Nokia’s 19-Customer AI-RAN Book Is Impressive—Now Show Us the Money

Nokia has spent the better part of two years telling anyone who would listen that AI-RAN is real, that it is commercial, and that it is happening now. This week, with its Q3 FY 2026 results landing alongside a fresh wave of NVIDIA-and-Nokia announcements out of GTC Berlin, we finally have enough evidence to grade that claim. The verdict is uncomfortable for both the hype merchants and the skeptics: the technology is genuinely further along than the cynics said it would be, and the business case is genuinely further behind than the believers want to admit.

Let me be precise about what actually happened, because the headlines have a way of blurring the important lines.

What Nokia and NVIDIA Actually Delivered

On October 21, at the opening day of GTC Berlin, NVIDIA and Nokia said nearly 20 operators worldwide have now announced AI-RAN collaborations built on Nokia’s anyRAN software and NVIDIA’s Aerial RAN Computer (ARC). Nokia’s own Q3 disclosure puts a firmer number on it: 19 customers engaged in its AI-RAN program. That is not a pilot here and a lab there. That is a customer book.

The flagship proof point is T-Mobile, which completed what the companies are calling the world’s first AI-RAN call on a live commercial network in Bellevue, Washington. Read the technical detail carefully, because it matters: the call ran on commercial ARC servers, Nokia’s CUDA-accelerated anyRAN software, and Massive MIMO radios, connected to an existing 5G Standalone core and a commercial smartphone. No lab rig. No prototype handset. A real call on a real network with real radios.

Around that anchor, the field trials stack up. Elisa completed Europe’s first AI-RAN field trial, running voice, mobile data and vision AI workloads simultaneously on a single accelerated computing platform over 5G in Helsinki. Deutsche Telekom validated the Nokia-NVIDIA commercial platform in Bonn and Oulu labs, including AI-enhanced uplink channel estimation that showed typical uplink throughput improvements of 10–15%, with larger gains under more complex radio conditions. BT is preparing live AI-RAN trials in Birmingham in 2027. Orange is testing in both lab and real network conditions. Nokia has reported spectral-efficiency improvements above 20% in live environments.

And the vendor is not standing still on hardware. NVIDIA introduced new Aerial RAN Computer configurations aimed at high-capacity sites, and nine EMEA operators are building and expanding NVIDIA-powered infrastructure across centralized and distributed locations. Operators and software partners are also picking up NVIDIA’s Nemotron open models to build agentic services for autonomous network operations.

That is a serious body of work. Anyone still writing off AI-RAN as a slideware fantasy is not paying attention.

The Q3 Tell: Demand Outrunning Supply Is a Nice Problem Until It Isn’t

The results, though, tell a messier story. Nokia’s Radio Networks quarter benefited from deliveries pulled forward from Q4 FY 2026, while Core Software faced server-related project delays. Read that again: the AI-RAN narrative is surging, but the financials are being shaped by timing shifts and supply constraints rather than by a clean, linear ramp of AI-RAN revenue.

Nokia CEO Justin Hotard’s framing is telling. “Together with T-Mobile and NVIDIA, we completed the world’s first AI-RAN call in a live operator network in line with the commitment we announced last year,” he said, adding that “as we move towards commercial deployment, the features and spectral efficiency delivered by our platform are uniquely differentiated in the market.”

That is a commitment-keeping story, not yet a revenue story. And the Futurum analysis of the quarter lands the key caveat squarely: operator adoption will depend on whether capacity gains and software flexibility justify deployment costs across commercial networks. Translation—the 20%-plus spectral efficiency numbers are impressive in a trial, but a CFO needs to see them hold up across a whole network, across seasons, across traffic mixes, before signing a multi-year capex line.

This is the oldest problem in RAN economics, and AI-RAN has not solved it. It has merely restated it in GPUs.

The Skeptic’s Case Is Stronger Than It Looks

The most intellectually honest take on the Bellevue milestone came not from a vendor press release but from a geopolitics-of-6G analysis published on October 23. Its assessment: the achievement “establishes an important commercial-network integration milestone, but does not yet demonstrate production-scale economic superiority, sustained multi-workload performance or autonomous 6G functionality.”

That is exactly right, and it is the sentence every operator’s strategy team should pin to the wall. A single successful call proves integration. It does not prove that a GPU-accelerated RAN beats a purpose-built one on total cost of ownership at scale, under sustained multi-workload load, over a five-year depreciation cycle. Those are different questions, and the industry keeps answering the easy one while implying it has answered the hard one.

Consider what is genuinely unproven. Elisa ran voice, mobile data and vision AI together on one platform—a real demonstration of workload convergence, but in a field trial, with distributed deployment and commercial AI-RAN service evaluation explicitly pushed to 2027. BT’s live trial is a 2027 event. Deutsche Telekom’s move from lab validation to field trials is also a 2027 plan. Even T-Mobile’s live call, for all its significance, is a single call on a commercial network, not a production traffic regime.

So the honest timeline is this: 2026 is the year of the live proof point. 2027 is the year of the field trial at scale. Commercial deployment at meaningful volume is a 2027-and-beyond conversation. Anyone claiming otherwise is selling something.

The Strategic Question Nobody Wants to Answer

Here is where I part company with both camps. The skeptics are right that the economics are unproven. But they are wrong to conclude that means AI-RAN is a fad. The believers are right that the technology works. But they are wrong to pretend the customer has already decided.

The real story in this week’s news is not the call. It is the shape of the buying decision that Nokia is quietly engineering. Running anyRAN software on NVIDIA’s accelerated computing platform creates a path for adding radio capabilities through software, shifting part of the purchasing decision toward continuing performance improvements. That is the sentence that should worry Nokia’s competitors and excite operators simultaneously.

If RAN capability becomes a software upgrade on general-purpose accelerated compute, then the value migrates from the radio hardware to the software stack and the silicon underneath it. That is a profound shift for an industry that has spent decades competing on bespoke ASICs and radio engineering. It is also precisely why the AI-RAN fight is really a fight over where intelligence lives—on the GPU, in the software layer, or distributed across both.

For Open RAN operators, this cuts both ways. On one hand, a software-first RAN aligns beautifully with the disaggregated, multi-vendor vision that Open RAN has always promised. On the other, the Nokia-NVIDIA pairing is a vertically integrated bundle dressed in open clothing. Nokia’s anyRAN plus NVIDIA’s ARC is a two-vendor stack, and operators who adopt it wholesale may find themselves trading one form of lock-in for another—just with better spectral efficiency.

That tension is the real story of 2026, and it is not going away.

What Would Actually Change My Mind

I am not a reflexive AI-RAN doomer. I am a demanding customer. Here is what would move me from cautious to convinced, and operators should demand the same evidence before they commit capex.

First, sustained multi-workload performance in production, not a field trial. Elisa’s Helsinki demonstration is a start, but I want to see a commercial network running voice, data and AI inference simultaneously for months, with published reliability and capacity data, not a press release describing a successful afternoon.

Second, a credible total-cost-of-ownership model. The 10–15% uplink throughput gains from AI channel estimation and the 20%-plus spectral efficiency improvements are real and meaningful. But GPUs are expensive, power-hungry, and depreciate on a different curve than radios. Someone needs to publish the math that shows those efficiency gains outrun the cost of the accelerated compute. So far, nobody has.

Third, evidence that the software-upgrade path actually delivers. The promise of anyRAN on ARC is that operators buy the platform once and add capabilities through software. If that holds, the economics change fundamentally. If it turns into a treadmill of hardware refreshes to keep pace with the software, the promise is hollow. Watch the 2027 field trials for this specifically.

Fourth, multi-vendor reality. A two-vendor stack is not Open RAN. It is a closed bundle with open interfaces. Operators should insist on genuine interoperability—competing software stacks on competing accelerators—before they declare victory.

The Closing Take

Nokia deserves real credit. Nineteen customers, a live commercial call, a credible 2027 field-trial roadmap, and spectral efficiency gains above 20% in live environments is a stronger position than almost anyone predicted two years ago when the NVIDIA partnership was announced. Hotard is right that the platform is differentiated. The company kept the commitment it made last year, and that matters in an industry that has heard a lot of AI-RAN promises.

But let us not confuse momentum with money. The Q3 results show demand testing supply capacity, deliveries pulled forward, and project delays—the financial signature of a business still finding its footing, not one that has arrived. The 2027 trials at BT, Deutsche Telekom, Elisa and Orange will be the real referendum. If the capacity gains and software flexibility justify the deployment costs across commercial networks, AI-RAN becomes the defining architecture of the 6G era. If they do not, we will have built a very impressive demonstration platform and called it a market.

The technology is no longer the question. The economics are. And the industry has spent two years celebrating the first while dodging the second. It is time to show the money—or stop pretending the deal is done.

Sources

• NVIDIA, Nokia and Telecom Operators Advance AI-RAN and Distributed Intelligence — The Fast Mode
• The week in AI-native telco (17-23 Oct 2026) — TelecomTV
• AI-RAN and the Geopolitics of 6G — Debug Lies News
• Nokia, Nvidia accelerate AI-RAN, physical AI commercial deployment — Computer Weekly
• Nokia Q3 FY 2026: AI Network Demand Tests Supply Capacity — Futurum
• The week in telecom: European telcos take on satellite market, Nvidia rekindles AI RAN hopes — Telecompaper
• AI Infrastructure Race: 9 Models in 5 Weeks — Tech Insider