VodafoneThree and Ericsson Put Hardware Under the AI-RAN Hype

VodafoneThree and Ericsson Put Hardware Under the AI-RAN Hype

Executive briefing memo | 6 November 2026 | Open RAN blog

Bottom line up front

The most consequential AI-RAN news this week is not a forecast, a consortium, or a demo. It is VodafoneThree putting Ericsson’s latest massive MIMO radio kit into its network and readying AI-powered radio optimisation tests on top of it. The UK operator is doing what most of the industry has so far only promised: building the physical layer foundation that any credible AI-RAN deployment requires, then layering intelligence onto real infrastructure rather than lab benches.

For Open RAN operators and their technology suppliers, the signal is straightforward. AI-RAN value will be unlocked by operators who sequence the work correctly — radio modernisation first, automation and intelligence second. The vendors that can supply both halves of that equation, and prove the integration, will capture the near-term spend. The rest will be selling slides.

What actually happened

VodafoneThree, the UK joint venture formed from the merger of Vodafone UK and Three UK, has installed Ericsson’s latest massive MIMO equipment and is preparing AI-powered radio optimisation tests. The story is deliberately unglamorous: hardware underneath the AI-RAN ambition, with the intelligence layer to follow.

That sequencing matters. Massive MIMO is the physical substrate that makes AI-driven radio optimisation worth doing. Beam management, interference coordination, and traffic-aware scheduling all depend on antenna arrays and radio units capable of being steered at software speed. Without modern radios in the field, AI optimisation algorithms have little to optimise. VodafoneThree appears to understand that the intelligence layer is only as good as the aperture beneath it.

The timing is also telling. The announcement lands in the same week that Verizon confirmed a nationwide deployment of the Ericsson Intelligent Controller as part of its own AI-enabled RAN automation push, and days after Ericsson was reported to be working with VodafoneThree on AI-RAN more broadly. Ericsson is quietly assembling a portfolio of operator relationships in which the RAN, the controller, and the optimisation software come from a single vendor. That is a coherent commercial strategy, and it is one Open RAN operators need to weigh carefully.

Why this is the story of the week

Compare the alternatives. SNS Telecom & IT published a forecast that AI-enabled RAN automation and optimisation will grow at roughly 9% annually to reach US$1.1 billion by 2030 — a useful number, but a market-sizing exercise rather than an operational decision. Nokia’s European CTO used a Yerevan summit to forecast that AI could account for more than 40% of network traffic by 2030, with autonomous AI agents becoming the dominant network consumers; provocative, but speculative. Ethio Telecom and ZTE completed Ethiopia’s first pre-commercial AI-RAN deployment, a genuine milestone for that market but one with limited read-across for operators in mature markets. Samsung’s worst 5G quarter is a financial story with AI-RAN implications, but it is a story about vendor pain, not operator progress.

VodafoneThree’s move is different in kind. It is an operator with real subscribers, real spectrum, and real capex commitments making a concrete infrastructure decision that positions it for AI-driven radio optimisation. It is the kind of decision that shows up in earnings and in network performance data, not just in press releases. For a technical and business audience, that is the story worth briefing on.

The architecture question hiding in the announcement

There is a strategic tension embedded in this announcement that operators should not miss. Ericsson’s massive MIMO radios and its Intelligent Controller form a tightly integrated stack. That integration is a strength for deployment speed and performance assurance. It is also a question mark for operators who have bet on multi-vendor openness.

Open RAN’s promise was that intelligence — whether expressed as xApps, rApps, or the newer dApps operating through the proposed E3 interface — could sit above radio hardware from any vendor. The SNS Telecom & IT report notes that reluctance among distributed unit and centralised unit vendors has held back adoption of xApps and the near-real-time RIC, and that some vendors are instead pursuing proprietary AI-enabled real-time optimisation delivered as subscription software upgrades to existing RAN platforms. That is precisely the model Ericsson is executing with VodafoneThree and Verizon.

The implication for Open RAN operators is not that proprietary integration is wrong. It is that the window for proving multi-vendor AI optimisation at scale is narrowing. If the integrated vendors deliver measurable gains first, the burden of proof shifts decisively onto the open ecosystem. Operators running multi-vendor RANs need to demand comparable evidence — and soon.

The economics: where the money actually goes

The SNS Telecom & IT figures are worth reading closely because they separate hype from spend. AI-for-RAN automation and optimisation software and services are expected to reach US$1.1 billion annually by 2030 at a 9% compound annual growth rate. The wider RAN automation market — including traditional SON and proprietary network-level automation, AI-for-RAN upgrades, Open RAN SMO and RIC with d/x/rApps, RF planning tools, and automated test and measurement — grows at only about 4% over the same period.

Two conclusions follow. First, AI-for-RAN is the fastest-growing slice of a slow-growing pie. Second, the absolute numbers are modest relative to RAN capex. This is not a market that will rescue flat RAN economics on its own. It is a market that improves the return on infrastructure already deployed — which is exactly why VodafoneThree’s hardware-first approach is commercially rational. The operator is not buying AI as a standalone product; it is buying a better yield on radios it needed anyway.

That framing should shape operator procurement. AI optimisation software sold as a subscription upgrade to existing platforms has a clear value proposition: incremental performance for incremental opex, with no new capex. AI optimisation that requires a rip-and-replace is a much harder sell. Vendors who can demonstrate gains on installed base will win the first wave.

The autonomy ladder and what it means for procurement

Nokia’s Ricardo Pinto estimated that autonomous networks currently average around Level 2, with the strategic goal being Levels 4 and 5 — substantially more autonomous infrastructure management. That gap is the real roadmap for AI-RAN, and it is a useful procurement anchor.

Operators should ask vendors where their proposed solutions sit on that ladder today and what evidence supports the claim. Level 2 automation is largely rule-based with human oversight. Level 4 implies closed-loop control across multiple functions with minimal intervention. The distance between them is not a software update; it is a sequence of validated capabilities, each requiring field data, testing, and governance.

VodafoneThree’s AI-powered radio optimisation tests are best understood as an early rung on that ladder, executed on real infrastructure. That is the right place to start. Operators who skip straight to autonomy claims without the underlying radio modernisation and data foundations will find themselves debugging physics rather than algorithms.

The latency wall and the compute debate

The week’s news also surfaces the technical fault line that will define AI-RAN architecture: where intelligence runs, and on what silicon.

The E3 interface, as reported, provides a standardised mechanism for dApps to interface directly with Open RAN baseband nodes for closed-loop control beyond the 10-millisecond latency limits of xApps — enabling lower-layer performance optimisation, geolocation, integrated sensing, spectrum sharing, and interference cancellation. That is the open ecosystem’s answer to proprietary real-time optimisation: push intelligence closer to the radio, below the latency wall.

Meanwhile, the compute debate is unresolved. Samsung, per Light Reading, argues that CPUs can carry AI-RAN workloads, citing Intel Xeon 6’s Advanced Matrix Extensions and noting that relevant processing must happen within a transmission time interval of roughly 500 microseconds — which implies lightweight models that may not need a GPU. Nokia and Nvidia, by contrast, position GPUs as the natural AI-RAN platform.

For operators, this is not an abstract architectural argument. It determines power consumption, space requirements, cost, and vendor lock-in. VodafoneThree’s hardware-first move keeps options open on compute, because the radio layer is agnostic to whether the optimisation intelligence eventually runs on CPU, GPU, or specialised accelerators. That optionality is worth more than any single vendor’s benchmark.

Risks and watch items

Integration risk. AI-powered radio optimisation on live networks must be proven under real traffic, real interference, and real weather. Test programmes are not deployments. Watch for published performance data, not trial announcements.

Vendor concentration risk. The Ericsson relationship spans radios, controller, and optimisation software. That delivers speed but reduces multi-vendor leverage. Operators should contract explicitly for open interfaces and data portability.

Measurement risk. AI optimisation gains are notoriously hard to attribute. Operators should insist on baseline measurements and independent verification before committing to subscription pricing.

Market timing risk. With AI-for-RAN reaching only US$1.1 billion by 2030, vendors expecting a windfall will be disappointed. The winners will be those selling incremental value on installed base.

  1. Audit your radio layer first. AI optimisation cannot exceed the capability of the underlying massive MIMO and radio units. Sequence modernisation before intelligence.

  2. Demand evidence on the autonomy ladder. Require vendors to state their current level and the field data supporting it.

  3. Contract for openness. Even in single-vendor AI-RAN programmes, insist on standard interfaces, data export, and exit terms.

  4. Keep compute optional. Avoid architectural commitments that foreclose CPU, GPU, or accelerator choices before the latency and power trade-offs are proven at scale.

  5. Model the economics on incremental yield. Treat AI optimisation as a return-on-installed-base play, not a capex story.

The takeaway

VodafoneThree and Ericsson have done something the industry talks about more than it does: they have put hardware under the AI-RAN ambition. That is less exciting than a forecast of 40% AI traffic or a US$1.1 billion market, and considerably more useful. For Open RAN operators, the lesson is that intelligence without infrastructure is a demo, and infrastructure without intelligence is a missed return. The operators who sequence both correctly — and contract for openness while doing so — will be the ones whose AI-RAN programmes survive contact with the field.

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.