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

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

On October 21, 2026, T-Mobile completed what NVIDIA, Nokia, and the operator itself called the world’s first AI-RAN call on a live commercial network in Bellevue, Washington. The call ran over T-Mobile’s production 5G Standalone core, using commercial massive MIMO radios, Nokia’s CUDA-accelerated anyRAN software, and NVIDIA’s Aerial RAN Computer (ARC) platform. It was, in the words of Omdia analyst Gabriel Brown, a “wow moment” — not because a phone call is remarkable, but because the base station making it was running AI workloads on general-purpose accelerated computing rather than purpose-built silicon.

That single call did not happen in a vacuum. It is the latest marker on a road the telecom industry has been walking for more than a decade: the slow, contentious, and often disappointing journey from fixed-function radio hardware to software-defined, disaggregated, and now AI-native networks. Understanding why this moment matters — and why it is still more of a beginning than an arrival — requires looking back at how the RAN got here.

From ASICs to Software: The Open RAN Prelude

For most of the 2G, 3G, and 4G eras, the radio access network was a monument to proprietary integration. Base stations were built on custom ASICs and DSPs, running vendor-specific software tightly coupled to vendor-specific hardware. Flexibility was low, multi-vendor deployment was nearly impossible, and innovation moved at the pace of hardware refresh cycles rather than software releases.

Centralized RAN (C-RAN) in the early 4G and 5G densification years introduced a degree of centralization, but it did not break the fundamental vendor lock-in. The real shift came with virtualization and Open RAN, which argued that radio software could run on standard x86 or Arm servers and that interfaces between radio units and baseband units could be opened up. That movement spent the better part of a decade proving, often painfully, that disaggregated, multi-vendor RAN was technically feasible. It also proved that feasibility alone does not guarantee economics or performance parity.

The industry’s collective experience with Open RAN — the promises, the delays, the incremental commercial wins — set the stage for the next question. If radio software could run on generic compute, could that generic compute be the same GPU-accelerated infrastructure that already dominates AI training and inference? That is the question AI-RAN attempts to answer.

The AI-RAN Hypothesis: Putting Idle Compute to Work

The core premise of AI-RAN is deceptively simple. Traditional RAN infrastructure is dimensioned for peak load, which means a significant amount of compute sits idle most of the time. If that same infrastructure could run AI inference workloads — either AI-for-RAN algorithms that improve radio performance or RAN-for-AI services that generate new revenue — operators could extract more value from the same footprint.

As one industry analysis put it, AI-RAN is best understood as the next step in the same direction as vRAN and Open RAN, not a break from it. 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 now also run AI workloads, turning idle compute capacity at cell sites into a usable resource rather than overhead that only exists to handle peak network load.

That shift builds directly on groundwork laid by 5G Advanced and 3GPP Release 18, which already pushed more intelligence and flexibility into the radio layer. It also builds on a hardware trend that has been accelerating outside telecom: the rise of GPU-accelerated computing as the default platform for AI. NVIDIA’s argument — that accelerated computing now belongs in places that used to be considered pure CPU territory — has been tested in data centers and is now being tested in the RAN.

The Partnership That Made the Call Possible

The Bellevue call was the product of a specific partnership timeline. In September 2024, T-Mobile announced an AI-RAN Innovation Center in Bellevue to explore how AI and RAN could benefit each other. In 2025, Nokia and NVIDIA formed a partnership to add NVIDIA-powered, commercial-grade AI-RAN products to Nokia’s radio access network portfolio, and T-Mobile was named as an operator working with both companies to integrate AI-RAN technologies into its 6G development process. Earlier in 2026, Nokia announced successful lab testing of its anyRAN software on NVIDIA’s GPU-accelerated platform with T-Mobile and other operators.

The October 2026 call moved that validation from the lab into a live commercial network environment — real spectrum, commercial infrastructure, a commercial 5G core, and a commercial smartphone. Nokia described it as not just the world’s first AI-RAN call but the world’s first software-defined RAN call using accelerated computing, a moment when RAN reached parity with the cloud and software worlds.

That framing is ambitious, but it captures the significance. For four decades, the RAN was the part of the network most resistant to the software-defined revolution that transformed computing and the core network. The Bellevue call is the first credible evidence that the RAN can join that revolution.

Beyond One Call: A Wave of Operator Commitments

The T-Mobile milestone was announced at NVIDIA’s GTC event in Berlin, and it came alongside a broader wave of operator commitments. Deutsche Telekom has validated Nokia and NVIDIA’s commercial AI-RAN platform in Bonn and Oulu labs, including AI-enhanced uplink channel estimation that showed typical uplink throughput improvements of 10–15%, with plans to move to field trials in 2027. Elisa completed Europe’s first field trial of AI-RAN in Finland, running voice, mobile data, and vision AI together on a single accelerated computing platform over 5G. BT is preparing live AI-RAN trials in Birmingham, England, in 2027. Orange is testing AI-RAN in lab and real network conditions and introducing new voice intelligence services for enterprises and emergency responders.

Globally, nearly 20 telecom operators have publicly announced AI-RAN collaborations using Nokia anyRAN software and NVIDIA ARC. New infrastructure deployments were announced by CRA in the Czech Republic, etisalat in the UAE, Silknet in Georgia, and TIM in Italy, while existing programs expanded at Cassava in Africa, Deutsche Telekom in Germany, Fastweb + Vodafone in Italy, Kazakhtelecom in Central Asia, and Orion Telecom in Serbia. Separately, SoftBank and Nokia are targeting 2027 for a commercial AI-RAN trial in Japan, combining Nokia’s AI-native anyRAN Cloud RAN software with SoftBank’s AITRAS platform and NVIDIA’s ARC hardware.

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. It suggests the AI-RAN category has moved from a slide-deck pitch to a list of operators willing to put their name next to a joint initiative — even if most have not yet disclosed trial results the way Elisa has.

The Ericsson Counterpoint: AI in RAN Without GPUs

The same week as the T-Mobile AI-RAN call, T-Mobile also revealed a different AI achievement in its live 5G-Advanced network in New York City with Ericsson. In a large-scale commercial trial, T-Mobile implemented Ericsson’s AI-managed beamforming, which uses an AI model trained on T-Mobile’s network data running directly on Ericsson’s RAN compute and silicon to predict demand and dynamically shift sounding reference signal (SRS) resources. In activated New York City sites, T-Mobile clocked a 16% improvement in downlink throughput, 12% better spectral efficiency, 10% increase in downlink data volume, and 50% more advanced beamforming using full-band SRS during peak hours. That followed an earlier AI-native scheduler with link adaptation trial that showed 10% better spectral efficiency and a 15% increase in downlink throughput.

The contrast is instructive. Ericsson’s approach delivers measurable AI gains without GPUs, using custom silicon and software updates to deployed equipment. NVIDIA and Nokia’s approach bets that GPU-accelerated platforms will deliver larger gains and, crucially, enable new AI workloads alongside the RAN. Both are real. Both are being tested on live networks. The industry has not yet resolved which economics will win at scale.

The Open RAN Question: Convergence or Competition?

The relationship between AI-RAN and Open RAN is one of the most consequential questions for operators. AI-RAN, as implemented by Nokia and NVIDIA, is not inherently open in the Open RAN sense. It is a tightly integrated stack — Nokia’s anyRAN software running on NVIDIA’s ARC hardware — and it does not require multi-vendor disaggregation. In that sense, it could be seen as a competitive alternative to Open RAN’s vendor-neutral vision.

But the two are not mutually exclusive. Open RAN proved that radio software could run on generic infrastructure. AI-RAN takes that one step further by asking whether that generic infrastructure can be GPU-based accelerated computing rather than a separate category of telecom-specific merchant silicon. If AI-RAN succeeds, it could strengthen the case for software-defined RAN generally, even if the first commercial implementations are single-vendor. The industry’s history suggests that tightly integrated first implementations often pave the way for later disaggregation — provided the interfaces and economics allow it.

There is also a competitive dynamic to watch. Ericsson will face growing pressure from investors and analysts to clarify whether its custom-silicon RAN strategy can match NVIDIA and Nokia’s claimed efficiency gains, 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 their core differentiator. The next two years of trials will be as much a commercial argument as a technical one.

What Comes Next: 2027 and the Field Trial Wave

The Bellevue call is a milestone, not a finish line. The next phase of field trials will extend to new AI-for-RAN algorithms and AI workloads running alongside radio access networks on a shared accelerated computing platform. BT will test whether gains reported in Nokia’s wider AI-RAN platform testing — including spectral-efficiency improvements of more than 20% — can be replicated with live customer traffic. Deutsche Telekom will move from lab validation to field trials. Elisa expects distributed deployment and evaluation of customer use cases and commercial AI-RAN services in 2027. SoftBank and Nokia target 2027 for their commercial trial in Japan.

Nokia’s stated ambition is a path to 2x spectral efficiency, which would put more capacity from the same spectrum at the heart of what comes next. Whether that ambition survives contact with real-world traffic, power constraints, and cost models is the open question. The industry has been here before with Open RAN, where lab results often outpaced commercial reality. The difference this time is that the AI workload itself may justify the compute investment in ways that pure RAN virtualization sometimes struggled to do.

The Historical Frame: A Software-Defined RAN at Last

Seen from a decade’s distance, the Bellevue call looks less like a sudden breakthrough and more like the convergence of two long trends. The first is the telecom industry’s slow, fitful march toward software-defined networks, which began with core virtualization, moved through Open RAN, and has now reached the radio itself. The second is the broader computing industry’s shift to accelerated, GPU-based infrastructure for AI, which is now pushing into every corner of the network.

For four decades, the RAN was the exception to the software-defined rule. It was the part of the network where proprietary hardware and embedded software reigned, where innovation moved at hardware speed, and where vendor lock-in was the default. The first AI-RAN call on a live commercial network does not end that era overnight. But it does provide the first real evidence that the RAN can be software-defined, accelerated, and AI-native — and that operators are willing to put their names and their spectrum behind the bet.

The next twelve months will determine whether that bet pays off. The trials are scheduled, the operators are named, and the hardware is in the field. What remains to be seen is whether the economics, the performance, and the ecosystem can sustain the momentum that one call in Bellevue has generated. History suggests the answer will be neither as fast nor as simple as the hype implies — but it also suggests that the direction of travel is now clear.

Sources

Related Posts

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.

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

A skeptical exec's Q&A on T-Mobile's live AI-RAN call, Elisa's European field trial, and what Nokia and NVIDIA's 2027 commercial timeline really means.