VodafoneThree, Orange and the AI-RAN Silicon Question: A Decade of Disaggregation Meets the GPU-or-ASIC Fork

VodafoneThree, Orange and the AI-RAN Silicon Question: A Decade of Disaggregation Meets the GPU-or-ASIC Fork

The week of 17 November 2026 may not be remembered as the moment AI-RAN arrived. It will more likely be remembered as the week the industry stopped pretending the hard part was the AI, and admitted the hard part is the silicon.

In the space of roughly 48 hours, VodafoneThree confirmed lab and field trials of AI-RAN technology built on Nokia’s RAN software, Nvidia’s Aerial RAN compute and Red Hat’s OpenShift cloud layer; Vodafone Group opened an AI-RAN test centre in Malaga, Spain, where the same facility previously hosted open RAN testing; and Orange’s RAN leadership publicly worked through the GPU-versus-ASIC fork that now defines the vendor landscape. Light Reading, TelecomTV and The Mobile Network all covered variants of the same story. That convergence is itself the signal.

To understand why this week matters, it helps to remember how we got here — because the AI-RAN conversation is really the third act of a story that began with a very different promise.

Act One: The Disaggregation Promise

Open RAN’s founding argument was never about artificial intelligence. It was about interfaces. Break the radio access network into disaggregated, interoperable components — the distributed unit, the centralised unit, the radio unit — connect them through open fronthaul and open management interfaces, and competition would migrate from proprietary hardware to software and merchant silicon. The bet was that general-purpose compute, riding the cost curves of the IT industry, would eventually outrun purpose-built telecom ASICs.

For years, that bet was contested but stable. Ericsson and Nokia defended integrated, optimised designs. New entrants championed virtualised RAN on x86. Operators ran trials, published blueprints, and argued about performance gaps measured in single-digit percentages.

Then AI arrived and changed the question. If the RAN of the future must run machine learning inference — channel estimation, link adaptation, scheduling, energy optimisation — then the compute substrate stops being a cost decision and becomes a strategic one. The RAN is no longer just a pipe. It is a compute platform that happens to transmit radio.

Act Two: The GPU Enters the Tower

The Nokia-Nvidia partnership moved that argument from theory to product. Nvidia’s Aerial RAN computer platform paired with Nokia’s RAN software promised something the industry had talked about but rarely demonstrated: the ability to run RAN workloads and AI workloads on the same accelerated infrastructure. T-Mobile in the US completed the first voice call on a live commercial network using Nokia’s RAN software and Nvidia’s Aerial platform last month. Ronnie Vasishta, who leads Nvidia’s telecom activities, told journalists that commercial deployments of the platform with Nokia are expected to start next year.

That timeline is now being stress-tested across Europe. BT plans to trial the Nvidia platform next year in EE’s live mobile network, with the first AI-RAN sites from Nokia and Nvidia expected in Birmingham. Deutsche Telekom, Elisa and Orange are all planning AI-RAN field trials with the same pairing. And this week VodafoneThree joined the list.

Act Three: VodafoneThree’s Trial, and What It Actually Tests

The VodafoneThree announcement is easy to skim as another operator press release. It is more interesting than that, because of what the operator says it will measure.

The trials will focus on “AI for RAN” — using machine learning for channel estimation and link adaptation, so the network can more accurately assess changing radio conditions and adapt data transmission in real time. Results will be benchmarked against VodafoneThree’s commercial network, comparing capacity, reliability and spectral efficiency to determine where AI is effective and how it could be implemented across the network.

That benchmarking discipline is the detail worth noting. For years, AI-RAN claims have been presented as laboratory multipliers. VodafoneThree is explicitly framing its trials as a comparison against a live commercial baseline. That is the difference between a demo and a procurement decision.

The operator has also identified a broader use-case map: radio optimisation, mobility, energy efficiency, anomaly detection, coverage and capacity optimisation, quality-of-service assurance, and — crucially — the sharing of accelerated compute between RAN and AI workloads. That last item, sometimes called AI-and-RAN, is one of the three pillars identified by the AI-RAN Alliance. It is also the most commercially consequential, because it determines whether an operator’s AI-RAN investment can host third-party AI revenue or merely subsidise its own radio efficiency.

On the architecture side, the trial brings together Nokia’s AnyRAN software, Nvidia’s Aerial RAN computer and Red Hat OpenShift to create what the operator describes as a common platform capable of supporting both RAN and AI workloads. Multi-vendor shades are visible throughout.

Malaga: The Lab That Changed Its Question

The VodafoneThree news coincided with Vodafone Group opening an AI-RAN test centre in Malaga, Spain. The location is not incidental. It is the same facility where open RAN testing was carried out — a physical reminder that the industry’s research infrastructure is being repurposed from an interface question to an intelligence question.

And Vodafone is not treating the silicon decision as settled. At Malaga, the group will evaluate multiple RAN hardware platforms: Nokia and Nvidia’s GPU-based compute, purpose-built ASICs, and x86 processors — specifically Intel’s Granite Rapids. It will also be able to test Nokia’s target of doubling spectral efficiency by 2028.

That multi-platform approach is consistent with Vodafone’s public posture. The company’s head of RAN evolution has been clear that Vodafone is investigating the benefits of different network chips for AI-RAN, and Light Reading has reported that Vodafone is sold on AI-RAN but not yet on the need for GPUs. In other words: the operator believes the intelligence layer is coming, but is refusing to pre-commit to the accelerator underneath it.

That is a more sophisticated position than either the GPU camp or the ASIC camp tends to acknowledge. It treats the silicon question as an empirical one — to be answered by benchmarks in Malaga and field data in the UK, not by vendor slideware.

Orange and the Opinionated Vendors

If VodafoneThree’s news is the what, Orange’s commentary this week is the why. Speaking to The Mobile Network, Orange’s RAN leadership described a process that started at the beginning of the year and has now produced a good understanding of the different Open RAN configurations under consideration. The next step, they said, is understanding how AI-RAN implementations can increase the capacity and performance of those architectures — and the offerings from some vendors are only now becoming much clearer.

Orange’s conclusion is notable for its even-handedness. The operator is convinced it will need to work with both Ericsson and Nokia. “There are pros and cons to both,” the Orange executive said, adding that what is appealing about both approaches is that the companies are opinionated — they can explain clearly why they believe their way of thinking is best.

That framing captures the current state of the AI-RAN silicon debate better than any marketing document. Put over-simply: Ericsson sees the benefits of tightly integrating RAN software with its multi-core custom silicon, arguing that custom ASICs deliver highly optimised performance and, thanks to multi-core architecture, have plenty of compute to run AI workloads. Nokia is moving away from custom silicon toward general-purpose and merchant solutions, arguing it can benefit from ongoing cost and performance improvements in general-purpose computing and create an open platform capable of running advanced AI algorithms — many of which have not yet been invented.

Both arguments are coherent. They are also incompatible at the level of procurement. An operator cannot hedge indefinitely.

The Memory Crisis and the Cost Question

There is a less glamorous factor in this story that deserves attention: memory. Orange noted it is keeping an eye on potential impacts of the memory crisis, observing that all vendors are affected and that this is a concern because operators want to make sure it does not have too much impact on future costs.

This matters because the entire AI-RAN thesis rests on an assumption of declining compute costs. GPUs are expensive. Accelerated infrastructure is power-hungry. If memory and component costs remain elevated, the economics of putting AI compute in the RAN — particularly for the AI-and-RAN resource-sharing model — become harder to justify against simpler alternatives. The silicon debate is not only about performance. It is about who absorbs the cost of an uncertain demand curve.

Nokia’s Open Radio Ambition

One thread in Orange’s commentary deserves separate treatment, because it points beyond AI-RAN to something larger. Nokia’s Harri Holma confirmed during a panel that the company wants to create a more open radio platform, allowing third parties to develop applications using information generated within the RAN — including interference, positioning and sensing data.

That is a significant ambition. It reframes the RAN not just as a compute host for AI, but as a sensor and data source for an application ecosystem. If Nokia succeeds, the value of the RAN shifts from connectivity to data. If it fails, the open radio platform becomes another interface standard looking for developers.

Either way, it explains why Nokia is willing to bet on merchant silicon. An open platform with third-party applications requires a compute substrate that improves on the IT industry’s cadence, not a bespoke ASIC roadmap measured in telecom generations.

Why This Week Matters

Step back, and the pattern across this week’s news is clear. The AI-RAN conversation has moved through three phases: first, whether AI could improve the RAN; second, whether the RAN could host AI; and now, third, what hardware should sit underneath both.

The first two phases produced consensus and a long list of trials. The third phase will produce winners and losers. VodafoneThree’s trials with Nokia, Nvidia and Red Hat, benchmarked against a commercial network, are a test of the GPU thesis in a live UK environment. Vodafone’s Malaga facility, comparing GPUs, ASICs and x86, is a test of whether the industry can even agree on how to measure the answer. Orange’s careful refusal to pick a side prematurely is a reminder that the largest operators will hedge until the data is unambiguous.

And the data will not be unambiguous for a while. Nokia’s target of doubling spectral efficiency by 2028 is a useful marker: if achieved, it would be the strongest argument yet for the merchant-silicon, software-first path. If it slips, the ASIC camp’s case strengthens.

For the Open RAN community, the deeper lesson is that disaggregation was never the destination. It was the precondition. The interfaces had to be opened before the intelligence could be placed. Now that the intelligence is arriving, the question is where it lives — in custom silicon optimised for radio, or in general-purpose accelerators optimised for everything else.

That question will not be settled by press releases. It will be settled by benchmark results, cost models and, eventually, procurement decisions. This week, three operators and two vendors showed us how seriously they are taking it. That is progress — even if the answer remains, for now, genuinely open.

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.