Vodafone Is Sold on AI-RAN but Not on GPUs: A Skeptical Exec's Q&A on the Silicon Question

Vodafone Is Sold on AI-RAN but Not on GPUs: A Skeptical Exec’s Q&A on the Silicon Question

Every few months the AI-RAN conversation resets itself around the same unresolved question: does intelligence in the radio access network actually require a GPU? This week, Vodafone—one of the operators most closely watched on RAN strategy—gave the clearest signal yet that the answer is “not necessarily.” Speaking at an AI-RAN symposium hosted by Omdia (a Light Reading sister company), Kyriakos Exadaktylos, Vodafone’s head of RAN evolution, endorsed AI’s role in improving RAN performance while conspicuously declining to endorse the GPU-centric architecture that Nvidia and its partners have been promoting.

That nuance matters enormously for anyone building or buying Open RAN. The industry has spent two years conflating “AI-RAN” with “GPU-RAN,” and the two are not the same thing. Below, we work through the story the way a skeptical operator executive would—question by question—using what was actually said and reported this week.

What exactly did Vodafone say, and what did it not say?

Start with the substance. Exadaktylos was quoted saying, “I think the route to improve is quite significant, and AI will help for that.” That is a genuine endorsement of AI as a lever for RAN performance—spectral efficiency, link adaptation, resource management, and the rest. Vodafone is not a skeptic about the technology’s potential.

What Vodafone has not signed up for is the proposition that this improvement must run on GPUs. The headline framing from Light Reading captures it precisely: Vodafone is sold on AI-RAN but not yet on the need for GPUs. For an operator, that is a procurement stance as much as a technical one. GPUs carry power, cost, and supply-chain implications that differ sharply from the ASIC and merchant-silicon paths that most RAN equipment already runs on.

Why would an operator hesitate on GPUs if the performance claims are real?

The hesitation is not about doubting AI. It is about where the inference runs, and whether the hardware premium is justified by the gain. The most instructive counterpoint in the reporting comes from Ericsson, one of Vodafone’s preferred vendors, which is running an AI-native form of link adaptation—an important RAN task—on internally designed ASICs.

Gabriel Foglander, Ericsson’s head of RAN strategy, explained the constraint that makes this a silicon decision rather than a software one: “The really critical part here is that you have about 40 microseconds before you need to make a decision on link optimization. The inferencing speed is really critical – and the deterministic nature of that. So, for this functionality, we’ve used our neural network model.”

Forty microseconds. That is the kind of latency budget that forces a conversation about determinism, not just raw throughput. A general-purpose GPU can be fast, but the operator’s question is whether it is fast and predictable enough, at the right cost and power envelope, to justify displacing silicon that is already in the radio. Ericsson’s answer—a neural network model running on custom ASICs—is a direct challenge to the assumption that AI inference in the RAN defaults to a GPU.

Is there evidence that non-GPU silicon can deliver the same AI-RAN gains?

Yes, and it is arguably the most important data point in this week’s reporting. A recent AI-RAN demo carried out by Nokia with the UK’s Virgin Media O2, running on Marvell silicon, showed a 20% improvement in spectral efficiency—the same level Nokia previously talked about achieving this year on GPUs.

Read that again. A 20% spectral efficiency gain is a headline number in AI-RAN marketing. The fact that it was demonstrated on Marvell silicon, not GPUs, undercuts the idea that GPUs are a prerequisite for the performance operators actually care about. It also fits a broader pattern: Nokia has been signaling a software-first, merchant-silicon-confident evolution in its AI-RAN messaging, even as it maintains a GPU-based product line co-developed with Nvidia.

For a skeptical exec, this is the crux. If the same class of gain is achievable on both silicon paths, then the GPU question becomes a question of economics, roadmap, and ecosystem—not physics.

So is this a blow to Nvidia?

It is a complication, certainly. Nvidia popularized the broad AI-RAN concept, and its commercial thesis depends on operators accepting GPU-based compute in the RAN. Any telco resistance to investing in GPUs hinders that thesis. The reporting notes that Nvidia has bigger worries following the abrupt exit of CEO Jensen Huang last week—and, separately, that Huang is reportedly weighing a return as supporters press the board. Leadership turmoil at the company at the center of the AI-RAN narrative adds uncertainty to any operator’s multi-year silicon bet.

The more acute pain, though, may be felt by Nokia. Nvidia took a $1 billion stake in Nokia this time last year, with Nokia announcing plans to base future RAN products exclusively on Nvidia GPUs. That “exclusively” is now the phrase to watch. Nokia is expected to continue supporting the latest products in commercial networks—co-developed with Marvell Technology—for several years, and it will not take custom silicon development forward. In other words, the Marvell-based line that just delivered 20% spectral efficiency with Virgin Media O2 remains commercially relevant even as the GPU-only future is being positioned.

What does this mean for Open RAN specifically?

Open RAN’s core promise is choice: disaggregated hardware and software, multi-vendor interoperability, and the ability to mix components. A GPU-only AI-RAN future narrows that choice at the compute layer. An ASIC-and-merchant-silicon path—where AI inference runs on vendor silicon or on general-purpose processors already in the RAN—preserves more of it.

That is why Vodafone’s stance is more than a procurement detail. Vodafone is a major Open RAN proponent, and its RAN evolution chief saying AI will help but GPUs are not yet proven necessary is a signal that the operator intends to keep its options open. For the Open RAN ecosystem, that is healthy. It means AI-RAN does not have to become a single-vendor compute story, and it keeps pressure on vendors to demonstrate gains on whatever silicon the operator already owns.

Aren’t vendors making grand claims about AI-RAN improvements?

They are, and the reporting is candid about it: vendors have made some grand claims about the percentage improvements AI-RAN could deliver. The 20% spectral efficiency figure is the one with the most concrete recent backing, appearing in both the Nokia–Virgin Media O2 demo on Marvell silicon and in Nokia’s earlier GPU-based claims for this year.

A skeptical exec should treat a 20% spectral efficiency gain as meaningful but not transformative on its own, and should ask three follow-ups: Under what traffic and radio conditions was it measured? Is it repeatable across vendors and bands? And what does it cost in power and silicon to achieve? The reporting does not answer these, and operators should not assume the headline number translates directly to their network.

Is the industry converging on one answer?

No—and that is the story. Look at the broader vendor landscape reported this week. Nokia introduced an AI-native RAN platform with Nvidia in July 2026 and in October shared further AI-RAN progress with operators including BT, Deutsche Telekom, Elisa, and Orange, focused on smarter radio control and better resource use. Samsung was named sole global vendor for KT’s AI-RAN project and main vendor for SK Telecom’s project in September 2026. Ericsson is pushing AI-native link adaptation on its own ASICs. SoftBank is preparing to showcase its AITRAS platform—which runs AI and RAN workloads on the same hardware—and its Large Telecom Model at MWC27 Barcelona, positioning itself as a vertically integrated AI platformer.

That is a spectrum of approaches, not a consensus. Some vendors bet on GPUs, some on ASICs, some on running AI and RAN on shared general-purpose hardware. The market data reinforces that this is still early: Open vRAN and vRAN made up 7.3% of RAN revenue in 2025, with a projected 9.3% share in 2026. AI-RAN sits on top of a still-modest virtualized base.

What should an operator actually decide now?

The practical takeaway is to decouple two decisions that the industry has been bundling together. Decision one: do we believe AI improves RAN performance enough to invest? Vodafone’s answer is yes. Decision two: does that AI have to run on GPUs? Vodafone’s answer, at least for now, is not necessarily—and Ericsson’s 40-microsecond ASIC link adaptation and Nokia’s 20% Marvell-silicon result give that position technical cover.

For operators, that suggests a posture of architectural neutrality: demand AI-RAN gains be demonstrated on multiple silicon paths, keep GPU and non-GPU options open in procurement, and treat vendor exclusivity claims—including Nokia’s GPU-only positioning—as a roadmap to be validated rather than a settled fact. For vendors, it means the burden of proof has shifted. The question is no longer whether AI-RAN works; it is whether it works on the silicon the operator can afford to deploy, at the power and latency the radio demands.

The bottom line

Vodafone’s position is the most consequential AI-RAN signal of the week precisely because it is unglamorous. It accepts the technology’s promise while refusing to accept its most expensive architectural assumption. Ericsson’s ASIC-based link adaptation and Nokia’s Marvell-silicon demo show that the GPU question is genuinely open. And with Nvidia’s leadership in flux and Nokia’s GPU-exclusive bet facing a multi-year Marvell-supported commercial reality, the silicon debate in AI-RAN is entering its most interesting phase yet.

For Open RAN, that is good news. A multi-silicon AI-RAN future is a more open one—and openness, after all, is the whole point.

Sources

Related Posts

Vodafone's Málaga Lab and the Silicon Fork Behind AI-RAN

Vodafone opens a multi-vendor AI-RAN hub in Málaga while VodafoneThree trials Nokia and NVIDIA — and Orange weighs GPU versus ASIC.

VodafoneThree's AI-RAN Lab Goes Live: Inside the Málaga Hub Where AI-for-RAN, AI-and-RAN, and AI-on-RAN Get Measured

VodafoneThree opens a Málaga AI-RAN hub with Nokia and NVIDIA, testing AI-for-RAN, AI-and-RAN and AI-on-RAN on shared compute ahead of 2027 field trials.