Nokia’s Big Bet: GPUs in the Cell Tower and the Fight Over How Networks Get Smarter

Something genuinely interesting is happening in the world of mobile networks right now, and it’s the kind of thing that sounds technical but is actually about money, strategy, and two very different visions of the future. Nokia is going all-in on putting graphics chips — the same kind of hardware that powers AI chatbots — inside cell towers. Ericsson, its biggest rival, is politely declining. And the rest of the industry is watching to see who’s right.

Let’s break down what’s actually going on.

The Basic Idea: What Is AI-RAN?

First, some quick context. The radio access network, or RAN, is the part of a mobile network that connects your phone to the internet. It’s the antennas, the base stations, the equipment at the edge. Traditionally, that equipment has been built around specialized, custom-designed chips made for one job: moving radio signals around reliably and efficiently.

AI-RAN is the idea that you could use AI-capable hardware — specifically GPUs, the chips Nvidia dominates — to do that radio work and run AI workloads at the same time. Instead of a box that only does radio, you get a box that does radio plus whatever else an operator wants to sell or run at the edge of the network.

That’s the pitch, anyway. And it’s a big pitch, because it implies replacing a lot of specialized hardware with more general-purpose computing. It also implies that operators might one day make money from AI services running in their networks, not just from selling data plans.

Nokia’s Gamble

According to reporting from Light Reading, Nokia has been aggressively advertising telco interest in its forthcoming AI-RAN products. At its recent Radio World event in Oulu, Finland — a city near the Arctic Circle that’s long been a Nokia engineering hub — the company said 19 operators have signed up for what CEO Justin Hotard called “initial deployments.” Notably, the three companies he named in a LinkedIn post aren’t actually telecom operators, which is the kind of detail worth flagging.

Nokia has also been leaning on its partnership with Cohere, a startup whose technology Nokia believes could help it hit an ambitious target: doubling spectral efficiency by 2028 on much of the existing infrastructure. Spectral efficiency, in plain terms, is how much data you can squeeze through a given slice of radio spectrum. It matters because spectrum is finite and expensive. Vodafone, Telstra in Australia, and Bell Canada have all publicly endorsed the technology after field trials, and all three have invested in Cohere.

The broader commercial picture, per EME Outlook Magazine, is that in 2026 Nokia launched what it calls the industry’s first commercial AI-RAN platform, built on its anyRAN software and Nvidia’s Aerial AI-RAN platform. Nokia says it demonstrated GPU-accelerated AI-RAN with operators including BT, Elisa, Vodafone Group, and NTT DOCOMO, validating that AI inference workloads and radio processing can run concurrently on shared infrastructure. The company’s stated ambition is more than 100 percent spectral efficiency gains by 2028 — in other words, roughly doubling capacity without a traditional hardware refresh.

That last part is the pitch that matters to operators: your existing investment carries you into the 6G era instead of forcing another rip-and-replace cycle.

Ericsson’s Cautious Counterpoint

Ericsson is taking what EDN describes as a more evolutionary approach. Rather than betting on GPUs as a general-purpose substitute for custom silicon, Ericsson is embedding AI capabilities into the RAN infrastructure operators already have. The goal is better performance — particularly around scheduling and radio resource management — without a costly, large-scale hardware replacement.

Early trial results, according to EDN, point to real gains in efficiency and throughput. And there’s a concrete data point from the broader ecosystem: in August 2026, SoftBank and Ericsson conducted Japan’s first validation of AI-in-RAN technology on a commercial 5G network. The AI-native scheduler achieved up to roughly 25 percent higher spectral efficiency and 50 percent higher downlink user throughput compared with conventional technology. That’s a meaningful improvement, and it came from software intelligence layered onto existing infrastructure rather than from swapping in new chip architectures.

This is the crux of the disagreement. Nokia says the future is GPU-powered, general-purpose compute in the RAN. Ericsson says you can get a lot of the benefit by making existing RAN smarter with AI, and you don’t have to bet the farm on a new hardware paradigm to do it.

Why This Is a Real Bet, Not Just Marketing

The reason Nokia’s approach is described as a “massive gamble” — Light Reading’s words — comes down to physics and economics. GPUs are powerful, but they’re also power-hungry. EDN notes plainly that putting more GPUs into the RAN brings higher power consumption, cooling requirements, and infrastructure costs.

That’s not a small thing. Cell sites are often constrained by power and cooling budgets. Adding racks of GPUs to thousands of towers is a very different proposition from adding software to existing hardware. Operators will ultimately have to decide whether the revenue from running AI workloads at the edge is enough to justify those costs.

And that’s the real question: is there actually a business here? If operators mostly want better network performance — faster speeds, more capacity, fewer dropped connections — then adding AI to existing RAN infrastructure may be enough. The case for GPU-based AI-RAN becomes much stronger if operators see an opportunity to turn that extra compute into a genuinely new business around distributed AI workloads. That could mean running AI inference for enterprise customers at the edge, closer to where data is generated, with lower latency.

But that business doesn’t exist at scale yet. It’s a bet on the future.

The Nvidia Factor

You can’t tell this story without Nvidia. The chipmaker has backed its vision with a $1 billion investment in Nokia, and the two companies already have live trials running with T-Mobile. Nvidia has assembled its own framework for a RAN software stack, branded Aerial, which runs on its CUDA software platform. The idea is that other developers, provided they’re CUDA-literate, can take Aerial and customize it.

This is a familiar Nvidia playbook: make your platform the default, make it broadly accessible, and let the ecosystem build on top of it. It’s worked spectacularly in data centers. Whether it works in the RAN — a market with long product cycles, strict reliability requirements, and deeply entrenched competitors — is a different question.

There’s also a strategic dimension. Nvidia positioning GPUs as a general-purpose substitute for custom silicon in the RAN is, in effect, a challenge to the specialized chip vendors who’ve dominated that space. If Nvidia succeeds, it opens up a whole new market. If it doesn’t, it’s an expensive experiment.

The Path to 6G

All of this is happening against the backdrop of 6G development. Per EDN, the standards process has key milestones ahead: Release 20 is expected to conclude in 2027, followed by the normative 6G specification in Release 21 and a final protocol freeze in March 2029. After that, the focus shifts to operators and equipment manufacturers to turn specifications into reliable systems.

AI-RAN is widely viewed as one of the stages toward increasingly AI-native networks in the 6G era. Nokia’s own leadership has pointed to 2026 as the first year of commercial trials for 6G, with commercial release targeted for next year. The company’s longer-term ambition, per EME Outlook, is to be the partner that unifies intelligence across every layer of the network — RAN, core, IP, and optical — so they work as one coordinated system with transparency, traceability, and safety controls built in from design.

That’s a big vision. And it explains why Nokia is willing to make a bold hardware bet now: it wants to be positioned as the vendor that carries operators from today’s networks into the AI-native 6G era without forcing a disruptive reset.

What Operators Are Actually Doing

The picture across the industry is mixed, which is itself informative. Some operators are running proof-of-concept and field trials of AI-RAN using Nvidia’s Aerial RAN Computer — Nokia says trials have been conducted in North America, Europe, Asia-Pacific, and the Middle East, with the platform showing more than 20 percent spectral efficiency improvement in testing.

Others are taking a more measured path. Deutsche Telekom and Nokia, for example, are advancing O-RAN-aligned management using open O1 interfaces, with work covering configuration management and integration into Deutsche Telekom’s internal service management platform. Their AI-native RAN program includes cloud, edge, and radio-domain integration and trials of AI-based network optimization. That’s the unglamorous but essential plumbing work that makes multi-vendor networks manageable.

Meanwhile, in South Korea, carriers are racing to secure overseas patents in 6G and related technologies. SK Telecom filed a PCT international application for technology that links radio access networks with distributed edge data centers to allocate AI computing workloads efficiently — distributing work based on conditions across the network and multiple edge data centers. The company is also feeding research results into global standards discussions alongside its AI-RAN field trials. LG Uplus is pursuing low-earth-orbit satellite network technology with KAIST. KT has obtained patents in the US and Europe covering quantum key distribution.

Why does this matter? Because patents included in international standards can open the door to large licensing markets. LexisNexis IP estimates the 5G patent licensing market at about $15 billion a year. As that market grows, telecom and equipment companies are competing to patent candidate technologies — including AI-based network control, RAN, and edge computing — before standards are finalized.

That’s the real stakes behind the technical debate. It’s not just about which chip architecture wins. It’s about who owns the intellectual property that will underpin the next generation of networks.

The Honest Assessment

So who’s right? Honestly, it’s too early to say, and anyone claiming certainty is selling something.

The case for Nokia and Nvidia’s approach is that AI workloads are going to move closer to users, and the RAN is a natural place to host them. If that happens, having GPU-capable compute in the network becomes a strategic asset, not just a cost center. The 20 percent spectral efficiency gains Nokia reports are real and meaningful. The 2028 target of doubling efficiency is ambitious but not obviously impossible.

The case for Ericsson’s approach is that operators are pragmatic. They have existing infrastructure, tight budgets, and a strong preference for incremental improvements over disruptive overhauls. If you can get 25 percent better spectral efficiency and 50 percent better throughput from software, as the SoftBank trial suggests, that’s a compelling near-term story. And it doesn’t require betting on a new business model that may or may not materialize.

The likely outcome is that both approaches coexist for a while. Some operators, particularly those with edge computing ambitions or dense urban networks where capacity is most valuable, will experiment with GPU-based AI-RAN. Others will focus on AI-enhanced RAN that squeezes more out of what they already own. Over time, as 6G standards solidify and the economics become clearer, the industry will converge on whatever actually works.

What’s clear is that the RAN — that unglamorous collection of antennas and base stations that most people never think about — is becoming one of the most strategically important battlegrounds in tech. The decisions being made now about chip architectures, software platforms, and standards will shape how mobile networks work for the next decade.

And Nokia, for better or worse, has decided to make a very big bet on one particular version of that future. Whether it pays off is the question the entire industry is waiting to answer.

Sources

• Nokia puts new open RAN tech in 5G revival bid as Ericsson demurs
• 5G to 6G: AI moves into the network - EDN
• Nokia Doubles Down on AI, Partners with Microsoft and NVIDIA for 6G
• Telecom System Integration Market (2035): 5G Network
• Nokia Advances AI and Partners with Microsoft, NVIDIA for 6G
• Nokia : Intelligent Networks by Design - EME Outlook Magazine
• Korean Telecoms Race to Secure Overseas Patents in 6G, Quantum