For years, operators have been asked to believe that the next upgrade to the radio network will unlock better economics. Nokia’s latest AI-native RAN push is interesting because it reframes that promise: not as another hardware cycle, but as a software-and-compute shift that aims to extract more capacity from the same spectrum and sites. The opportunity is real, but so are the caveats. Nokia, in its own press release, says it has moved beyond concept demos into operator testing and early commercial positioning, yet the hardest questions still concern energy, total cost, governance and whether new edge revenues will arrive fast enough to justify the architecture.
The radio access network has long been the part of telecoms most associated with scale which involves more spectrum, more sites, more radios, more power, more capex. Nokia’s current argument is that this logic is reaching its limits. In July 2026, the company launched what it calls the industry’s first commercial AI-native RAN platform, built on its anyRAN software and NVIDIA’s Aerial AI-RAN platform, with the explicit claim that operators should be able to get materially more capacity from spectrum and infrastructure they already own rather than relying only on new hardware cycles. That is a significant shift in narrative, and it lands at a moment when the wider industry is wrestling with flat RAN growth, slowing capex and pressure to make networks support AI workloads as well as connectivity.
The important question is not whether Nokia can produce another wave of product messaging. It is whether AI-native RAN offers a credible answer to three operator problems at once: rising traffic, stubborn economics and the search for new edge-era revenues. Nokia’s recent operator engagements with T-Mobile US, Indosat, SoftBank, Deutsche Telekom, Orange, Telia, Virgin Media O2 and Taiwan Mobile suggest the company has enough momentum to make this worth serious attention, but the public record still supports a cautious reading rather than a triumphalist one.
The timeline above is assembled from Nokia and operator announcements spanning March to July 2026. It shows that Nokia’s AI-RAN story is not based on a single July launch, but on a sequence of ecosystem, operator and commercial milestones that built towards that launch.
Why conventional RAN economics look increasingly unsustainable
The case for AI-native RAN starts with an awkward industry truth: traditional mobile economics are not especially generous. Dell’Oro expects worldwide telecoms capex to fall by 2% in 2026, while carrier revenue is projected to rise only modestly at around 2% CAGR through to 2030; the same forecast points to wireless capital intensity falling towards 11% by 2029. Separately, Dell’Oro says the RAN market is not a long-term growth market and expects worldwide RAN revenues to grow at only 1% CAGR over the next five years, with 6G investment not expected to meaningfully ramp up until later in the forecast period. Operators still need to improve network performance, but the financial room for doing so through brute-force expansion is limited.
That tension shows up even more clearly in automation research. Dell’Oro notes that mobile data traffic grew by 16% in 2025, while operator revenue grew at just 1% CAGR over the past decade, leaving operators with little room to expand capex and opex simply to manage more complexity. GSMA has framed the same problem more bluntly: the future of 5G is about intelligence and programmability, and “spectral efficiency alone does not guarantee returns”; meanwhile, AI innovation cycles move in months while telecoms standardisation still moves in years. The strategic implication is that operators do not just need a better radio. They need a more adaptive cost model.
Nokia’s own recent results help explain why the company is leaning into this story now. In Q1 2026, Nokia said sales to AI and cloud customers grew 49% year on year and accounted for 8% of group sales, while network infrastructure net sales grew 6%. In Q2 2026, Nokia said sales to AI and cloud customers grew 105% and network infrastructure net sales grew 12%; Reuters separately reported AI and cloud customer net sales of €446 million in the quarter. Nokia is therefore pursuing AI-native RAN against a backdrop in which conventional telco spend is constrained, but AI-related infrastructure demand is visibly stronger.
This chart uses Nokia’s latest public quarterly figures. It compares two different but related disclosures: sales growth to AI and cloud customers, and growth in the Network Infrastructure segment. They are not like-for-like revenue categories, so the chart should be read as directional evidence of Nokia’s changing momentum rather than as a clean segment comparison. Public data does not yet isolate AI-RAN revenue as a standalone line item.
What Nokia means by AI-native RAN
Nokia’s definition of AI-native RAN is more ambitious than adding another layer of analytics to network operations. According to its July launch announcement, the platform combines Nokia anyRAN software with NVIDIA’s Aerial AI-RAN platform and is designed to run AI “at radio timescales”, across 4G, 5G and future network evolution, while also supporting existing Nokia radios and O-RAN-compliant radio units. Nokia also says the offer will arrive through a software subscription model, with pilot deployments starting at the end of 2026 and commercial availability in 2027, which matters because it suggests a change in commercial packaging as well as architecture.
That is an important distinction. In the older framing, AI helps operate the network more efficiently from the side. In Nokia’s newer framing, AI becomes part of how the radio network itself is built, tuned and monetised. The clearest public description of this shift came in Nokia’s March MWC update, which described AI-RAN as a “foundational step” toward cognitive, software-defined wireless networks and highlighted joint work with operators including T-Mobile US, Indosat, SoftBank, BT, Elisa, NTT DOCOMO and Vodafone Group. Omdia’s post-MWC summary also described AI-RAN as the main RAN theme of MWC Barcelona 2026, with “AI-ready radios” becoming a visible part of vendor roadmaps. Nokia is therefore moving with the market, but it is also trying to define the commercial language of that market before rivals do.
The operator partnerships tell us what this means in practice. Nokia and Deutsche Telekom said in March that they were expanding work on open fronthaul integration, cloud RAN and multivendor flexibility as part of Deutsche Telekom’s O-RAN strategy. Nokia and Telia said they would jointly develop and test AI-RAN use cases with a focus on commercial applications, including mission-critical scenarios. Nokia and Orange said their work with NVIDIA would explore predictive optimisation, spectral efficiency and new services such as integrated sensing and communication. Taken together, these announcements suggest Nokia sees AI-native RAN not as a single product box, but as a stack combining radio optimisation, cloudification, multivendor integration and new service logic.
Why the spectral-efficiency claim matters
The headline claim in Nokia’s July announcement is that its AI-native RAN platform could deliver more than 100% spectral-efficiency gains by 2028, after the more than 20% gains already demonstrated and a stated path to 50% by 2027. Even if one treats those figures cautiously, they matter because spectral efficiency is where the technical story meets the balance sheet. If an operator can transport materially more traffic over the same spectrum holdings and installed radio assets, the effective value of those assets rises without the same immediate pressure for new spectrum purchases or dense site expansion.
Still, the right reading is not “Nokia has doubled spectrum value” but “Nokia is trying to prove that software-led radio improvements can become economically meaningful”. GSMA’s warning that spectral efficiency alone does not guarantee returns is useful here, because better radio performance only translates into better business if it reduces cost per bit, delays capex, improves experience in congested cells, or supports premium services that customers will pay for. Analysys Mason has made the same point from a different angle, arguing that operators will only deploy AI in the RAN if they can see a transformed cost base rather than a more fashionable but costlier version of the same network problem.
The evidence today still falls short of full independent validation. Nokia’s own March and July materials show meaningful progress, but they remain vendor-led disclosures. That makes operator context especially important. Virgin Media O2’s March announcement of multi-year RAN modernisation deals with Nokia and Ericsson for thousands of sites, alongside Nokia’s separate announcement of its own role in that programme, suggests that operators are still making large-scale investment decisions around conventional 5G performance, capacity and coverage even as AI-native ideas advance. Nokia’s challenge is to show that AI-native RAN can sit inside those real-world transformation programmes rather than remain a parallel innovation track.
How RAN and AI workloads could share infrastructure
This is where Nokia’s current story becomes genuinely more interesting than a standard RAN upgrade pitch. In its March update, Nokia said that with T-Mobile US and NVIDIA it had tested GPU-accelerated AI-RAN workloads in T-Mobile’s over-the-air AI-RAN Innovation Centre, combining radio workloads and AI applications on a single NVIDIA Grace Hopper 200 server. T-Mobile’s own statement in March described the same direction in more operator-centric language: distributed edge AI networks that can support “physical AI” applications over AI-RAN-ready infrastructure. That is the most concrete public sign that Nokia’s AI-native RAN thesis is really about converged compute, not just smarter scheduling.
Nokia has pointed to two further examples that sharpen the argument. With Indosat, it said it achieved Southeast Asia’s first AI-RAN-powered Layer 3 5G call, using an open, cloud-native network with Nokia RAN software accelerated by NVIDIA GPUs. With SoftBank, it demonstrated how spare AI-RAN compute capacity identified by SoftBank’s orchestration layer could be used for third-party AI tasks, framing the RAN as a platform for AI services beyond connectivity. If these models hold up at scale, a RAN site stops being a single-purpose network cost centre and starts to look more like a distributed compute asset.
Analyst work suggests operators are already positioning infrastructure in that direction. Omdia says 70% of surveyed operator decision-makers view support for AI and ML workloads as a key factor in cloud infrastructure decisions, and 58% identify GPUs among the options they are deploying or considering for AI processing and inferencing. The same Omdia research argues that many operators are cloudifying RAN and core into distributed architectures spanning central, regional and edge locations, which naturally supports different AI workloads across different latency profiles. Nokia’s architecture, in other words, is not emerging in isolation. It is trying to match an operator infrastructure trend that analysts already see developing.
Where Nokia sees commercial upside at the edge
The most credible near-term upside is still operational rather than transformational. If AI-native RAN improves spectral efficiency, energy management, fault handling and congestion performance, operators can justify it as a network economics play before they need to prove entirely new revenue lines. Dell’Oro’s current market view supports that reading: most AI-RAN activity today is centred on AI-for-RAN, improving performance and efficiency inside the existing mobile model, and AI RAN is expected to surpass US$10 billion and reach roughly one-third of the total RAN market by 2029, though that is not all net-new revenue. This should temper some of the more exuberant talk around AI-RAN as an immediate revenue revolution.
Where the commercial story becomes more interesting is at the edge, where Nokia is effectively testing three monetisation ideas at once. The first is shared infrastructure, where spare compute can host non-RAN AI tasks, as shown in SoftBank’s public demonstration with Nokia. The second is industry-focused use cases, which Nokia and Telia explicitly said they will explore, including mission-critical applications. The third is service orchestration, where Nokia’s agentic network slicing work with AWS, du and Orange aims to use AI to translate conditions such as traffic, events and location into adaptive network slices that could underpin premium services. None of these models is fully proven at scale, but together they show how Nokia is trying to move the discussion from “better radio” to “programmable, revenue-aware network edge”.
Omdia’s edge outlook again supports a measured rather than promotional interpretation. Its 2026 “Trends to Watch” work on telco edge and AI explicitly points to AI inferencing at the edge, neocloud partnerships and AI monetisation on RAN as central themes for operators this year. That does not mean monetisation is solved. It means the direction of travel is now clear enough that large infrastructure vendors such as Nokia are aligning product strategy around it.
The risks operators should keep in view
The first risk is cost and energy. AI-native RAN is often sold as a more efficient way to run the network, but it introduces additional compute demands that can narrow or even erase efficiency gains if badly designed. GSMA has warned that, under a base scenario, AI-driven traffic could raise operator energy consumption by around 25% by 2030. Dell’Oro’s analysis of AI RAN is similarly careful: it sees long-term momentum, but also emphasises persistent constraints around power budgets, strict cost controls and the practicality of supporting non-telco workloads at scale, especially at distributed RAN sites. In plain terms, bigger claims about spectral efficiency will not matter much if the compute bill grows faster than the capacity benefit.
The second risk is strategic dependency. Omdia’s survey work notes that hardware choices for AI-RAN can become heavily shaped by vendor partnerships, and explicitly contrasts proprietary ASIC-led approaches with software-defined strategies using commercial off-the-shelf servers and partners such as Intel, NVIDIA and AMD. Omdia also warns that low operational costs and deterministic performance can come with significant upfront investment and vendor lock-in. Nokia’s pitch is more open than a single-vendor ASIC story, but it is still deeply tied to NVIDIA’s accelerated computing ecosystem in its current public form. That is commercially important for operators that spent the last few years reducing dependency through Open RAN and broader cloud optionality.
The third risk is governance. Once AI is pushed deeper into network control, the operational problem is no longer just model accuracy. It is traceability, accountability and policy control across a live network. Nokia’s broader June announcements show that the company itself recognises this: it expanded work with Google Cloud to embed Gemini-based AI agents into its assurance software, added agentic AI capabilities across its autonomous-network portfolio, and deepened collaboration with AWS around autonomous networks for the AI era. Those are not RAN releases, but they matter because they show AI-native RAN is likely to sit inside a wider shift toward machine-speed decisioning across network operations. Operators will need stronger controls, not fewer.
Key AI-RAN considerations for telcos:
- Assess AI-native RAN as an economics question first. Ask what it does to cost per bit, deferment of spectrum or site spend, and energy per carried gigabyte before treating it as a 6G story.
- Treat shared infrastructure as the decisive architectural test. If RAN and AI workloads can coexist on common accelerated platforms without unacceptable latency, energy or reliability penalties, the commercial logic becomes far stronger.
- Scrutinise dependency and portability. Open interfaces, cloud portability and hardware choice will matter even more in AI-RAN than they did in first-generation cCloud RAN debates.
- Link RAN discussions to wider autonomous-network governance. Nokia’s concurrent moves in agentic assurance, cloud-based autonomy and AI-driven slicing suggest that RAN intelligence will not remain a siloed domain problem.
- Assume public data remains incomplete. There is still no clean public revenue breakout for AI-RAN itself, and the strongest efficiency claims remain vendor-reported rather than independently benchmarked at scale.
Is this really a path to 6G
Nokia clearly wants AI-native RAN to be read as the bridge from 5G-Advanced to 6G. Its March MWC update described AI-RAN as paving the way to AI-native 6G, while its July launch framed the platform as the practical path to AI-native networks. GSMA’s May 2026 6G progress material likewise places AI integration, programmability and intelligent network evolution near the centre of the 6G discussion rather than at its margins. From that perspective, Nokia’s thesis is plausible: if 6G begins as a software, automation and compute transition before it becomes a fully new radio cycle, then AI-native RAN is exactly the kind of intermediary architecture operators should be testing now.
The more disciplined conclusion, though, is that AI-native RAN is better understood today as a route to a different operating model rather than a guaranteed route to 6G success. Dell’Oro expects AI RAN and Cloud RAN to play major roles in the second half of the 5G cycle and the early 6G era, but it also stresses that adoption paths will depend on trade-offs involving flexibility, performance, energy efficiency, total cost of ownership and speed to market. That is probably the right final frame: Nokia has made AI-native RAN one of the most serious conversations in mobile infrastructure this year, but the conversation is still about operator proof, not vendor proof.
For telecoms professionals, the practical takeaway is straightforward. Nokia’s latest developments are worth following not because they prove the future has arrived, but because they make the next strategic questions much harder to avoid.
Marion Webber