VMO2 Pits Ericsson Against Nokia on AI-RAN: A Skeptical Exec’s Guide to the GPU-vs-ASIC Decision

VMO2 Pits Ericsson Against Nokia on AI-RAN: A Skeptical Exec’s Guide to the GPU-vs-ASIC Decision

Virgin Media O2 (VMO2) has reportedly put Ericsson and Nokia through an AI-RAN evaluation, according to Light Reading. The story lands at a moment when AI-RAN claims are flying fast—T-Mobile’s live AI-RAN call, Nokia’s GPU redesign, Ericsson’s ASIC counterpunch—and operators are being asked to pick a lane. For a skeptical executive, the VMO2 bake-off is less about who wins and more about what it reveals: the industry’s quiet split over where intelligence should live in the RAN, and whether Open RAN’s promise of multi-vendor flexibility can survive the compute wars.

Below, we walk through the questions a skeptical exec should be asking, answered from the research and reporting available today.

What exactly did VMO2 do, and why does it matter?

Light Reading reports that VMO2 “judges Ericsson vs. Nokia on AI-RAN.” The details are thin—VMO2 has not published a full scorecard—but the significance is clear: a major European operator is formally comparing two fundamentally different approaches to injecting AI into the radio access network. On one side is Nokia, which has partnered with NVIDIA to redesign its 5G products to run on GPUs instead of Marvell’s silicon. On the other is Ericsson, which continues to market its own application-specific integrated circuits (ASICs) as a suitable RAN platform for AI and claims to have AI-capable ASICs already installed in commercial networks.

The evaluation matters because it is a real-world test of a question that has divided the industry: do you need general-purpose GPUs to run AI in the RAN, or can specialized silicon deliver the goods? The answer will shape capex, vendor lock-in, and the pace of AI-RAN innovation for years.

Why is Nokia betting on GPUs, and what does NVIDIA get out of it?

Nokia’s GPU bet is the most visible hardware shift in AI-RAN. Under a deal struck about a year ago, NVIDIA invested $1 billion in Nokia, and Nokia has been redesigning its 5G products to work on GPUs instead of Marvell’s silicon. The logic: GPUs, which became the default for training large language models in data centers, can replace traditional algorithms with AI and boost spectral efficiency. Nokia claims improvements exceeding 20% in laboratory testing, though live-network results are still emerging.

For NVIDIA, the investment is a beachhead into the RAN, a market that has historically been dominated by purpose-built silicon. If Nokia’s GPU-based approach gains traction, NVIDIA becomes a supplier to the telecom infrastructure market at scale. That’s a big if—operators are notoriously conservative about power, cost, and reliability—but the VMO2 evaluation is exactly the kind of validation NVIDIA needs.

What is Ericsson’s counterargument, and is it credible?

Ericsson’s CEO has been vocal about championing his silicon over NVIDIA chips in the RAN. The company argues that its ASICs are already AI-capable and deployed. On a recent earnings call, CEO Per Narvinger said around 60,000 basestations operated by its customers now support an AI successor to link adaptation, a RAN algorithm, and that it is boosting spectral efficiency by around 10%. That’s a tangible, shipping result—not a lab number.

Ericsson’s credibility rests on the fact that it controls its own silicon roadmap and can optimize for RAN-specific workloads. ASICs are generally more power-efficient than GPUs for fixed functions, and they don’t require a wholesale redesign of the baseband architecture. For operators worried about energy costs and physical footprint, that matters. The counterargument is flexibility: if AI models evolve rapidly, a fixed-function ASIC may become a liability, while a GPU platform can be reprogrammed.

What does T-Mobile’s live AI-RAN call tell us about the state of the art?

T-Mobile, Nokia, and NVIDIA recently completed what they describe as the world’s first AI-RAN call routed through a commercial cell site in Bellevue, Washington. The call ran over T-Mobile’s live production network using a Nokia Massive MIMO radio, Nokia’s 5G anyRAN software, and servers built on NVIDIA’s Aerial RAN Computer (ARC) platform. It connected to T-Mobile’s existing 5G Standalone core and a commercial smartphone—not a lab rig.

That’s a genuine milestone, but it’s important to separate the hype from the evidence. The call itself doesn’t prove that AI makes your calls better; it proves that the compute platform can sit in a live network and handle a call. The most concrete performance numbers for the Nokia-NVIDIA platform ironically came from a lab: Deutsche Telekom reported typical uplink throughput gains of 10% to 15% from AI-based channel estimation in initial tests in Bonn and Oulu, with a field trial still to come.

Meanwhile, T-Mobile is hedging. A day before Nokia’s post, T-Mobile’s chief network officer Ankur Kapoor published results from AI-managed beamforming trials with Ericsson in New York City, with the AI model running on Ericsson’s own silicon rather than NVIDIA GPUs. At peak hours, T-Mobile reported a 16% improvement in downlink user throughput and 12% better spectral efficiency. So T-Mobile is backing both horses—a sign that the GPU-vs-ASIC debate is far from settled.

How does Open RAN fit into this? Isn’t AI-RAN supposed to be open?

The tension between AI-RAN and Open RAN is real. Open RAN’s promise is multi-vendor interoperability through open interfaces, allowing operators to mix and match components from different suppliers. AI-RAN, as currently implemented by Nokia and Ericsson, tends to be tightly coupled to the vendor’s compute platform—whether that’s Nokia’s GPU-based anyRAN or Ericsson’s ASIC-based RAN compute. If AI algorithms are embedded in proprietary silicon or software stacks, the openness of the RAN could erode.

That’s why the VMO2 evaluation is worth watching. If VMO2 concludes that both vendors can deliver comparable AI performance, it may push for standardized interfaces that allow AI workloads to run on either platform. If one approach clearly wins, the operator may accept a more integrated solution—and Open RAN’s multi-vendor vision could take a back seat to performance.

What about the smaller players and open-source efforts?

The AI-RAN ecosystem is broader than the two Nordic giants. Parallel Wireless, for example, recently won a deal with Telecall in Colombia to deploy its Unity AI Open RAN solution across thousands of sites, embedding AI in the RAN to adapt to real traffic and radio conditions. The deal includes GreenRAN technology, which pairs AI-enabled Open RAN with energy-efficiency features. That’s a reminder that AI-RAN is not just a developed-market game; it’s being used to extend coverage in underserved areas.

On the open-source front, DeepSig has opened an OCUDU lab in Arlington, Virginia, to accelerate open-source AI-RAN for 6G. The lab gives operators, engineers, and government partners a place to see AI running live on OCUDU. These efforts matter because they could lower the barrier to entry and foster a more diverse supplier ecosystem—something that aligns with Open RAN’s original goals.

What’s the market size for AI-RAN automation, and does it justify the hype?

SNS Telecom & IT forecasts annual spending on AI-RAN automation and optimisation will reach $1.1 billion by 2030. That covers software and services that use AI and machine learning to automate and optimise mobile radio networks. The research points to a shift from older self-organising network technology to more open, application-based architectures built around RAN intelligent controller (RIC) and service management and orchestration (SMO) platforms. In this model, software applications known as rApps, running on the non-real-time RIC layer, are emerging as a preferred route to network-wide automation across multi-vendor environments.

$1.1 billion is not trivial, but it’s also not the multi-billion-dollar bonanza that some AI-RAN narratives imply. It suggests that the near-term opportunity is in software and automation—not necessarily in ripping out and replacing hardware. That’s consistent with the idea that AI-RAN will be an evolutionary overlay on existing RAN investments, at least initially.

Which operators are actually adopting AI-RAN automation today?

Adoption is increasing among operators including AT&T, Verizon, Telus, Vodafone, Deutsche Telekom, Swisscom, and Telstra. The report adds that several operators, including Rakuten Mobile, China Mobile, STC, and TDC NET, have already reached a limited level of highly autonomous operation for specific use cases in live RAN environments. That’s a diverse set—US, European, Asian, and Middle Eastern operators—which suggests that AI-RAN automation is not confined to a single region or business model.

But “limited level” is doing a lot of work in that sentence. Highly autonomous operation for specific use cases is a long way from fully autonomous networks. The gap between the two is where the real engineering and business challenges lie.

What should a skeptical exec take away from the VMO2 evaluation?

First, the GPU-vs-ASIC debate is not academic. It will determine power consumption, cost, and flexibility for years. Nokia’s GPU approach offers reprogrammability and potentially faster iteration on AI models, but it comes with the overhead of general-purpose compute. Ericsson’s ASIC approach offers efficiency and a proven shipping path, but may be less adaptable as AI algorithms evolve.

Second, T-Mobile’s dual-track strategy is a signal. The operator is working with both Nokia/NVIDIA and Ericsson, and its executives describe AI-RAN as “a continuum, not a single product.” That’s a pragmatic stance that other operators may emulate—at least until the economics become clearer.

Third, the Open RAN dimension is at risk of being sidelined. If AI-RAN becomes a vendor-specific compute platform war, the multi-vendor, interoperable vision of Open RAN could be compromised. Operators who care about openness should demand standardized interfaces for AI workloads and avoid locking themselves into a single vendor’s silicon roadmap.

Fourth, the numbers are still early. Lab results of 10-20% spectral efficiency gains are promising, but live-network results are only beginning to emerge. The business case will hinge on whether those gains translate into real capex or opex savings—and whether they can be achieved without a forklift upgrade.

What’s the bottom line for operators weighing their options?

The VMO2 evaluation is a useful bellwether, but it’s not a verdict. The AI-RAN market is still in its early innings, and the technology is evolving rapidly. Operators should watch the VMO2 outcome closely, but they should also run their own trials, insist on open interfaces, and avoid betting the farm on a single compute architecture until the evidence is overwhelming.

In the meantime, the $1.1 billion automation market by 2030 suggests that the software layer—rApps, RIC, SMO—may be the safest bet. It’s less sexy than GPUs in the tower, but it’s where the near-term returns are likely to be. And it’s where Open RAN’s multi-vendor principles can still be preserved.

For a skeptical exec, the question is not “GPU or ASIC?” but “How do I keep my options open while capturing the efficiency gains?” The operators that answer that question well will be the ones that emerge from the AI-RAN hype cycle with a network that is both smarter and more open.

Sources

Related Posts

VMO2 Pits Ericsson Against Nokia on AI-RAN: 5 Things the GPU-vs-ASIC Verdict Means for Open RAN Operators

VMO2 is benchmarking Ericsson and Nokia on AI-RAN, pitting GPU-based spectral efficiency against custom ASIC silicon. Here are 5 things operators should weigh.

The Week AI-RAN Grew Up: T-Mobile's Live Call, O2's Paddington Trial, and the $1.1 Billion Question

From a live AI-RAN call in Bellevue to a 20% spectral efficiency gain at London Paddington, AI-RAN's proof points are arriving faster than its business case.