AI-RAN’s Commercial Call: 5 Things the Nokia-NVIDIA GTC Berlin Wave Means for Open RAN Operators

For two years, AI-RAN has lived in the comfortable ambiguity of the lab demo: impressive spectral-efficiency charts, carefully staged conference keynotes, and a polite silence about what happens when the technology meets a commercial network with real customers on it. That ambiguity ended on October 21, 2026, at NVIDIA’s GTC Berlin event. Nokia and NVIDIA used the opening day to announce that nearly 20 telecom operators worldwide have now publicly committed to AI-RAN collaborations built on Nokia’s anyRAN software and NVIDIA’s Aerial RAN Computer (ARC), and that the first of those commitments has crossed the line from laboratory validation to a live commercial network.

The headline milestone: T-Mobile completed what the companies describe as the world’s first AI-RAN call on a live commercial network in Bellevue, Washington. The 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. Around that anchor announcement, Elisa disclosed Europe’s first AI-RAN field trial — running voice, mobile data and vision AI together on a single accelerated computing platform over 5G — while BT, Deutsche Telekom and Orange outlined trial plans stretching into 2027.

For a technical and business audience watching Open RAN mature, this is not just another vendor press release. It is the moment the AI-RAN conversation shifts from “does it work?” to “where does it pay, and who controls the architecture?” Here are five things that shift means for operators.

1. The ‘first live call’ milestone matters less than what it proves about integration

It is tempting to file T-Mobile’s Bellevue call under marketing. Resist that temptation. The technical significance is not the call itself — operators have been making voice calls over commercial 5G for years. The significance is the integration stack the call validated: commercial ARC servers (not lab hardware), Nokia’s anyRAN software accelerated through NVIDIA CUDA, massive MIMO radios, an existing 5G Standalone core, and an off-the-shelf commercial smartphone.

That combination matters because AI-RAN’s central promise is that the same accelerated computing platform can run radio workloads and AI inference side by side. Proving a live call end-to-end on commercial equipment — rather than a bespoke testbed — is the first credible evidence that the “distributed intelligence platform” framing is an engineering reality and not a slide. It also signals that the GPU-accelerated RAN stack has cleared the basic interoperability bar with a standalone core, which is where many multi-vendor efforts historically stumble.

For Open RAN operators, the takeaway is narrower but real: the integration risk of putting general-purpose accelerated compute into the RAN path is being retired incrementally, in public, by operators willing to attach their names to it. That de-risks the architecture question for everyone else.

2. The operator roster is now the story — and it is broader than the US

Nearly 20 operators have publicly announced AI-RAN collaborations on Nokia anyRAN software and NVIDIA ARC. That is a meaningful base of named commitment for a technology that, as recently as two years ago, existed mostly as a concept at industry conferences. The roster spans regions and use cases:

• T-Mobile (US): first live AI-RAN call on a commercial network.
• Elisa (Finland): Europe’s first AI-RAN field trial, running voice, mobile data and vision AI on a single accelerated platform, plus an enterprise voice-assistant application at a distributed site.
• BT (UK): live AI-RAN trials planned in Birmingham, England, in 2027.
• Deutsche Telekom (Germany): lab validation in Bonn and Oulu, with field trials planned for 2027.
• Orange (France): hosting open models including NVIDIA Nemotron as a managed service on its Live Intelligence platform, and deploying ARC for real-time multilingual voice intelligence for emergency and field teams.

Beyond the trial operators, NVIDIA cited new infrastructure deployments by CRA in the Czech Republic, e& (etisalat) in the UAE, Silknet in Georgia and TIM in Italy, alongside expansions by Cassava in Africa, Deutsche Telekom in Germany, Fastweb + Vodafone in Italy, Kazakhtelecom in Central Asia and Orion Telecom in Serbia.

The breadth is the point. AI-RAN is no longer a US-carrier curiosity or a single-vendor showcase; it is becoming a multi-region procurement pattern. For operators still on the sidelines, the peer-pressure dynamic is now real — and so is the risk of being a late adopter of an architecture your competitors are already shaping.

3. The first hard numbers are uplink numbers — and that reframes the business case

Among the most concrete data points to emerge from the GTC Berlin wave is Deutsche Telekom’s lab result: AI-based uplink channel estimation for the Physical Uplink Shared Channel (PUSCH) improved data throughput by 10–15 percent, with gains increasing as radio channel complexity increased. Nokia provided its commercial 5G RAN software and NVIDIA its Aerial RAN Computer for the test.

This is a quietly important detail. Much of the AI-RAN narrative has focused on downlink spectral efficiency and coverage — the metrics that dominate marketing. Uplink is where many networks feel the sharpest pain, particularly as AI-driven applications, video and industrial workloads push more data from the edge toward the core. An AI technique that delivers double-digit uplink gains, and whose gains scale with channel complexity, is exactly the kind of “worthiness” evidence operators need to justify the additional compute, cost and operational complexity of accelerated infrastructure.

It also aligns with what US operators have said publicly. AT&T has trialed Ericsson’s AI-native link adaptation and reported significant improvements in throughput and spectral efficiency; Verizon has described downlink spectral-efficiency gains from AI-based link adaptation with Ericsson, plus coverage improvements in another field trial. The pattern across vendors and operators is consistent: AI in the RAN is producing measurable, if workload-dependent, gains — and the industry is now arguing about economics rather than physics.

4. The architecture fight is moving below the RAN: dApps, E3, and the app economy

Buried in the same GTC Berlin news cycle was a development that may matter more to Open RAN’s long-term structure than any single trial. Nokia said it will offer dApp software plug-ins as part of its AI-RAN ecosystem via the Open RAN standardized E3 interface. The dApp RAN specification builds on existing market adoption of rApps, which rely on the non-real-time RAN Intelligent Controller.

This is the next front in the battle over where intelligence lives. rApps operate in non-real-time control loops; dApps, delivered through E3, target more distributed, instantaneous applications closer to the radio. Nokia’s move follows planned product support for dApps by Cohere Technologies. Verizon, for its part, has described a dual RAN AI strategy combining centralized rApps in a non-real-time environment with local AI integrated into the gNodeB for real-time optimization.

The strategic implication for Open RAN operators is significant. If the AI-RAN application layer fragments into proprietary plug-in ecosystems tied to specific accelerated-compute platforms, the flexibility that Open RAN promised could be quietly re-centralized — not at the hardware layer, but at the software and interface layer. Conversely, if E3-based dApps become genuinely multi-vendor, operators gain a new lever for differentiation and a new class of suppliers to play against incumbents. The interface standards being written now will determine which outcome prevails.

5. The 2027 pipeline is the real test — and the clock is already running

Strip away the announcements and the calendar tells the honest story. T-Mobile has a live commercial call. Elisa has a field trial. Deutsche Telekom has lab validation with field trials planned for 2027. BT plans live trials in Birmingham in 2027. Orange is in trials. Most of the roster has not yet disclosed trial results the way Elisa and Deutsche Telekom have.

That gap — between named commitment and published outcome — is where AI-RAN’s credibility will be won or lost over the next twelve to eighteen months. Scaling beyond trials will require open architectures, automation, and disciplined lifecycle management across distributed networks. It will also require operators to answer the worthiness question rigorously: does AI deliver performance benefits that justify the additional compute, cost and operational complexity at scale, not just in a lab?

The competitive backdrop sharpens the stakes. Ericsson will face growing pressure to clarify whether its custom-silicon RAN strategy can match the efficiency gains claimed by the NVIDIA-Nokia camp without committing to GPU-based infrastructure. Samsung and Mavenir are likely to lean harder into hardware-agnostic Open RAN positioning as a contrast to NVIDIA’s GPU-centric pitch. And NVIDIA itself is broadening the stack: new MGX 2U and MGX 4U configurations target compute-intensive AI and radio workloads at centralized RAN facilities, while its Nemotron open models are being integrated into Nokia’s Autonomous Networks Suite and Cisco’s Crosswork AI platform, with BT, NTT DATA and TCS building telecom AI agents on them.

What operators should do now

For Open RAN operators, the practical agenda is less about whether to believe the AI-RAN hype and more about sequencing decisions deliberately:

• Demand workload-specific proof. Uplink channel estimation gains of 10–15 percent are a starting benchmark. Ask vendors for numbers against your own traffic mix, spectrum and cell density — and for the compute overhead required to achieve them.
• Watch the interface layer as closely as the hardware. E3 and dApps will determine whether AI-RAN preserves Open RAN’s multi-vendor promise or recreates lock-in one layer up.
• Treat 2027 as a decision deadline, not a distant horizon. The BT, Deutsche Telekom and Orange trials will produce the first comparative, real-world dataset. Operators without an evaluation plan by then will be choosing architecture by default rather than by design.
• Keep the economics honest. The worthiness gate applies to every operator, not just the two US carriers who articulated it most clearly. AI must pay for its compute.

The bottom line

The GTC Berlin wave did not answer every question about AI-RAN. It did something more useful: it moved the debate from possibility to proof. T-Mobile’s live call, Elisa’s field trial, Deutsche Telekom’s uplink numbers and a roster approaching 20 named operators together mark the point where AI-RAN stopped being a vendor thesis and became a procurement reality. The next chapter — scale, economics and interface governance — will be written by operators who insist on evidence. That is exactly the discipline Open RAN has always demanded, and it is the right lens through which to judge what comes next.

Sources

• AI in the RAN moves toward the scale test — RCR Wireless
• NVIDIA, Nokia and Leading Telecom Operators Turn Networks Into Distributed Intelligence Platforms — CXOToday
• NVIDIA, Nokia AI-RAN Deal Signs Nearly 20 Telcos — Tech-Insider
• T-Mobile’s AI-RAN Call Pulls In 3 More Carriers — Shattered
• NVIDIA and Nokia Expand AI-RAN Trials With Global Telecom Groups — N24
• NVIDIA, Nokia AI-RAN Reaches Nearly 20 Telcos — Shattered
• How Is NVIDIA Expanding AI-Powered Telecom Networks Following T-Mobile’s First AI-RAN Call? — Yahoo Finance
• NVIDIA, Nokia and Operators Advance AI-RAN on Live Networks — Unite.AI
• Nokia to offer dApp software features via Open RAN E3 interface — Telecompaper
• Deutsche Telekom tests AI to boost RAN throughput with Nokia, Nvidia — Telecompaper