Supermicro, Nokia and NVIDIA Push AI-RAN From the Cell Site to the Core
The three vendors say running radio and AI workloads on shared infrastructure lifts spectral efficiency and opens edge services. The servers ship now; the operator economics remain unproven.
Supermicro, Nokia and NVIDIA used NVIDIA’s GTC Berlin event to announce that they will deploy AI-RAN, or artificial-intelligence radio access networks, from distributed cell sites through aggregation points to core network hubs. The stack combines Nokia’s RAN software with NVIDIA’s Aerial RAN Computer platform on Supermicro’s servers. The pitch is that operators can run radio and AI workloads on the same infrastructure, which lifts spectral efficiency, automates network operations, cuts energy use, and opens a path to selling edge AI services such as voice AI, vision AI and physical AI.
The hardware ships now. Supermicro says its servers for NVIDIA ARC are generally available, with one configuration arriving later this year and another, the ARC-Pro, moving to customer testing before general availability in the first half of 2027. The announcement landed alongside a wider NVIDIA and Nokia push in which T-Mobile completed what the companies call the world’s first AI-RAN call on a live commercial network, and several European operators announced trials. The question that matters for an operator is not whether AI-RAN works, but which parts of it pay.
What the three companies announced
The solution spans three network layers, and Supermicro right-sizes the compute for each. At distributed cell sites, where proximity to radios and users matters most, it offers the ARC MGX 2U short-depth platform, a dual-node ARC-Pro that pairs an NVIDIA Grace C1 CPU with an RTX PRO 4500 Blackwell GPU per node in a 429 mm chassis and delivers up to 1.6 PFLOPS of FP4 performance, and the ARC-Compact for edge networking functions such as service management and orchestration.
At centralised MSO and aggregation sites, the ARC MGX 2U handles higher-throughput coordinated inference, with a choice of dual NVIDIA Vera CPUs or dual x86 processors and two RTX PRO 6000 Blackwell GPUs, and it supports multi-cell workloads including coordinated scheduling and beamforming. And, t core hubs, the ARC MGX 4U packs up to eight RTX PRO 6000 Blackwell GPUs, dual x86 CPUs, up to 6 TB of DDR5 memory and an optional ConnectX-8 switchboard with PCIe Gen 6.
Nokia Supplies the RAN Software
The three companies split the work in a way that says something about the market. Nokia supplies the RAN software, including the CUDA-accelerated anyRAN product it has developed with NVIDIA; NVIDIA supplies the Aerial RAN Computer platform and the accelerated computing underneath it; and Supermicro supplies the servers and the modular designs that package them for different parts of the network. NVIDIA also announced two ARC configurations of its own, the MGX 2U and MGX 4U, and said Dell Technologies, Quanta Cloud Technology and Supermicro will offer ARC-ready servers, so Supermicro is one of several system builders rather than the sole supplier. The announcement arrived at GTC Berlin, NVIDIA’s European developer event, which has become the venue for its telecom and AI infrastructure news.
Shared infrastructure is the harder half of the idea. A radio unit has strict requirements for latency, availability and determinism, and a large language model has none of those, so running both on one server demands careful separation. The industry’s answer is to isolate AI workloads from the RAN protocol stack through the user plane, to keep them in separate trust domains, and to schedule compute so that an inference job cannot delay a radio function. Supermicro’s stack addresses this by matching system size to location, but the harder question is software: which scheduler decides when an AI workload may use capacity the RAN might need, and how the operator enforces that boundary. The release describes the hardware and the intent, not the mechanism.
What AI-RAN means, and which part pays
The industry divides AI-RAN into three parts, and the distinction explains where the near-term money sits. AI-for-RAN uses AI to improve the radio network itself, through better channel estimation, scheduling, mobility and energy management. AI-and-RAN shares infrastructure between radio and AI workloads, so idle compute does something useful. AI-on-RAN puts AI services at the network edge, close to users and devices, to create new offerings. Supermicro’s release leans on all three, but the measured evidence the industry has published so far comes almost entirely from the first.
The measured gains, and where they come from
That evidence is worth setting out, because it is more modest than the marketing implies. Optus, working with Ericsson, reported spectral efficiency improvements reaching 10 per cent across trial sites and 25 per cent under certain conditions. Rakuten Mobile reported energy savings exceeding 20 per cent on its virtualised network, along with TM Forum Level 4 autonomous network validation for the energy case. Samsung’s Energy Saving Manager achieved 35 per cent total energy reduction in a United States commercial deployment, and its RAN Speed Optimizer lifted average downlink throughput by up to 12 per cent on a live 5G network. Turkcell makes the economic point plainly: better use of existing capacity can delay capital spending. These are real numbers from real networks, and they all come from optimisation rather than from new services.
The edge-services case, which is where the new revenue is meant to come from, remains unproven. Analyst firm ABI Research expects AI-RAN to generate more than 6.1 billion dollars in investment revenue by 2032, but notes that most of it will not arrive until 2029 and that the shortage of operators among AI-RAN Alliance members signals some industry scepticism about the near-term value. Its analyst puts the condition bluntly: real growth will come only when performance benchmarks are validated in the field. The AI-RAN Alliance itself, which has grown past 140 member organisations, now says its priority is implementation rather than demonstration, and its chair notes that not every AI model should run at a cell site and not every workload needs a GPU.
The hardware, and the operators
The Supermicro stack reflects that caution by matching compute to location rather than putting the largest system everywhere. The largest configuration sits at core hubs, where power and cooling are easiest and multi-cell workloads justify the density, while smaller systems sit closer to radios. On the operator side, the evidence is mostly trials. T-Mobile’s AI-RAN call ran on commercial ARC servers, Nokia’s CUDA-accelerated RAN software, massive MIMO radios and an existing 5G standalone core.
Elisa ran voice, mobile data and vision AI together on one accelerated platform in a European field first. BT plans live trials in Birmingham in 2027, Deutsche Telekom plans field trials in 2027 after lab validation, and Orange is testing spectral efficiency on live networks. Roughly 20 operators have announced AI-RAN work with Nokia and NVIDIA, which is meaningful interest and not yet commercial scale.
The operator trials also show how early the deployment still is. T-Mobile’s call proves that the pieces interoperate on a commercial network, which is a real milestone, but a single call is not a service. Elisa’s field trial shows that voice, data and vision AI can share one platform, which is the AI-and-RAN case in practice. BT and Deutsche Telekom have both scheduled their live trials for 2027, which means the largest European operators will not have production evidence for at least a year. The roughly 20 announced collaborations describe intent rather than scale, and none of the operators has published a commercial AI-RAN service with pricing.
What operators should ask
For an operator weighing it, the questions are practical. Which workloads actually need to run at the cell site, and which belong in a core hub? What spectral efficiency or energy gain can we measure on our own network, against a clear baseline? How do third-party AI models stay isolated from critical RAN functions, and who carries the risk if they interfere? What does the compute cost per site, and how does that compare with the capacity or energy it saves? What is the software upgrade path to 6G, and does it protect the hardware we buy now? And what evidence exists that customers will pay for edge AI services, since that is where the growth case rests?
What it signals
Supermicro, Nokia and NVIDIA have built a coherent, shipping stack for a real idea: that radio and AI workloads can share infrastructure, and that the network edge is a sensible place to run inference. The near-term value is credible because operators have measured it, and it comes from doing the same job better rather than from selling something new. The longer-term value, the edge-services business that would justify much larger investment, still rests on assumptions that no operator has yet validated at scale. That is the honest state of AI-RAN: a technology with proven efficiency gains and unproven new revenue, shipping into networks that need the first and hope for the second. That gap between capability and commerce defines the moment. Vendors have shipped the hardware and proven that the technology works in trials, and operators have measured the efficiency gains on their own networks.
What nobody has demonstrated is a customer paying for an AI service delivered from the radio edge, and that is the revenue that would justify building AI-RAN at scale rather than optimising existing networks with it. The near-term case is strong enough to buy, and the long-term case is strong enough to keep watching. An operator deciding today can bank the efficiency gains and treat the edge-services business as optional, which is exactly what most of them appear to be doing.