VMO2 Pits Ericsson Against Nokia on AI-RAN: 5 Things the GPU-vs-ASIC Verdict Means for Open RAN Operators
Virgin Media O2 has done something the rest of the industry has mostly talked about but rarely executed at this level of visibility: it is judging Ericsson and Nokia head-to-head on AI-RAN, according to reporting from Light Reading. The framing is deceptively simple. Nokia arrives with GPUs, courtesy of its deepening alliance with Nvidia. Ericsson arrives with its own application-specific integrated circuits (ASICs), which it insists are already AI-capable and already deployed in commercial networks.
The timing matters. This comparison lands in the same week that BT confirmed live AI-RAN trials with Nvidia and Nokia on its EE network in Birmingham for 2027, and as SNS Telecom & IT pegged annual AI-RAN automation and optimisation spending at USD $1.1 billion by 2030. In other words, the AI-RAN conversation has moved from conference keynotes into procurement evaluations. VMO2 is effectively running the bake-off that will shape how operators across Europe think about the silicon underneath their intelligent RAN.
Here are five things this head-to-head means for Open RAN operators.
1. The AI-RAN argument has quietly become a silicon argument
Strip away the marketing and the VMO2 evaluation is not really about whether AI belongs in the RAN. Everyone from AT&T and Verizon to Vodafone, Deutsche Telekom, Swisscom and Telstra has accepted that premise; SNS Telecom & IT’s research tracks adoption across exactly that cohort, alongside operators like Rakuten Mobile, China Mobile, STC and TDC NET that have already reached limited levels of highly autonomous operation for specific live use cases.
The real question is where the compute lives and what it is made of. Nokia’s position, built on its partnership with Nvidia, is that general-purpose GPUs can replace traditional algorithms with AI and deliver meaningful spectral efficiency gains. Nvidia has invested $1 billion in Nokia, and Nokia has been redesigning its 5G products to run on GPUs rather than Marvell’s silicon. Ericsson’s counterargument is that custom ASICs are highly optimised for every RAN task, including the physical layer, the most demanding category of functions.
That is the fork in the road VMO2 is testing. And because the RAN accounts for such a large share of mobile infrastructure cost and energy use, the answer has direct capex and opex consequences.
2. Ericsson’s ASIC claim is not vaporware — but the numbers deserve scrutiny
Ericsson has been unusually specific about its AI-in-silicon progress. 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 roughly 10%.
That is a real, deployed figure — not a lab result. It also sets a concrete bar for the comparison. Nokia, by contrast, has reported improvements exceeding 20% in spectral efficiency during laboratory testing, according to coverage of BT’s planned trials. Lab numbers and live-network numbers are not the same thing, and any operator reading this comparison should treat them accordingly.
There is a subtlety worth flagging. Ericsson’s GPU trials have focused on using GPUs to support only forward error correction (FEC) — described by Nvidia as a “deterministic” function — alongside non-RAN workloads, with all other functions in a virtual RAN scenario running on a CPU. That is a materially different architecture from Nokia’s broader GPU-centric redesign. So when VMO2 compares the two vendors, it is not comparing like with like. It is comparing two philosophies of where intelligence should sit.
3. Nokia’s GPU bet needs telcos to say yes — VMO2 is one of the first to answer
Nvidia’s own framing is instructive. A company spokesperson told Light Reading that GPU-accelerated AI can add substantial value through intelligent bandwidth allocation, traffic prediction, network optimisation and autonomous operations, and that DPUs present opportunities in virtualised broadband gateways, security and programmable networking. Notably, Nvidia said it has no plans for a GPU attack on custom silicon in fiber, arguing the opportunity is “not necessarily to replicate AI-RAN (PHY acceleration) in fixed networks.”
That is a candid admission that the RAN is the hardest target and the one where the GPU case is most contested. Nokia’s commercial future in AI-RAN hinges on operators validating that GPUs can carry PHY-layer workloads at scale without unacceptable power or latency penalties. VMO2’s evaluation is therefore more than a vendor scorecard — it is a referendum on the Nokia-Nvidia thesis.
For Open RAN operators, the stakes are structural. If GPUs win on economics, the RAN becomes a software-upgradeable compute platform with a broad merchant supply chain. If ASICs hold, the incumbents’ silicon advantage persists and the open ecosystem’s differentiation shifts further up the stack.
4. The verdict will shape how operators sequence their 2027 trials
BT’s plan is the clearest template so far. The operator intends to begin live AI-RAN trials with Nvidia and Nokia on its EE mobile network in Birmingham in 2027, assessing whether the lab gains can be reproduced on a live network. BT’s own framing is refreshingly sober: AI-RAN remains an emerging technology and is unlikely to make a material contribution to near-term earnings, though successful deployment could eventually support network efficiency and reduce the need for additional infrastructure investment.
That is the right posture, and VMO2’s comparison should be read the same way. These are not 2027 revenue stories. They are architecture decisions with five-to-ten-year tails. Operators watching from the sidelines — the Vodafones, Deutsche Telekoms and Telstras already named in the automation research — will use early verdicts like VMO2’s to decide whether to run parallel vendor tracks or commit early to one silicon philosophy.
The practical lesson: run the comparison on live traffic, with your own spectral efficiency baselines, and be explicit about whether you are measuring link adaptation alone or the full PHY stack.
5. AI-RAN automation spending is real, but the money follows the architecture
The $1.1 billion annual figure for AI-RAN automation and optimisation by 2030 is not trivial, but it is also not the whole prize. SNS Telecom & IT’s research points to a market shift from older self-organising network technology to more open, application-based architectures built around the RAN intelligent controller and service management and orchestration platforms. In that model, rApps running on the non-real-time RIC layer are emerging as the preferred route to network-wide automation across multi-vendor environments.
That is the quiet Open RAN story inside the VMO2 bake-off. If intelligence lives in rApps above a standardised interface, the underlying silicon matters less to the operator’s long-term flexibility — and the multi-vendor promise holds. If intelligence gets baked into vendor silicon, the RIC becomes a thin veneer over a familiar lock-in.
So the question operators should ask VMO2’s peers is not simply “which vendor won?” but “what did the comparison reveal about portability?” A GPU-based Nokia deployment and an ASIC-based Ericsson deployment can both be excellent. Only one of them, however, is likely to make it easier to swap in a third vendor’s radio or a third-party rApp later.
The bottom line
VMO2’s decision to judge Ericsson against Nokia on AI-RAN is the most concrete signal yet that the GPU-versus-ASIC debate has moved from marketing to measurement. Ericsson has a deployed, quantified 10% spectral efficiency gain to defend. Nokia has lab results above 20% and a 1.1 billion annually by 2030.
For Open RAN operators, the takeaway is to treat vendor AI-RAN claims as hypotheses to be tested on live networks, to insist on clarity about which layers are accelerated, and to keep the RIC and SMO layers open enough that today’s silicon verdict does not become tomorrow’s lock-in.
Sources
- VMO2 judges Ericsson vs. Nokia on AI-RAN — Light Reading
- BT Group (LON: BT.A) shares price forecast for 2026-2028: Analysts see 224p target — Beinsure
- BT Group shares target GBX 220 as AI-RAN plans support momentum — Traders Union
- AI-RAN automation spending to hit USD $1.1bn by 2030 — IT Brief UK
- Nvidia has no plans for a GPU attack on custom silicon in fiber — Light Reading