AI-RAN Leaves the Lab: A Skeptical Exec's Q&A on T-Mobile's Live Call, Elisa's Field Trial, and Nokia-NVIDIA's 2027 Playbook

October 21, 2026
10 mins
Technology
ai ran elisa nokia nvidia open ran t mobile telecom

AI-RAN Leaves the Lab: A Skeptical Exec’s Q&A on T-Mobile’s Live Call, Elisa’s Field Trial, and Nokia-NVIDIA’s 2027 Playbook

At GTC Berlin on October 21, 2026, NVIDIA and Nokia did something the AI-RAN conversation has been promising for years but rarely delivering: they put a date-stamped, named-operator, commercially grounded milestone on the table. T-Mobile completed what the pair describe as the world’s first AI-RAN call on a live commercial network in Bellevue, Washington. Elisa ran Europe’s first AI-RAN field trial in Helsinki, carrying voice, mobile data and vision AI simultaneously on one accelerated platform over its 5G Advanced network. BT, Deutsche Telekom and Orange are queued behind them. Nearly 20 operators have now publicly announced AI-RAN collaborations built on Nokia’s anyRAN software and NVIDIA’s Aerial RAN Computer (ARC).

That is a lot of headline. So rather than repeat the press release, let’s do what a skeptical executive would do: interrogate it. Below are the questions a CTO, CFO or network strategy lead should be asking — and what the disclosed facts actually support.

”A ‘first AI-RAN call’ — isn’t that just a marketing label on a demo?”

It is fair to be suspicious, but the technical details here matter. According to NVIDIA’s announcement, the T-Mobile call ran on commercial ARC servers, Nokia’s CUDA-accelerated RAN software, and massive MIMO radios, connected to an existing 5G Standalone core and a commercial smartphone. In other words, it was not a bespoke lab rig with a proprietary handset — it used the operator’s own core and an off-the-shelf device.

That does not mean the network is running AI-RAN at scale. It means the fundamental question — can accelerated computing carry a standards-based call on commercial infrastructure? — has a demonstrated answer. Jensen Huang, speaking at GTC Berlin, framed it as the culmination of nearly eight years of work. Treat the milestone as a gate that has been passed, not a product that has shipped.

”What did Elisa actually prove that T-Mobile didn’t?”

Different thing, and arguably more interesting for European operators. Elisa’s Helsinki trial ran voice, mobile data and video analytics at the same time on the same AI-RAN platform over a live 5G Advanced network. That is the “AI on RAN” half of the equation — the network doing AI workloads, not just using AI to optimize radio.

Elisa’s COO Sami Komulainen described the goal as a distributed AI inference network in Finland, bringing AI closer to users, enterprises and data, and called the trial a foundation for AI-native 6G. Nokia’s Emma Falck framed the same result in commercial terms: combining connectivity and AI computing lets operators use infrastructure more efficiently and create new revenue from edge AI services.

For a skeptical exec, the takeaway is that the two headline trials cover different halves of the AI-RAN thesis. T-Mobile proved the radio works on accelerated compute. Elisa proved the compute can do something other than radio. You need both before the business case closes.

”Fine — but when does anyone actually make money?”

The honest answer from the announcements themselves: not this year. Elisa expects distributed deployment, customer use-case evaluation and a path toward commercial AI-RAN services in 2027. BT is planning live-network trials in Birmingham, England in 2027. Deutsche Telekom has validated the Nokia-NVIDIA commercial AI-RAN platform in Bonn and Oulu labs and plans to move to field trials in 2027. Orange plans to progress testing in both lab and real network conditions.

So 2026 is the year of live proof points; 2027 is the year of field trials and early commercial evaluation. Anyone building a 2026 business case around AI-RAN revenue is ahead of the disclosed timeline. Anyone building a 2028-2030 case around it is roughly aligned with what operators are saying publicly.

”Where’s the measurable performance gain? I need numbers, not vision.”

There is one concrete figure worth anchoring on. Deutsche Telekom and Nokia are developing AI-based channel estimation, and initial laboratory measurements showed typical uplink throughput improvements of 10–15%, with greater gains under more complex radio conditions. The next phase extends this to additional AI spectral-efficiency algorithms and field validation.

That is a real, if modest, number — and note it is lab-measured, not field-proven. It also sits squarely in the “AI for RAN” bucket: using AI to squeeze more out of existing spectrum. The 10–15% uplink gain is the kind of figure a CFO can work with when weighing whether accelerated compute earns its place in the radio site. The larger “AI on RAN” revenue story — edge inference, vision AI, new enterprise services — remains largely qualitative at this stage.

”Isn’t AI-RAN just a Nokia-NVIDIA duopoly play that locks me in?”

This is the sharpest question, and the answer is mixed. On one hand, the platform is explicitly a two-vendor stack: Nokia’s anyRAN software on NVIDIA’s ARC hardware. Nearly 20 operators have signed up to that combination. On the other hand, the infrastructure layer is broadening. Supermicro announced it is bringing NVIDIA Aerial RAN Computer solutions to the network edge with Nokia, with modular, right-sized systems spanning distributed cell sites, aggregation locations and core hubs. Supermicro’s Charles Liang positioned the play around flexibility and deploying AI capacity precisely where it is needed.

NVIDIA’s Soma Velayutham framed the Supermicro collaboration as helping operators deploy ARC across diverse environments “from distributed cell sites to centralized RAN locations.” That is a hardware ecosystem forming around the platform, which is healthier than a single-box story — but it is still an ecosystem anchored on one accelerator vendor and one RAN software vendor. Operators should read “nearly 20 operators” as evidence of momentum, and also as a question about how much bargaining power they retain.

”What about the software layer — is this open or proprietary?”

Here the picture gets more nuanced and, for an Open RAN audience, more consequential. Nokia is bringing its anyRAN software onto NVIDIA ARC for trials at BT, Deutsche Telekom, Elisa and Orange. Separately, NVIDIA is pushing its Nemotron family of open AI models into telecom network management: Nokia is incorporating Nemotron into its Autonomous Networks Suite, and Cisco is integrating fine-tuned versions into its Crosswork AI platform. TCS, BT and NTT DATA are adopting Nemotron to build AI agents for network operations, targeting network management, site inspections and fault remediation.

So the model layer is being positioned as open, while the RAN software and accelerator layers remain vendor-specific. That is a familiar pattern, and it is worth watching whether the open-model story translates into genuine multi-vendor flexibility at the RAN layer or functions mainly as a software on-ramp to the Nokia-NVIDIA stack.

”The sovereign AI angle — is that real or just a European buzzword?”

It appears to be more than buzzword. NVIDIA is reportedly helping European telecom operators build sovereign AI factories, with around 10 EMEA operators committed to using its accelerated computing hardware for industrial AI clouds and AI-native RAN. New infrastructure announcements include CRA in the Czech Republic, etisalat in the UAE, Silknet in Georgia and TIM in Italy. Expansions cover Cassava in Africa, Deutsche Telekom in Germany, Fastweb + Vodafone in Italy, Kazakhtelecom in Central Asia and Orion Telecom in Serbia.

Deutsche Telekom CEO Tim Höttges tied the theme to industrial relevance — bringing intelligence closer to high-tech industries, and arguing that companies perform best with optimal networks and smart AI. For operators, the sovereign framing matters because it reframes AI-RAN spending as strategic infrastructure rather than a RAN capex line item. That can help unlock budget, but it also raises governance and data-residency questions that the technical trials do not answer.

”What’s the actual architecture story here — AI for RAN, AI on RAN, or both?”

Both, and the announcements are unusually explicit about it. AI-RAN extends accelerated computing into the radio access network, turning it into an intelligence workload that runs on the same platform as advanced services, so operators can get more out of their spectrum while serving demand from sites close to users. Nokia describes the platform as built around three capabilities: AI-led spectral efficiency to extract more capacity from existing spectrum, integrated sensing, and an expanding infrastructure ecosystem for physical-AI and edge-AI applications.

That three-part structure is the clearest articulation yet of how the vendors want operators to think about the investment. Spectral efficiency is the near-term, measurable justification. Sensing and edge AI are the longer-term optionality. A skeptical exec should fund the first and demand evidence before betting the farm on the second and third.

”What’s the competitive context — is anyone else close?”

The announcements position Nokia and NVIDIA as the ones with live-network proof points and a named 2027 trial roster. That is a meaningful lead in mindshare. But the broader context includes SpaceX’s stated plan to turn Starlink into a major U.S. mobile carrier, which has already prompted analyst caution on incumbent operators — Bernstein reportedly cut its T-Mobile target citing that pressure. AI-RAN is, in part, an incumbent’s answer to new competition: if you cannot out-build a satellite constellation, you can try to make your terrestrial network smarter and more valuable per site.

That framing matters for how operators justify the spend internally. AI-RAN is not only a spectral-efficiency play; it is a defensive move to make existing infrastructure do more as new entrants reshape the competitive map.

”So what should I actually do with this information?”

A few practical takeaways for the skeptical executive:

Treat 2026 as proof, 2027 as trial, 2028+ as commercial. The disclosed timelines from Elisa, BT, Deutsche Telekom and Orange all point to 2027 for field trials and early commercial evaluation. Plan your investment gates accordingly.

Anchor on the 10–15% uplink figure, and verify it in your own conditions. It is the one hard performance number on the table, it is lab-measured, and it is explicitly slated for field validation. Make that validation a condition of scaled spend.

Separate the two halves of the thesis in your business case. AI for RAN (spectral efficiency, channel estimation) has near-term measurable value. AI on RAN (edge inference, vision AI, new services) is where the growth story lives but where evidence is thinnest. Fund them as distinct bets.

Watch the ecosystem, not just the headline. Supermicro’s entry and the Nemotron model integrations suggest the stack is broadening. Track whether that broadening reaches the RAN software layer or stays at the infrastructure and model edges.

Ask hard questions about lock-in. Nearly 20 operators on one Nokia-NVIDIA combination is momentum, but it is also concentration. Negotiate accordingly.

AI-RAN has moved from conference-slide ambition to live-network calls and named field trials. That is genuine progress, and it deserves to be taken seriously. But the gap between a first call and a commercial service is where most technology promises die — and on the disclosed evidence, that gap is still roughly a year wide, with the revenue case still largely unproven. The operators who win will be the ones who use 2026’s proof points to ask better questions, not to write bigger checks.