an ad-hoc solution today, he said, and should become native to the network. But radio alone isn’t enough — the system needs to accommodate other sensors with pattern detection close to them, which can’t be done in the cloud.
Leboucher closed by noting that this is a new operating model, and that the change is not just technology. Network operations staff move from doing activities that can be automated to supervising a much more autonomous network — understanding intent, knowing the safeguards, knowing what has to be checked and secured. He wants Orange’s existing network API work connected to the radio platform evolution, because combining radio performance, latency, sensing and geolocation is where he sees value the network can actually monetize.
Table of Contents
