Photo by alexander ermakov on Pexels

The Gap Between the Lab and the Live Network

At the recent Intelligent RAN Forum, Nokia made headlines with a striking claim: a 20% near-term improvement in spectral efficiency, scaling toward a full 2x gain by 2028 through the application of artificial intelligence and machine learning across its radio access network portfolio. It’s the kind of headline that turns heads in boardrooms and at analyst briefings alike. But for operators like Orange, the more pressing question isn’t what the numbers look like on a slide deck — it’s what they look like on a live cell site in central Paris or suburban Lyon.

The contrast between Nokia’s forward-looking projection and Orange’s measured skepticism encapsulates one of the central tensions in today’s telecom industry: vendors racing to define the AI-native RAN future, while operators demand accountability in the present.

What Nokia Is Actually Claiming

Nokia’s spectral efficiency roadmap is tied to its AI RAN strategy, which involves embedding machine learning models directly into the RAN stack — from the baseband unit down to the radio unit — to dynamically optimize beamforming, scheduling, interference management, and energy use in real time. The company has outlined a phased approach: incremental efficiency gains in the near term, building toward transformative improvements as AI models mature and training datasets grow richer with live network telemetry.

The 2x spectral efficiency figure represents the theoretical ceiling of what coordinated, AI-optimized multi-layer networks could achieve by 2028, assuming continued advancements in massive MIMO, network slicing, and cloud-native RAN architectures. Nokia has pointed to trials and controlled lab environments where early AI RAN features have already delivered measurable throughput improvements and notable reductions in energy per bit — a critical metric as operators face mounting pressure over operational costs and sustainability commitments.

The Role of AI in the RAN Stack

To understand the potential, it helps to look at where AI is actually being applied. In Nokia’s roadmap, AI touches several layers: predictive interference coordination between cells, intelligent handover optimization that reduces ping-pong between towers, AI-driven scheduler enhancements that better allocate radio resources under variable load conditions, and closed-loop automation that adjusts antenna tilt and power dynamically. Each of these individually offers modest gains — it’s the compounding effect across all layers simultaneously that vendors believe will eventually unlock the larger efficiency multiplier.

Orange Wants Proof, Not Promises

France’s incumbent operator Orange has been vocal about what it needs before committing to large-scale AI RAN deployment: demonstrable, reproducible results in live production networks. The operator’s position reflects a broader caution among Tier 1 European carriers who have been burned before by features that performed well in vendor-managed trials but underdelivered when handed off to operations teams working with mixed-vendor infrastructure and real-world traffic variability.

Orange’s concerns are not just technical. They are also operational. Deploying AI-driven optimization across tens of thousands of active sites requires new tooling, new expertise, and new ways of managing network behavior that is increasingly autonomous. Questions around explainability — can the network’s AI decisions be audited and understood by human engineers? — are central to Orange’s evaluation criteria. So is interoperability: does Nokia’s AI RAN solution play well in a multi-vendor environment, or does efficiency require a mono-vendor lock-in that operators have spent years trying to avoid?

The Open RAN Question Lurking in the Background

Orange’s push for live validation also carries implicit weight in the broader Open RAN conversation. As operators have invested in disaggregated, open interfaces — partly to avoid exactly the kind of vendor dependency that proprietary AI models could reintroduce — any AI RAN framework needs to demonstrate compatibility with O-RAN Alliance specifications. Nokia has made commitments in this direction, but operators are watching closely to see whether the most powerful AI features remain locked inside proprietary silicon or whether they can be exposed through standardized interfaces that third-party application developers and system integrators can build upon.

A Familiar Industry Dynamic

This is not the first time the telecom industry has navigated the gap between transformative vendor claims and operator ground-level caution. The early days of network function virtualization (NFV) saw similar dynamics, with vendors promising capex and opex savings that took years longer to materialize than projected. 5G itself was sold on use cases — ultra-low latency industrial automation, network slicing revenue streams — that have been slower to commercialize than the industry once hoped.

That history makes operators like Orange understandably deliberate. Large capital commitments to AI RAN infrastructure need to be justified by business cases that hold up under scrutiny, not just vendor-sponsored KPIs.

Industry Outlook: Patience as a Strategic Asset

The trajectory toward AI-native RAN is almost certainly the right direction — the question is one of timeline and transparency. Nokia’s 2028 target for 2x spectral efficiency is ambitious but not implausible, particularly as GPU-accelerated baseband processing matures and AI training pipelines become more standardized. However, the operators who will ultimately fund that vision are right to insist on staged, verifiable milestones in live environments.

For the broader ecosystem, Orange’s stance may ultimately benefit everyone. Vendor claims tested against real-world operator scrutiny produce better products, more honest roadmaps, and more durable business relationships. If Nokia can deliver even a consistent 20% spectral efficiency improvement across diverse live deployments by 2025 or 2026, it will have built the credibility needed to make the 2028 headline believable — and investable.

The intelligent RAN era is coming. But between here and there, the most important frequency to get right isn’t measured in gigahertz. It’s the frequency of trust between vendors and the operators who actually run the networks.