The AI-RAN Call That Mattered: T-Mobile, Nokia and NVIDIA Turn a Live 5G Network Into a Compute Platform — and Nearly 20 Operators Follow

The AI-RAN Call That Mattered: T-Mobile, Nokia and NVIDIA Turn a Live 5G Network Into a Compute Platform — and Nearly 20 Operators Follow

On October 21, 2026, at the opening day of GTC Berlin, NVIDIA and Nokia used a single announcement to reframe the AI-RAN conversation. T-Mobile completed what the companies describe as the world’s first AI-RAN call on a live commercial network in Bellevue, Washington — running on commercial Aerial RAN Computer (ARC) servers, Nokia’s CUDA-accelerated anyRAN software and Massive MIMO radios, connected to an existing 5G Standalone core and a commercial smartphone. In the same breath, NVIDIA and Nokia disclosed that nearly 20 telecom operators globally have now publicly announced AI-RAN collaborations built on that same software-and-hardware pairing, with BT, Deutsche Telekom, Elisa and Orange named as the next wave of trials.

For an Open RAN audience, the headline is not the call itself. It is the architecture behind it: accelerated computing — the GPU platforms that dominate AI training and inference — is being pulled into the radio access network, positioning the RAN as an intelligence workload rather than a purpose-built, fixed-function appliance. That is a structural claim with real consequences for how operators think about vendor flexibility, spectrum economics and the Open RAN investment case. It also deserves the skepticism that any first-of-its-kind milestone invites. This note separates what was demonstrated from what was asserted, and lays out what operators should watch next.

What Was Actually Demonstrated

The T-Mobile achievement is best understood as a commercial-network integration milestone, not a production-scale economic proof point. The call ran on commercial ARC servers with Nokia’s NVIDIA CUDA-accelerated RAN software and massive MIMO radios, connected to an existing 5G Standalone core and a commercial handset. That combination matters: it shows the AI-RAN stack can interoperate with a live operator core and off-the-shelf user equipment rather than a lab bench. Independent commentary — including a detailed geopolitical analysis of the announcement — was careful to note that the milestone does not yet demonstrate production-scale economic superiority, sustained multi-workload performance, or autonomous 6G functionality. That is the right framing. A first call proves the plumbing works; it does not prove the business case.

Omdia senior principal analyst Gabriel Brown, quoted by Light Reading, called it a “wow moment” and emphasized the more durable point: AI models deployed directly in the basestation can deliver measurable performance gains through software updates to deployed equipment. That is the sentence operators should underline. If intelligence can be added to radios via software, the upgrade economics of the RAN change — and that is precisely the argument Open RAN has been making for a decade.

The Operator Map: Who Is Doing What

The announcement was not a single-carrier story. NVIDIA and Nokia laid out a pipeline that spans the US and Europe, and it is worth mapping carefully because the maturity levels differ materially.

  • T-Mobile (US): Completed the first live AI-RAN call on a commercial network in Bellevue, WA, on Nokia anyRAN software and NVIDIA ARC.
  • Elisa (Finland): Completed Europe’s first AI-RAN field trial, running voice, mobile data and vision AI simultaneously on a single accelerated computing platform over 5G in Helsinki. Elisa has also deployed ARC at a distributed network location running a voice intelligence application as an enterprise assistant, and is using open models to build edge applications.
  • Deutsche Telekom (Germany): Validated the Nokia-NVIDIA commercial AI-RAN platform in Bonn and Oulu labs, including AI-enhanced uplink channel estimation. Initial lab measurements reportedly showed typical uplink throughput improvements of 10–15%, with greater gains under more complex radio conditions. Field trials are planned for 2027. Separately, Deutsche Telekom is expanding its NVIDIA-powered Industrial AI Cloud, hosting open models including Nemotron and Soofi and serving robotics companies such as Agile Robots.
  • BT (UK): Preparing live AI-RAN trials in Birmingham, England, in 2027, using Nokia’s 5G software on NVIDIA ARC and assessing capacity, reliability and spectral efficiency against Nokia’s reported platform gains.
  • Orange (France): Progressing on AI-RAN testing in both lab and real network conditions, hosting open models including Nemotron as a managed service on its Live Intelligence trusted AI platform, and deploying ARC at distributed sites for voice intelligence using Nemotron Speech — including real-time multilingual translation for emergency and field services teams.

Beyond the trial cohort, NVIDIA said new infrastructure deployments were announced by CRA in the Czech Republic, etisalat in the UAE, Silknet in Georgia and TIM in Italy, with expansions at Cassava in Africa, Deutsche Telekom in Germany, Fastweb + Vodafone in Italy, Kazakhtelecom in Central Asia and Orion Telecom in Serbia. The company also flagged nine operators across Europe, the Middle East and Africa developing or expanding NVIDIA-based infrastructure. The distinction between “AI-RAN trial” and “accelerated computing deployment” matters — much of that list is the latter, not the former.

The Numbers Behind the Hype

The performance claims circulating around this announcement are directionally consistent but not yet comparable across carriers. Nokia has reported spectral-efficiency improvements of more than 20% in its wider AI-RAN platform testing. Deutsche Telekom’s lab work on AI-based channel estimation showed 10–15% typical uplink throughput gains, with greater gains under more complex radio conditions. Elisa’s field trial demonstrated that radio and AI workloads can coexist on one platform in a live network environment — an operational claim more than a throughput claim.

For context, Light Reading noted that T-Mobile separately revealed a different AI achievement the day before: an AI-managed beamforming feature running on Ericsson RAN compute and silicon in New York City, with results including a 16% improvement in downlink throughput, 12% better spectral efficiency, a 10% increase in downlink data volume, and 50% more advanced beamforming using full-band SRS. That was the carrier’s second AI model deployed in Ericsson RAN this year, following an AI-native scheduler with link adaptation that showed 10% better spectral efficiency and a 15% increase in downlink throughput.

The juxtaposition is the most instructive data point in the entire news cycle. T-Mobile is running AI in RAN on two tracks: GPU-accelerated AI-RAN with Nokia and NVIDIA, and AI embedded in Ericsson’s custom silicon. Both produced measurable gains. Neither has yet published the cost-per-bit math that would settle the architecture debate. Operators should treat the percentage improvements as credible proof that AI in the RAN works, and treat the platform choice as still genuinely open.

Why This Matters for Open RAN

The AI-RAN narrative is often framed as a departure from Open RAN. It is better understood as the next step in the same direction. Open RAN spent a decade separating radio software from proprietary hardware so operators could mix vendors. AI-RAN asks whether the generic infrastructure underneath can be the same GPU-based accelerated computing that already dominates AI workloads, rather than a separate category of telecom-specific merchant silicon. Where vRAN and Open RAN were mainly about flexibility and vendor choice, AI-RAN adds a genuinely new capability: the same servers that process radio signals can also run AI inference, turning idle compute capacity at cell sites into a usable resource rather than overhead that exists only to handle peak load.

The strategic implication is that the RAN becomes a distributed intelligence platform. NVIDIA’s framing — adding accelerated computing to centralized data centers, metro sites and cell sites to push AI models physically closer to users, businesses and machines — is essentially an argument that the telco edge is the natural home for inference. Orange’s multilingual voice intelligence for emergency responders and Deutsche Telekom’s industrial robotics demonstrations are early examples of what that looks like in practice. If those use cases monetize, the RAN stops being a cost center and starts being a revenue platform. That is a far more powerful investment thesis than spectral efficiency alone.

The Competitive Read-Across

The announcement also sharpens the competitive map. Ericsson faces growing pressure to clarify whether its custom-silicon RAN strategy can match the efficiency gains NVIDIA and Nokia are claiming, even without committing to GPU-based infrastructure. Samsung and Mavenir are likely to lean harder into their Open RAN, hardware-agnostic positioning as a direct contrast to NVIDIA’s GPU-centric pitch, framing vendor flexibility as the core differentiator. That is a coherent counter-position: if AI-RAN requires a specific accelerator vendor’s platform, the multi-vendor promise of Open RAN narrows considerably.

There is also a software layer worth tracking. NVIDIA’s Nemotron family of open models is being integrated into network management: Nokia is incorporating Nemotron into its Autonomous Networks Suite, Cisco is integrating fine-tuned versions into its Crosswork AI platform, and Tata Consultancy Services, BT and NTT DATA are adopting Nemotron to build AI agents for telecom operations — targeting network management, site inspections and fault remediation. The open-model angle is strategically important for Open RAN operators because it keeps the intelligence layer portable even as the compute layer consolidates around a GPU vendor.

What to Watch Next

The next twelve months will test three things. First, whether BT, Deutsche Telekom and Elisa publish their own live-network call milestones mirroring T-Mobile’s Bellevue announcement as each pilot matures from lab validation to commercial-site testing — and whether the gains replicate under live customer traffic rather than controlled conditions. Second, whether the uplink channel estimation work at Deutsche Telekom translates from 10–15% lab improvements into field results in 2027. Third, whether operators disclose the economics: power per site, utilization of the AI inference capacity, and the total cost of ownership versus purpose-built RAN silicon. Elisa’s distributed deployment and evaluation of commercial AI-RAN services expected in 2027 will be the first real read on whether the platform pays for itself.

The honest assessment is that AI-RAN has crossed from slideware to a named list of operators willing to put their reputations behind joint initiatives, and one of them has made a call on a live commercial network. That is real progress. But the industry has not yet demonstrated production-scale economic superiority, and the competitive landscape is far from settled — Ericsson’s silicon-based AI gains are a live counterargument. For Open RAN operators, the right posture is neither dismissal nor capitulation. It is disciplined engagement: insist on multi-vendor interoperability at the intelligence layer, demand transparent economics before committing spectrum and capital, and treat the GPU-versus-silicon debate as a portfolio decision rather than a religious one. The intelligence is coming to the RAN either way. The question is who controls where it lives.

Sources

Related Posts

The Long Road to the First AI-RAN Call: How a Decade of Disaggregation Led to T-Mobile's Live 5G Milestone

T-Mobile's first live AI-RAN call with Nokia and NVIDIA caps a decade-long shift from proprietary RAN to software-defined, GPU-accelerated networks.

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

Nokia and NVIDIA now count nearly 20 operators on AI-RAN. Five takeaways from T-Mobile's live call, Elisa's field trial and the 2027 pipeline.