The Week AI-RAN Grew Up: T-Mobile’s Live Call, O2’s Paddington Trial, and the $1.1 Billion Question

Paddington Station sees roughly 191,499 passenger journeys on an average day. They arrive in waves — the morning crush, the lunchtime lull, the evening exodus — and the mobile network serving them has to absorb every surge and every silence. It is, in other words, exactly the kind of place where a radio access network either proves it can think on its feet or quietly falls over.

That is where O2, the UK operator owned by Virgin Media O2, chose to run what it calls the country’s first live trial of AI-RAN. The site, one of its so-called “Giga Sites” in London, used artificial intelligence to manage spectrum and capacity in real time. According to O2, the trial delivered a 20% boost in spectral efficiency, improved throughput, and — perhaps most importantly for anyone who has ever tried to send a photo from a packed train platform — a network that stayed stable through sustained periods of high demand.

“Mobile data demand continues to grow, so we need smarter ways to make the most of the spectrum and infrastructure we already have,” O2 chief technology officer Jeanie York said of the trial. “This live trial with Nokia showed how AI can improve network efficiency and help us respond more effectively to changing customer demand.”

That single sentence, buried in a week of AI-RAN announcements, may be the most important thing anyone said about the technology all month. Not because 20% is a spectacular number — it isn’t, in isolation — but because of where the gain came from: software, on infrastructure the operator already owned, in one of the busiest and most unpredictable environments in Britain.

A Week of Proof Points, Not Promises

The last several days have produced an unusual density of AI-RAN news, and the through-line is that the technology has stopped being a lab curiosity. T-Mobile, Nokia, and NVIDIA announced what they describe as the world’s first AI-RAN call, routed through a commercial cell site in Bellevue, Washington. The call ran over T-Mobile’s live production network using a Nokia Massive MIMO radio, Nokia’s 5G anyRAN software, and servers built on NVIDIA’s Aerial RAN Computer platform. It connected to T-Mobile’s existing 5G Standalone core and a commercial smartphone — not a test handset on a lab rig.

Separately, T-Mobile’s chief network officer, Ankur Kapoor, published results from AI-managed beamforming trials with Ericsson in New York City. At peak hours, the operator reported a 16% improvement in downlink user throughput and 12% better spectral efficiency. The AI model in that case ran on Ericsson’s own silicon rather than NVIDIA GPUs — a detail that matters enormously to anyone watching the architectural fight underneath the AI-RAN hype.

And on the market side, SNS Telecom & IT forecast that annual spending on AI-RAN automation and optimisation will reach USD $1.1 billion by 2030, with operators including AT&T, Verizon, Telus, Vodafone, Deutsche Telekom, Swisscom, and Telstra increasing adoption, and players like Rakuten Mobile, China Mobile, STC, and TDC NET already reaching limited levels of highly autonomous operation in live RAN environments for specific use cases.

Individually, each of these is a data point. Together, they sketch the outline of a transition: AI-RAN is moving from the demo circuit to the production network, and the arguments about it are shifting from “does it work?” to “what is it actually worth?”

The Continuum Argument

Kapoor’s framing deserves attention because it cuts against the way vendors like to sell AI-RAN. “AI-RAN is a continuum, not a single product,” he wrote, describing a north star of using AI to make the network smarter and more efficient today while building toward an intelligent, distributed network that can power the next generation of AI applications. He tied the work to 5G-Advanced, which he called “where we can put new intelligence to work today and prove the capabilities that will shape 6G.”

That is a notably sober position for an operator that has also opened an AI-RAN Innovation Center with NVIDIA, Ericsson, and Nokia, and that is developing AI-RAN software with both major RAN vendors simultaneously. T-Mobile is, in effect, hedging. It is running GPU-based AI-RAN experiments with Nokia and NVIDIA on one track, and silicon-native AI features with Ericsson on another. The Bellevue call is the flashy headline; the New York beamforming trial may be the more commercially consequential result, because it required no new accelerator hardware in the radio site.

Industry analyst Gabriel Brown of Omdia captured the significance bluntly, telling Light Reading that the milestone was “a wow moment,” and noting that the gains came from software updates to equipment already in the field. T-Mobile and Ericsson had previously reported close to 10% better spectral efficiency from a separate AI scheduler in May, again with no GPUs involved.

The GPU Question Hiding in Plain Sight

This is where the week’s news becomes genuinely interesting for a technical audience. The AI-RAN narrative pushed hardest by NVIDIA and its partners implies a future in which general-purpose GPUs sit inside — or alongside — the radio site, running AI workloads that share infrastructure with the RAN itself. That vision is what makes AI-RAN strategically exciting: it turns the cell site into a compute platform, capable of hosting third-party applications, edge inference, and new revenue-generating services.

But the most concrete performance numbers reported this week did not come from GPUs. They came from AI models running on existing RAN silicon, delivered as software updates. The Deutsche Telekom figures that circulated alongside the Nokia and NVIDIA announcement — typical uplink throughput gains of 10% to 15% from AI-based channel estimation — came from initial lab tests in Bonn and Oulu, with a field trial still to come. That is a meaningful distinction. Lab gains and live-network gains are different animals, and conflating them is how hype cycles get inflated.

To its credit, T-Mobile has not oversold the Bellevue call. It has not claimed that the AI makes voice calls sound better, or that customers will notice anything at all yet. When Fierce Network asked how many sites were involved in the AI-managed beamforming trials, T-Mobile declined to specify a number. Those are the right questions to ask, and the refusal to answer them is itself informative: the deployment footprint is still small.

What the Money Says

The SNS Telecom & IT forecast adds a useful reality check to the technical exuberance. USD $1.1 billion in annual AI-RAN automation spending by 2030 is real money, but it is a rounding error against global RAN capex, which runs into the tens of billions annually. The research also points to where that spending is likely to concentrate: a shift away from older self-organising network technology toward more open, application-based architectures built around the RAN intelligent controller and service management and orchestration platforms.

In that model, rApps running on the non-real-time RIC layer are emerging as the preferred route to network-wide automation across multi-vendor environments. That is a fundamentally Open RAN-shaped architecture — and it is the clearest signal yet that the AI-RAN story and the Open RAN story are converging, whether or not every vendor wants to admit it.

The economics make sense. Radio access networks account for a large share of mobile infrastructure costs and energy use, which makes them a natural target for operators seeking greater autonomy in network operations. If AI can shave energy consumption, reduce manual optimisation effort, and squeeze more capacity out of existing spectrum, the return does not require a new killer application. It shows up on the opex line.

But it also explains why the GPU question is so contentious. Software-only AI features that run on existing silicon are cheap to deploy and easy to justify. GPU-based AI-RAN platforms require new hardware, new power budgets, and new operational models — and their business case depends on workloads beyond RAN optimisation materialising at scale. That is a bet on the future, not a saving in the present.

Nokia’s Stakes

The week’s news also sharpened the competitive picture around Nokia. The Finnish vendor has been among the most vocal proponents of AI-RAN, and its efforts have accelerated since NVIDIA invested $1 billion into the company in October. That investment has given Nokia both capital and a powerful partner with an obvious interest in seeing its technology deployed by mobile operators worldwide as quickly as possible.

The Bellevue call with T-Mobile and the Paddington trial with O2 are exactly the kind of reference customers Nokia needs. But as Light Reading’s Iain Morris argued this week, Nokia badly needs big AI-RAN deals — and he floated the provocative idea that a conversation with Elon Musk might be a good place to start. The logic: if Starlink moves seriously into mobile network rollout, reportedly exploring access to as many as 80,000 tower sites and filing with the FCC about “towers” and “other structures,” a GPU-powered RAN partnership could reshape the competitive landscape. It is speculative, but it captures something real: AI-RAN’s commercial future depends on landing anchor customers at scale, not on accumulating impressive trial logos.

The Skeptic’s Ledger

So where does that leave an operator trying to decide what to do? The honest answer is that this week’s news strengthens the case for AI-RAN as an operational efficiency play while leaving the case for AI-RAN as a platform play unresolved.

The efficiency case now has live-network evidence: 20% spectral efficiency at Paddington, 16% downlink throughput improvement in New York, and a working end-to-end call on a commercial site in Bellevue. These are not trivial gains, and they arrive through software, which means they can be deployed incrementally, measured, and rolled back if they disappoint. For operators facing flat revenue and rising traffic, that is a compelling proposition.

The platform case — the vision of cell sites as general-purpose AI compute nodes hosting third-party workloads — remains largely prospective. The GSMA Intelligence podcast this week touched on exactly this gap, noting that while AI-RAN efficiency proof points are accumulating, the monetization angle has been “sort of under the radar.” Elisa’s multi-workload approach in Europe, which extends beyond RAN to voice, data, and video analytics, is one of the few examples pointing toward a broader model. Deutsche Telekom’s physical AI and industrial robotics work is another. But these are early.

There is also the question of architectural lock-in. Operators that adopt vendor-proprietary AI features running on vendor silicon get efficiency gains today but may find themselves with less flexibility tomorrow. Operators that wait for open, multi-vendor AI-RAN platforms may get a more durable architecture but risk falling behind on near-term performance. T-Mobile’s dual-track strategy — Ericsson silicon on one side, Nokia and NVIDIA GPUs on the other — is a reasonable hedge, but it is not a strategy most operators can afford to replicate at scale.

What to Watch Next

The coming months will test whether this week was an inflection point or a well-timed cluster of announcements. Three things are worth watching.

First, the field trial results. Deutsche Telekom’s lab figures need to be validated in live networks, and the industry should hold vendors to that standard. Lab gains of 10% to 15% in uplink throughput are promising; live-network gains at scale would be transformative.

Second, the rApp ecosystem. The SNS forecast’s emphasis on the RIC and SMO layer suggests that the real battle will be fought over who controls the application platform, not who supplies the radio. If rApps become the dominant route to automation across multi-vendor environments, the Open RAN architecture wins by default — even in networks that are not formally branded as Open RAN.

Third, the anchor deals. Nokia’s AI-RAN story is credible, but credibility is not the same as commercial traction. A large-scale commitment from a tier-one operator — or a disruptive entrant like Starlink — would change the calculus for everyone else.

For now, the most useful way to read this week is as a maturation moment. AI-RAN has produced its first genuine live-network proof points, its first credible market forecast, and its first serious architectural debate. The hype has not disappeared, but it has acquired a counterweight: numbers, measured in production networks, on real traffic, at real cell sites. That is how technologies grow up.

Sources

• The week in AI-native telco (17-23 Oct 2026), TelecomTV
• ‘This is a wow moment’: T-Mobile has made the world’s first AI-RAN call, TechRadar
• AI-RAN automation spending to hit USD $1.1bn by 2030, IT Brief UK
• Nokia badly needs big AI-RAN deals – Musk would be a great start, Light Reading
• What’s up with… Verizon, O2, Tata Communications, The AI-Native Telco, TelecomTV
• Episode 22: NVIDIA’s Telecom AI Push, Europe’s Satellite Alliance and the 6G Spectrum Race, GSMA Intelligence