NVIDIA and Nokia used the opening day of GTC Berlin on October 21, 2026, to announce that cell towers and telecom data centers are turning into AI infrastructure. T-Mobile, BT, Deutsche Telekom, Elisa, and Orange are the named operators now running or piloting what NVIDIA calls AI-RAN, a way of running artificial intelligence workloads directly on the radio access network hardware that carries phone calls and mobile data.
The headline claim from NVIDIA: T-Mobile has completed what the company describes as the world’s first AI-RAN call on a commercial network. Finland’s Elisa says it ran Europe’s first field trial combining voice, mobile data, and vision AI on a single accelerated computing platform. Nearly 20 telecom operators worldwide have now publicly announced AI-RAN collaborations built on Nokia’s anyRAN software running on NVIDIA’s Aerial RAN Computer, according to the companies.
For an industry that has spent two decades treating radio hardware and AI compute as separate worlds, that is a notable shift. This article breaks down what was actually confirmed at GTC Berlin, what remains unproven marketing language, and what it could mean for telecom operators, NVIDIA’s data center business, and the networks millions of people use every day.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Is AI-RAN and Why It Matters Now
RAN stands for radio access network, the layer of equipment, antennas, and processing hardware that connects a phone to the rest of a mobile network. For most of the industry’s history, that hardware ran fixed-function signal processing and nothing else. AI-RAN flips that assumption. It puts GPU-accelerated computing inside the radio network itself, so the same hardware that handles signal processing can also run AI models, from network optimization to computer vision to generative AI inference for enterprise customers.
NVIDIA’s pitch, repeated across its announcement, is that telecom operators already own three things AI companies desperately want: land, power, and connectivity spread across tens of thousands of sites. Turning that physical footprint into usable AI capacity, rather than building new data centers from scratch, is the core argument behind the GTC Berlin news. The concept echoes mobile edge computing, which has pushed processing power closer to cell sites for years, except AI-RAN adds GPU-class acceleration to that same edge. Jensen Huang, NVIDIA’s founder and CEO, put it plainly in the announcement: “Telecom networks were built to carry the world’s conversations. Now, they will also carry the world’s intelligence.”
Inside the GTC Berlin Announcement: Who’s Involved
The announcement names seven companies directly: NVIDIA, Nokia, T-Mobile, BT, Deutsche Telekom, Elisa, and Orange. NVIDIA supplies the compute layer through its Aerial RAN Computer platform and CUDA-accelerated RAN software. Nokia supplies the radio software layer through a product called anyRAN, designed to run across Nokia’s existing radio portfolio while plugging into NVIDIA’s accelerated hardware. According to a GlobeNewswire release detailing Nokia’s side of the announcement, Nokia is bringing that CUDA-accelerated anyRAN software to live trials at BT, Deutsche Telekom, Elisa, and Orange.
NVIDIA frames the rollout around a broader architectural idea: operators are adding accelerated computing across three tiers of their networks, centralized data centers, metro sites, and individual cell sites, so AI models and applications can run physically closer to the people, businesses, and machines that need them. That is the “distributed intelligence platform” language in NVIDIA’s own headline for the announcement, and it is meant to describe infrastructure that previously only carried traffic now also processing it.
T-Mobile Claims the World’s First AI-RAN Call
The most concrete technical claim in the announcement belongs to T-Mobile. NVIDIA says the carrier completed the world’s first AI-RAN call on a commercial network, and the setup details are specific enough to evaluate. The call ran on commercial Aerial RAN Computer servers using Nokia’s CUDA-accelerated RAN software, relied on massive multiple-input multiple-output (massive MIMO) radios, connected into T-Mobile’s existing 5G Standalone core network, and used an ordinary commercial smartphone rather than a lab prototype.
That combination matters because it signals the test was not a closed-lab demo running on experimental hardware. It ran on infrastructure already wired into a live 5G Standalone network, with off-the-shelf handsets on the other end. NVIDIA has not published independent performance data from the call beyond calling it a first, so readers should treat the claim as a milestone of architecture, not yet a benchmark of speed or capacity gains for T-Mobile customers.
Elisa Runs Europe’s First AI-RAN Field Trial
Finnish operator Elisa gets its own milestone in the announcement: Europe’s first field trial of AI-RAN, running voice, mobile data, and vision AI together on one accelerated computing platform over a live 5G network. Running three distinct workload types, two network-native and one AI-native, on shared hardware is the point NVIDIA wants to make with the Elisa trial specifically. It is meant to demonstrate that AI-RAN hardware does not just carry the AI workload alongside network traffic, it actively shares capacity with it in real time.
Elisa has a reputation in European telecom circles as an early mover on network virtualization and 5G Standalone deployment, which makes it a logical operator for NVIDIA and Nokia to showcase first. Neither company has published a specific commercial launch date for Elisa’s AI-RAN service, and this story treats that detail as unconfirmed rather than speculating on a timeline.
BT, Deutsche Telekom and Orange Join the Pilot Wave
Beyond the T-Mobile and Elisa milestones, NVIDIA and Nokia confirm that BT, Deutsche Telekom, and Elisa have each announced AI-RAN pilots on live networks using the same Nokia anyRAN software running on NVIDIA’s Aerial RAN Computer. Orange is named as a fourth operator where Nokia is bringing the software for trials. That is four major European operators, spanning the UK, Germany, Finland, and France, lined up behind a single technology stack within one announcement window.
Deutsche Telekom’s involvement comes with an actual performance figure, which is rare for this kind of multi-party announcement. According to reports on the announcement, Deutsche Telekom’s AI-RAN tests produced uplink throughput improvements of 10 to 15 percent. That is a modest but real number, and it is the only quantified performance claim tied to a named operator in the entire announcement, which makes it worth paying closer attention to than the broader distributed intelligence framing.
The Technology Stack Behind the Announcement
Three named products make up the technical core of the announcement. NVIDIA Aerial RAN Computer is the hardware and software platform designed to run radio signal processing and AI inference on the same GPU-accelerated servers. Nokia anyRAN is the software layer that lets Nokia’s existing radio hardware portfolio run on different kinds of underlying compute, including NVIDIA’s accelerated servers, rather than requiring Nokia’s traditional proprietary hardware. NVIDIA CUDA-accelerated RAN software is the piece that actually performs the signal processing workload using NVIDIA’s CUDA programming platform instead of fixed-function radio chips.
Jensen Huang described the technical approach directly in comments tied to the announcement: “With Nokia, we’re extending accelerated computing into the radio network itself, creating the foundation for AI-native 5G.” Nokia’s own statement, posted on its newsroom announcement, framed the work similarly, describing a joint effort to bring AI into both the radio network and network operations more broadly. Neither company has published detailed specifications, GPU counts, or pricing for Aerial RAN Computer deployments, so this piece does not estimate those figures. The GPU memory demands of that kind of accelerated compute are already straining suppliers like Micron, whose HBM4 sales jumped 331% in the run-up to this same October 21 window.
Nearly 20 Operators Are Already on Board
The scale number attached to this announcement is nearly 20 telecom operators worldwide that have publicly announced AI-RAN collaborations using Nokia anyRAN software and NVIDIA’s Aerial RAN Computer platform. That figure includes the named operators in the GTC Berlin announcement plus others that have signed on in earlier NVIDIA AI-RAN initiatives. It is a meaningful number for an infrastructure shift this early, since telecom operators are typically slow, risk-averse buyers that run multi-year procurement cycles before committing capital to new radio hardware.
What the nearly 20 figure does not tell readers is how many of those operators have moved past pilot status into paid commercial deployment. Based on what has been confirmed so far, T-Mobile and Elisa are the furthest along, with completed technical milestones. BT, Deutsche Telekom, and Orange are described as running pilots or trials, which is an earlier stage than a live commercial rollout serving paying customers at scale.
From Cell Towers to Data Centers: The Architecture Shift
NVIDIA’s distributed intelligence platform framing rests on a three-tier architecture: centralized data centers, metro-level sites, and individual cell sites, each running accelerated computing that previously only existed in the data center tier. That matters because it changes where AI inference can physically happen. A model running at a cell site can respond with the kind of low latency that a centralized cloud data center hundreds of miles away cannot match, which opens use cases like real-time computer vision for industrial robotics or augmented reality that depend on near-instant response times.
Huang leaned into this framing with a specific long-term forecast tied to the announcement: “Every single telco in the future, and I predict this, every single telco in the future, with 100,000 sites representing some 100 gigawatts of distributed telco base stations and data centers, will in the future evolve into a distributed AI network.” That is a forward-looking prediction from NVIDIA’s CEO, not a confirmed industry-wide commitment, and it should be read as NVIDIA’s own sales narrative for why operators should buy into the Aerial RAN Computer platform.
How RAN Got Here: A Short History
Radio access networks spent the 2G through 4G era running almost entirely on proprietary, fixed-function hardware built by a handful of vendors, with little room for third-party software or general-purpose compute. The 5G era brought two major shifts: virtualized RAN (vRAN), which moved baseband processing onto standard server hardware, and Open RAN, an industry push to disaggregate radio hardware and software so operators could mix vendors instead of locking into one supplier’s full stack.
AI-RAN is best understood as the next step in that same direction, 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 inference 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 the 3GPP’s Release 18 standard, which already pushed more intelligence and flexibility into the radio layer before AI-RAN arrived.
RAN Evolution: Traditional, Virtualized and AI-Native
| RAN Approach | Hardware Type | Software Flexibility | Native AI Capability | Example |
|---|---|---|---|---|
| Fixed-function RAN (2G-4G era) | Proprietary ASICs/DSPs | Low, single-vendor | None | Legacy macro cell sites |
| Centralized RAN (C-RAN) | Centralized baseband units | Moderate | None | Early 4G/5G densification |
| vRAN / Open RAN | Standard x86/Arm servers | High, multi-vendor | Limited, rule-based only | Multi-vendor 5G rollouts |
| AI-RAN (Aerial RAN Computer + anyRAN) | GPU-accelerated servers | High | Native, GPU-accelerated | T-Mobile, Elisa trials, Oct. 2026 |
| AI-native 6G RAN (emerging) | Unified AI-RAN platforms | Software-defined | Built in from design | Early industry roadmap work |
Who’s Involved: Companies and Confirmed Milestones
| Company | Role | Confirmed Milestone |
|---|---|---|
| NVIDIA | Compute platform provider | Aerial RAN Computer, CUDA-accelerated RAN software |
| Nokia | Radio software provider | anyRAN software deployed across trials |
| T-Mobile (US) | Operator | World’s first AI-RAN call on a commercial network |
| Elisa (Finland) | Operator | Europe’s first AI-RAN field trial, voice + data + vision AI |
| BT (UK) | Operator | Announced AI-RAN pilot on a live network |
| Deutsche Telekom (Germany) | Operator | AI-RAN pilot; 10-15% uplink throughput gain reported |
| Orange (France) | Operator | Nokia bringing anyRAN software for trials |
Competitive Landscape: Ericsson, Samsung and Open RAN
NVIDIA and Nokia are not the only companies chasing an AI-native radio network. Ericsson is betting on AI traffic management in markets like India, where 5G subscriber bases have grown fast, and the Swedish equipment maker competes directly with Nokia across most of the world’s 5G markets. Samsung also sells RAN equipment and chips into the same competitive set, and both companies have pushed their own flavors of AI-assisted network management for several years. A Telecoms.com analysis of the announcement noted that European operator enthusiasm for AI-RAN is a key part of NVIDIA’s pitch to the market.
What sets the NVIDIA-Nokia announcement apart is the specific claim of GPU-class accelerated compute running inside the radio network itself, rather than AI software layered on top of conventional RAN hardware for tasks like traffic prediction or energy savings. None of NVIDIA’s rivals have announced a comparable multi-operator deployment running on NVIDIA-class GPU acceleration at the cell site level, which gives NVIDIA and Nokia a head start in this specific technical approach, even though the broader AI-for-telecom race has multiple players already competing.
The Open RAN movement adds another wrinkle. Open RAN was built around the idea of disaggregating hardware and software so operators are not locked into one vendor’s full stack. NVIDIA’s Aerial RAN Computer platform paired with Nokia’s anyRAN software fits that disaggregation model in principle, since anyRAN is designed to run across Nokia’s radio portfolio on different underlying compute. Whether that openness extends to letting operators swap in GPUs from other vendors remains an open question that neither company has addressed publicly.
Market Impact and What Comes Next
For NVIDIA, the telecom push extends its data center business into a sector that has historically bought very little GPU compute. NVIDIA has spent the past two years building partnerships across cloud providers and enterprise customers, and AI-RAN represents a new category of buyer: telecom operators with physical infrastructure in tens of thousands of locations worldwide, a footprint that cloud hyperscalers do not have at the same density. That same GPU appetite is already visible elsewhere in NVIDIA’s business, including the 1.44-million-GPU buildout tied to SpaceX’s AI pivot and the recently launched RTX Spark desktop AI platform, both signs of how broadly NVIDIA is spreading its hardware across industries beyond traditional cloud computing.
For Nokia, the anyRAN software push is a bet on staying relevant as radio hardware itself becomes commoditized and the real value shifts toward software and compute partnerships. Nokia has leaned into multi-vendor, multi-compute flexibility with anyRAN rather than defending a closed hardware stack, a strategic choice that puts it in a cooperative position with NVIDIA instead of a competitive one.
For operators like T-Mobile, BT, Deutsche Telekom, Elisa, and Orange, the calculation is different again. These are capital-intensive businesses already spending heavily on 5G Standalone rollouts, and adding GPU-accelerated compute to that buildout is a real cost even before any new revenue from AI services materializes. Deutsche Telekom’s reported 10 to 15 percent uplink throughput gain is the clearest early signal of direct network benefit from the investment, separate from any future AI service revenue the operators might generate by renting out spare compute capacity to enterprise customers.
What Industry Leaders Are Saying
NVIDIA CEO Jensen Huang has been the most vocal figure tied to this announcement, framing it both as a near-term technical milestone and a long-term industry shift. On the scale of the opportunity, he said: “Leading global operators are turning their land, power and connectivity into a distributed AI infrastructure, bringing intelligence closer to the industries, people and robots that use it.”
“Telecom networks were built to carry the world’s conversations. Now, they will also carry the world’s intelligence.”
Jensen Huang, Founder and CEO, NVIDIA
Huang has made this argument before. In comments tied to NVIDIA’s earlier work with telecom partners on AI-native wireless networks, he described the underlying technical approach in blunt terms: “We’ve fused signal processing and AI.” That same philosophy carried through into the broader 6G-focused push NVIDIA has pursued with telecom industry leaders, where Huang said: “Next-generation wireless networks will be revolutionary, and we have an unprecedented opportunity to ensure AI is woven in from the start.”
Nokia’s own public statement on the announcement struck a similar note of partnership rather than competition, describing the company’s role as bringing a shared vision into reality alongside NVIDIA and the operators involved: “Together with Nvidia and leading operators, Nokia is turning this vision into reality by bringing AI into the radio network and across network operations.” A Yahoo Finance report on Huang’s remarks captured the scale of his long-term forecast for telco infrastructure becoming AI capacity.
Predictions: Where AI-RAN Goes From Here
- More operators beyond the nearly 20 already announced will likely join the Nokia anyRAN and NVIDIA Aerial RAN Computer ecosystem over the next year, since the GTC Berlin announcement functions as a public recruitment pitch as much as a product launch.
- Expect competitors, particularly Ericsson and Samsung, to respond with their own GPU-accelerated or AI-native RAN announcements rather than cede the category to NVIDIA and Nokia without a direct counter-offer.
- Deutsche Telekom’s 10 to 15 percent uplink throughput gain will likely become the reference data point other operators cite when evaluating whether AI-RAN investment is worth the capital cost, since it is currently the only quantified performance figure tied to a named operator.
- Pilot programs at BT, Elisa, and Orange are likely to generate more specific performance data over the next several months as trials mature, which will give analysts a clearer picture than the broad distributed intelligence platform framing offers today.
- The commercial case for AI-RAN will increasingly depend on whether operators can sell spare compute capacity at cell sites to enterprise customers for use cases like industrial vision AI and robotics, not just on network performance improvements alone.
Why This Matters Beyond Telecom
The AI-RAN push lands at a moment when NVIDIA’s $500 billion AI spending bet is under scrutiny across the entire technology sector, with data center capacity, power supply, and GPU availability all cited as bottlenecks for the broader AI buildout. Turning existing telecom infrastructure into usable AI compute, rather than building new greenfield data centers, is a materially different approach to solving that capacity problem, and it is one that plays to NVIDIA’s existing hardware advantage while giving telecom operators a new growth story beyond selling phone plans and broadband. A Unite.AI writeup of the GTC Berlin news and a 24/7 Wall St analysis both framed the announcement as a bet that cell towers could increasingly function as AI servers in the years ahead, though neither NVIDIA nor Nokia has published a confirmed commercial timeline to match that framing.
It also raises a practical question for everyday mobile users: does AI processing running on the same hardware as phone calls and data traffic improve the service people already pay for, or does it compete with that traffic for limited GPU and power resources at each site? Deutsche Telekom’s reported uplink throughput gain suggests the former, at least in that specific trial, but the industry is still in the early pilot stage, and five-digit site counts at gigawatt power scale, the kind Huang described in his long-term forecast, remain years away from any confirmed rollout timeline.
Frequently Asked Questions
What is AI-RAN?
AI-RAN refers to radio access network hardware that runs AI workloads directly alongside radio signal processing, using GPU-accelerated compute instead of fixed-function radio chips. NVIDIA’s Aerial RAN Computer platform, paired with Nokia’s anyRAN software, is the specific implementation announced at GTC Berlin on October 21, 2026.
What exactly did NVIDIA and Nokia announce?
NVIDIA and Nokia announced that telecom operators including T-Mobile, BT, Deutsche Telekom, Elisa, and Orange are running AI-RAN deployments or pilots using Nokia’s anyRAN software on NVIDIA’s Aerial RAN Computer platform. T-Mobile completed what NVIDIA calls the world’s first AI-RAN call on a commercial network, and Elisa ran Europe’s first AI-RAN field trial.
How many telecom operators are using AI-RAN?
Nearly 20 telecom operators worldwide have publicly announced AI-RAN collaborations using Nokia anyRAN software and NVIDIA’s Aerial RAN Computer, according to the companies. Most are still at the pilot or trial stage rather than full commercial deployment.
What did T-Mobile actually test?
T-Mobile’s AI-RAN call ran on commercial Aerial RAN Computer servers with Nokia’s CUDA-accelerated RAN software, using massive MIMO radios connected to T-Mobile’s existing 5G Standalone core network, with a commercial smartphone on the receiving end rather than lab equipment.
Does AI-RAN improve network speed for regular customers?
Deutsche Telekom reported uplink throughput improvements of 10 to 15 percent in its AI-RAN tests, the only quantified performance figure tied to a named operator in this announcement. Broader performance claims for other operators have not been published.
Is AI-RAN the same thing as 6G?
No. AI-RAN is being deployed on existing 5G Standalone networks today. NVIDIA has separately discussed AI-native wireless network design as part of its work with telecom industry leaders on next-generation networks, but that is a distinct, longer-term initiative rather than the same announcement covered here.
Which companies compete with NVIDIA and Nokia in this space?
Ericsson and Samsung both sell RAN equipment and pursue their own AI-assisted network strategies, though neither has announced a comparable multi-operator deployment built on GPU-class accelerated compute at the cell site level.
When will AI-RAN be available to everyday mobile customers?
No confirmed commercial launch date has been published for any operator’s AI-RAN service. T-Mobile and Elisa have completed technical milestones, while BT, Deutsche Telekom, and Orange are described as running pilots, an earlier stage than live commercial service.
