Nokia Puts AI Inside the Radio Signal Chain With NVIDIA, and BT, Deutsche Telekom, Elisa and Orange Are Already Testing It

Nokia announced in Berlin this month that its AI-RAN platform, built with NVIDIA accelerated computing, has moved from laboratory demonstration into live-network trials with four European operators. BT is preparing a live-network trial in Birmingham for 2027. Deutsche Telekom and Nokia are developing AI-based channel estimation that showed 10 to 15 percent uplink throughput improvement in laboratory measurements. Elisa ran voice, mobile data and video analytics simultaneously on the same AI-RAN platform in its Helsinki 5G network. Orange plans to progress from lab testing to real-network conditions.

The technical claim is specific: Nokia is applying AI within the 5G signal-processing chain itself, including channel estimation, beamforming, multi-user MIMO pairing and scheduling. This is not AI running alongside the radio as an overlay; it is AI operating inside the physical-layer processing that determines how spectrum gets used. Nokia AI-native anyRAN software runs on NVIDIA Aerial RAN Computer, which provides the accelerated compute for these functions.

The timing matters because the RAN market is flat. Ericsson’s chief executive Per Narvinger has warned that AI-driven component costs will pressure margins through 2027 and that growth remains hard to find. Nokia’s bet is that embedding AI into the radio chain creates a revenue path that is not dependent on new spectrum or new sites, but on software upgrades to existing base stations.

Ericsson has taken a different position. Narvinger defended the company’s custom ASIC strategy for AI RAN, citing 10 percent spectrum-efficiency gains from AI-native link adaptation deployed across more than 60,000 base stations with three operators. “AI is not something that will happen in 2028,” Narvinger said. “It is happening here and now, in our networks.” Ericsson keeps GPU options open for the future but argues that custom silicon delivers functionality faster and more cost-effectively today. The two vendors are now competing on architecture as much as on price.

Nokia’s sovereign agentic AI platform for telco autonomy, also announced in Berlin, extends the AI-RAN work into network operations. The platform uses governed AI agents across network domains, with the aim of moving operators from scripted automation toward autonomous network management. Nokia said the platform builds on the industry’s first commercial AI-RAN platform, announced in July.

The competitive dynamic is clear: Nokia is betting on NVIDIA’s GPU ecosystem for AI-RAN, while Ericsson is betting on its own silicon. Both are responding to the same market condition, which is that traditional RAN revenue is not growing and operators need a reason to upgrade. The question for operators is not whether AI-RAN works in a lab, but whether the efficiency gains justify the compute cost in a commercial deployment.

GSMA estimates that European operators face a cumulative investment shortfall of 205 billion euros against the network upgrades needed for the next decade. That shortfall is the economic pressure behind every AI-RAN pitch: carriers need to do more with existing spectrum and fewer people, and software upgrades to base stations are cheaper than new sites. Ofcom tracks 52 fixed network operators in the UK alone, and the regulator has flagged that autonomous network operations will be essential for managing 5G-Advanced complexity.

AI-RAN is still in trial phase, with commercial services expected in 2027. Stock positions on current-generation baseband units remain relevant for the next 18 months.

Supplier’s note: We hold Nokia base station and baseband units from the 2025 production batch in Shenzhen stock, roughly 50 units across seven SKUs, with lead time of five to seven working days on fresh orders. Cross-reference stock includes compatible power amplifiers, filters and mounting brackets from third-party vendors that fit the same tower configuration. Customers typically order in lots of two to four units for site upgrades, and we see repeat orders from the same system integrators every five to seven weeks. Pricing holds for 30 days from quotation. We also stock spare fans, power supplies and cable assemblies for the same chassis families, and can arrange on-site installation support through our partner network in Southeast Asia and the Middle East region.