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

The Setup: A Lab, Not a Launch

Strip away the press-release gloss and the VodafoneThree announcement of 18 November 2026 is a statement about measurement, not magic. The UK operator, working with Nokia and NVIDIA, has stood up an AI-RAN innovation and validation hub in Málaga, Spain, and laid out a programme that runs lab testing through the first half of 2027 before graduating to selected-site field trials. The headline number — trials starting in H1 2027 — is deliberately unglamorous. That is the point.

For an industry that has spent two years announcing AI-RAN partnerships faster than it has shipped AI-RAN features, the interesting content here is architectural. VodafoneThree is not simply bolting an AI model onto a conventional distributed RAN. It is testing three distinct propositions — AI-for-RAN, AI-and-RAN, and AI-on-RAN — on a shared, multi-vendor compute platform, and it is committing to independent, repeatable benchmarks to decide which of them actually pays. Kyriakos Exadaktylos, Head of RAN Evolution in Vodafone Group Technology, framed the lab’s purpose precisely: to establish “independent, repeatable benchmarks that show where AI delivers real value in mobile networks,” and to test “how AI and RAN workloads can share computing infrastructure securely and efficiently.”

That second clause is where the real analytical interest sits, and it is where operators should focus their diligence.

Why This Story Matters More Than the Headline

Within a three-day news window, this story competed with several adjacent items. Omdia’s AI RAN operator survey, published 20 November, offered useful directional data — a small lead group of roughly 12% of respondents have deployed or plan to deploy AI models directly in the L1/L2/L3 radio stack, with close to a quarter in live network trials. Orange’s commentary on the Ericsson-versus-Nokia silicon fork, reported 18 November, added colour on the strategic divergence between custom ASIC and merchant silicon approaches. Both are relevant. Neither is a discrete, verifiable engineering commitment.

The VodafoneThree programme is. It names the hardware, names the software stack, names the test methodologies, and names the use cases. In a market where AI-RAN claims are abundant and reproducible results are scarce, a lab with a published benchmark regime is a more useful signal than another memorandum of understanding. It also extends beyond the UK operator’s own footprint: the Málaga facility is positioned as a multi-vendor European validation hub, drawing on Vodafone’s existing Open RAN foundation to strengthen Europe’s role in AI development and standardisation.

The Three Pillars, Decoded

The AI-RAN Alliance’s three-pillar framework has become industry shorthand, but it is worth being precise about what each pillar actually asks of the network, because the answers have very different capital and operational consequences.

AI-for-RAN is the least controversial and the most immediately monetisable. Here AI improves the performance of the radio itself — machine learning-based channel estimation and link adaptation are the named initial use cases. In plain terms, the network gets better at judging changing radio conditions and adjusting how data is transmitted in real time. The operator’s stated ambition is improved capacity, reliability and spectrum efficiency, delivering a more consistent customer experience. This is where the Omdia survey data and the VodafoneThree programme converge: software updates to existing RAN equipment can improve coverage, energy consumption, uplink performance and spectral efficiency without a forklift upgrade.

AI-and-RAN is the strategically explosive pillar. It asks whether RAN workloads and AI workloads can share the same computing infrastructure — accelerated compute, purpose-built telecoms ASICs, and the x86 processors used in Open RAN — securely and efficiently. If the answer is yes, the economics of the radio edge change fundamentally: GPU capacity procured for RAN can be monetised against AI inference demand during off-peak radio hours. If the answer is no, or yes only with unacceptable isolation overhead, then the GPU-in-the-tower thesis loses its most compelling financial argument. VodafoneThree is explicitly testing resource sharing between RAN and AI workloads, and the Málaga lab is the right place to find out.

AI-on-RAN looks outward. It examines hosting AI applications — computer vision, robotics — closer to end users, using intelligence at the network edge to enable new customer services. This is the pillar with the longest commercial runway and the murkiest attribution, because the revenue accrues to enterprise customers rather than to the connectivity line item.

The Hardware: A Deliberately Heterogeneous Bench

The specification detail is where this announcement earns its credibility. The Málaga lab is equipped with QCT servers hosting the NVIDIA Aerial RAN Computer MGX 2U Short platform, alongside Intel Xeon 6 Granite Rapids servers. That is not a single-vendor reference design; it is a comparison bench. The operator intends to assess silicon performance using Standard Performance Evaluation Corporation benchmarks, AI performance through MLPerf, and RAN performance through the AI-RAN Alliance’s own methodologies. Running all three measurement regimes in parallel is the correct approach, and it is rarer than it should be.

On the software side, the trial stack combines Nokia’s anyRAN software, NVIDIA’s Aerial RAN platform, and Red Hat OpenShift to create what the operator describes as a common platform capable of supporting both RAN and AI workloads. The inclusion of OpenShift matters: container orchestration is the mechanism by which workload isolation, scheduling and security boundaries between RAN and AI functions will either hold up or fail. Any operator serious about AI-and-RAN should be watching how that layer performs under mixed load.

Nokia’s role deserves a note of context. The vendor is placing a substantial bet on general-purpose and merchant silicon, arguing that it can ride ongoing cost and performance improvements in general purpose computing to create an open platform capable of running advanced AI algorithms — many of which have not yet been invented. NVIDIA is both platform co-developer and an investor in Nokia. Ericsson, by contrast, argues that tightly integrated RAN software on multi-core custom silicon delivers highly optimised performance with ample compute for AI workloads. VodafoneThree is running this Nokia-NVIDIA track in parallel to separate work with Ericsson, which is the pragmatic choice for an operator that does not yet know which architecture wins.

ISAC and the Sensing Upside

Buried in the planned work is integrated sensing and communications, or ISAC — combining wireless connectivity with sensing capability for applications such as detecting intruders at sensitive sites. This is a genuinely differentiated use case and one that reframes the RAN as a sensor network rather than purely a transport pipe. It also explains why AI-on-RAN and edge compute belong in the same lab as AI-for-RAN: sensing generates data, edge AI processes it, and the commercial customer is an enterprise buyer with a security or safety budget, not a consumer with a data plan.

ISAC is early. Standards work is ongoing, and commercial deployments are some distance away. But the fact that a tier-one European operator is scheduling ISAC validation alongside channel estimation and link adaptation suggests the roadmap is being built as a continuum rather than a sequence of isolated experiments.

The Benchmark Question: Who Grades the Homework?

Here is where a skeptical analyst should press. The programme commits to independent, repeatable benchmarks — but independence is a spectrum. A lab funded and operated by an operator, using vendor-supplied platforms and vendor-co-developed test methodologies, is more independent than a vendor’s own marketing lab and less independent than a third-party test house. The AI-RAN Alliance’s performance methodology is still maturing, and MLPerf, while credible for AI workloads, was not designed for RAN-realistic latency and reliability constraints.

The genuinely valuable output from Málaga would be published, reproducible results that allow other operators to compare their own findings. If VodafoneThree delivers that, it will have done the industry a service well beyond its own network. If the results remain internal, the lab becomes a sophisticated procurement exercise — useful, but not the ecosystem contribution the framing implies.

The Economic Frame: Why 2027 Is the Right Date

The H1 2027 field trial timeline is a tell. It implies roughly two to three quarters of lab work before sites are selected, which is consistent with the maturity of the underlying stack rather than with a marketing calendar. It also aligns with the broader pattern: Omdia’s survey shows operators moving from trials toward deployment, and multiple European and Asian operators have announced AI-RAN programmes with Nokia and NVIDIA since MWC26.

The economic logic is straightforward. RAN capex has been flat to declining, and AI-RAN will not change that arithmetic on its own. What AI-RAN can change is the return on compute already deployed. If AI-and-RAN resource sharing works, the marginal cost of AI inference at the edge falls materially, and the business case for GPU-accelerated infrastructure shifts from “RAN performance improvement” — a hard sell against custom ASIC incumbents — to “RAN performance plus monetisable compute,” which is a far easier conversation with a CFO.

That is why the AI-and-RAN pillar, not AI-for-RAN, is the one to watch. AI-for-RAN delivers incremental spectral efficiency gains that will be competed away. AI-and-RAN determines whether the radio edge becomes a revenue-generating compute platform or remains a cost centre with better algorithms.

What Operators Should Take Away

First, treat the three pillars as three separate investment decisions with three separate payback profiles. Vendors that bundle them into a single AI-RAN narrative are obscuring the economics.

Second, insist on workload isolation evidence. The security and efficiency of sharing compute between RAN and AI functions is the crux of AI-and-RAN, and it is testable. Ask for the numbers.

Third, watch the orchestration layer. Red Hat OpenShift’s presence in this stack is not incidental; the container platform is where multi-tenancy, scheduling and security boundaries are enforced. Operators evaluating AI-RAN should be asking hard questions about orchestration performance under mixed RAN and AI load.

Fourth, note that VodafoneThree is hedging across silicon architectures. Running Nokia-NVIDIA in parallel with Ericsson is prudent given the unresolved GPU-versus-ASIC debate, and it preserves negotiating leverage. Operators that commit early to a single architecture are making a bet the evidence does not yet support.

Finally, value the benchmark commitment. In a market where AI-RAN claims outrun AI-RAN results, reproducible measurement is the scarcest commodity. Málaga’s most important output may not be a performance gain at all — it may be a methodology the rest of the industry can trust.

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

VodafoneThree's Nokia-Nvidia AI-RAN trials and Orange's vendor calculus show the industry's silicon debate is now the real Open RAN story.