0

Open and Cloud RAN re creating the programmable foundation for AI-RAN, with spectral efficiency, automation and orchestration emerging as key steps toward AI-native 6G networks

The radio access network (RAN) is entering a period of architectural convergence as AI-RAN takes shape. Open RAN, cloud-native infrastructure and artificial intelligence (AI) have largely developed as separate industry movements, but they are increasingly becoming components of the same transition toward a more programmable, adaptive and software-driven RAN.

That makes the current AI-RAN cycle the latest stage in a much longer evolution of network automation. Self-organizing networks provide useful historical context. SON developed alongside LTE as operators sought to automate optimization and reduce manual intervention across increasingly complex radio networks. By the middle of the last decade, machine learning algorithms were already being incorporated into SON functions, extending optimization beyond static, rules-based approaches.

Stéphane Téral, founder and chief analyst at Téral Research, described the longstanding objective as creating systems that “minimize human interaction to the lowest level possible.”

Open RAN subsequently expanded the architectural surface available for that intelligence. Disaggregating hardware and software, opening interfaces and introducing the RAN Intelligent Controller created new opportunities for supplier interoperability and software-based innovation through rApps and xApps.

The significance of openness, however, extends beyond building a RAN with the largest possible number of vendors. It creates architectural optionality: the ability to introduce different radios, software and applications where they deliver value.

Cloud RAN pushes the same transition further by virtualizing and containerizing baseband functions and moving them onto increasingly common compute infrastructure. As Téral put it: “You start by disaggregating software from hardware. Then you want to run functions on virtual machines. At some point, you containerize and then you move to the cloud.”

This is where the different technology trajectories begin to reinforce one another. Open interfaces create programmability. Cloud-native infrastructure increases software velocity and deployment flexibility. AI increases the sophistication of the decisions that software can make.

The near-term AI-RAN opportunity is centered primarily on improving the performance of the RAN itself. AI-driven algorithms can increase spectral efficiency, optimize network resources and support more advanced automation. Operators are also applying AI to root-cause analysis, intent-based optimization and closed-loop network operations.

But AI-RAN does not necessarily mean deploying GPUs at every cell site. Compute architecture will depend on workload, geography, latency requirements and economics. The relevant question is where additional intelligence produces enough performance or operational value to justify the infrastructure required to support it.

That question becomes more complicated as the industry moves toward shared AI and RAN infrastructure. Over time, radio workloads and AI applications could operate on common distributed compute platforms, extending intelligence closer to users and devices.

That makes orchestration one of the critical problems ahead. Operators will need to determine what workloads run where, how compute capacity is allocated and how increasingly heterogeneous infrastructure can be managed without overwhelming operational complexity.

This is also where today’s AI-RAN work begins to point toward AI-native 6G. The longer-term destination is not simply a RAN with more AI applications layered on top. It is a network in which connectivity, compute, software and intelligence are increasingly coordinated as a single system—capable of observing conditions, interpreting intent and continuously adapting resources around network, business and customer outcomes.

The AI-RAN outlook — operator priorities, vendor divergence

Nokia and Ericsson broadly agree on the destination for the radio access network (RAN): more intelligence embedded throughout the RAN, greater automation, improved use of scarce spectrum and, ultimately, infrastructure capable of supporting increasingly AI-centric traffic and services. Where they are beginning to diverge is on how operators should get there and, more specifically, how much the underlying RAN compute architecture needs to change.

Ericsson’s near-term proposition centers on embedding AI directly into existing RAN software and purpose-built hardware. Nokia is making a broader architectural bet around AI-accelerated computing, shared infrastructure and a programmable RAN platform designed to evolve toward AI-native 6G. The distinction is not absolute; Ericsson supports cloud and accelerated-compute architectures, while Nokia offers evolutionary paths for its installed base, but the center of gravity is increasingly different.

T-Mobile US makes the comparison particularly relevant. The operator is working with Ericsson on AI-native RAN software deployed against live commercial traffic while also working with Nokia and NVIDIA on GPU-accelerated AI-RAN infrastructure capable of running radio and AI workloads together. In effect, T-Mobile US is simultaneously testing two visions of where AI creates value in the RAN.

Ericsson: Put AI into the RAN operators already have

Ericsson’s approach begins with a straightforward operator proposition: improve network performance without requiring a wholesale change in infrastructure.

Its AI in RAN software suite embeds models directly into basebands and radios, using existing AI-ready Ericsson hardware rather than requiring operators to deploy GPUs at specific sites. Ericsson’s broader portfolio distributes intelligence across radios, RAN Compute, Cloud RAN and rApps, with its purpose-built silicon supporting AI inference close to the radio functions it is optimizing.

Ericsson and T-Mobile US moved AI-native Scheduler with Link Adaptation into large-scale trials on a live 5G Advanced network. The neural-network-based scheduler runs on Ericsson hardware and predicts changing radio conditions in real time. Ericsson reported close to a 10% improvement in spectral efficiency and up to a 15% increase in downlink throughput compared with legacy rule-based scheduling.

That is important because it makes the AI-RAN business case legible in familiar operator terms. Spectrum is expensive, capacity is valuable, and software that produces more throughput from already-deployed spectrum and infrastructure can be evaluated against a relatively conventional return-on-investment model.

The strategy also reflects the practical reality of the 5G investment cycle. Operators have already spent heavily on radios, basebands, spectrum and modernization. Ericsson’s message is that AI-RAN does not have to begin with another major infrastructure replacement. Its software can instead increase the intelligence — and potentially extend the economic value — of hardware already in the network.

That does not mean Ericsson is rejecting accelerated compute or Cloud RAN. Ericsson and T-Mobile have also demonstrated Cloud RAN software running on NVIDIA infrastructure. Rather, Ericsson’s immediate commercial emphasis is on applying the right AI model at the right point in the existing network architecture, with purpose-built silicon handling inference where it makes economic and technical sense.

Nokia: AI-RAN as an architectural transition

Nokia starts from a more expansive premise: AI is not simply another software capability to be added to the RAN; over time, it changes what the RAN itself needs to become.

Its AI-RAN proposition combines its anyRAN software with AI-accelerated computing and offers three deployment paths. Operators can add an AI-accelerated plug-in card to existing AirScale basebands, deploy dedicated accelerated AI-RAN nodes or move to cloud-native AI-RAN running on commercial off-the-shelf servers. All three sit on a common software foundation and roadmap toward 6G.

That range of options communicates to operators that AI-RAN requires an immediate move to a fully GPU-based network. But the architectural direction is explicit: RAN functions and AI workloads increasingly run on a shared compute foundation, creating infrastructure that can improve radio performance while also providing compute for AI applications.

In discussion with RCRTech, Nokia Vice President and Head of AI-RAN and Cloud RAN Aji Ed characterized the immediate value proposition around spectrum. “We bring twice the network capacity, twice the spectral efficiency compared to what we have today,” he said, describing a roadmap in which increasingly sophisticated algorithms and compute architectures drive progressive efficiency gains.

Nokia said it has already demonstrated more than 20% efficiency gains from AI-driven radio innovations and sees advanced AI models pushing spectral efficiency toward more than 2x current levels. Crucially, however, Nokia pairs that performance claim with a change in development cadence: a software-defined RAN that can continue improving through new models and algorithms rather than waiting for traditional hardware refresh cycles.

Ed described that shift as a move toward an architecture that “runs at the software speed, not just dependent on the silicon speed.”

T-Mobile US points to an operator-led synthesis

T-Mobile’s strategy suggests that operators may not view the emerging vendor approaches as mutually exclusive.

In an interview with RCRTech, T-Mobile Chief Network Officer Ankur Kapoor repeatedly framed AI around customer outcomes rather than technology deployment. “No automation is good enough unless it’s really impacting customer experience,” he said. “If it’s not delivering a differentiated customer experience, it’s only operational benefits.”

That framing helps explain why T-Mobile US can pursue Ericsson’s incremental software model and Nokia’s more architectural AI-RAN proposition at the same time.

Ericsson offers something operators can measure now: AI changing scheduling and radio optimization on infrastructure already carrying commercial traffic. Nokia’s proposition points toward what comes next: turning the cell site into part of a distributed compute architecture capable of processing both connectivity and AI workloads.

Kapoor’s own description of the future closely tracks that broader trajectory. T-Mobile US, he said, is trying to bring intelligence into cell sites that historically have simply “moved bits and bytes,” enabling them to support radio and AI workloads simultaneously.

“I’m not going to say that every tower is going to become a robot, but every tower is going to become an actual intelligent fabric,” Kapoor said. That local intelligence can feed T-Mo’s existing self-organizing network and Dynamic CX platforms, allowing distributed compute and centralized automation to work together.

The emerging Nokia-Ericsson divergence, then, is less about whether AI belongs in the RAN than where operators should begin and how quickly the compute architecture should change.

Ericsson is showing how AI-RAN can create incremental, quantifiable value inside today’s network. Nokia is positioning AI-RAN as the beginning of a deeper architectural transition in which connectivity, accelerated computing and distributed AI increasingly converge.

For operators, the answer may ultimately be both. The near-term requirement is better spectrum utilization, lower cost and improved customer experience. The longer-term opportunity is to turn widely distributed RAN infrastructure into an intelligent compute fabric. T-Mobile’s work with both vendors suggests the path to AI-native 6G will likely require operators to capture the first set of benefits without losing sight of the second.

Deep dive into Nokia’s AI-RAN strategy

As telecoms operators look to leverage the AI inflection point as a way to rethink operations and grow revenues, Nokia, as we touched on earlier in this article, sees the significance of AI-RAN as more than a way to make the radio access network perform better. The larger proposition is that AI changes what the RAN itself needs to become.

That distinction puts Nokia toward the more architectural end of the emerging AI-RAN spectrum. The immediate business case remains grounded in familiar operator priorities of spectral efficiency, capacity, power consumption and cost per bit. But Nokia is using those near-term benefits to support a broader transition from hardware-defined RAN infrastructure toward a programmable platform where radio and AI workloads increasingly share compute resources.

The company describes its AI-RAN platform as available for 5G and 5G-Advanced, software-upgradable toward 6G and fully Open RAN compliant. It supports three deployment paths: adding AI-accelerated capacity to existing AirScale deployments, introducing dedicated accelerated AI-RAN nodes, or deploying cloud-native AI-RAN on COTS infrastructure. The common denominator is Nokia’s anyRAN software stack and a shared software roadmap.

The range of options matters because Nokia is simultaneously advocating an architectural shift and acknowledging brownfield reality. Operators cannot forklift networks every time the compute model changes. Ed, the vice president and head of AI-RAN and Cloud RAN at Nokia, framed that pragmatism around placement of compute rather than attachment to a particular processor architecture. “It’s about bringing the right compute at the right place, and with the right configuration,” Ed said.

That creates an evolutionary path. Existing AirScale deployments can gain AI capability through a plug-in card; operators requiring greater AI headroom can use an accelerated AI-RAN node; and workloads demanding maximum flexibility can move onto cloud-native infrastructure. Nokia says all three approaches lead toward the same 6G destination.

The near-term rationale is much less abstract: operators have spent billions acquiring spectrum, making any improvement in bits per second per hertz economically meaningful. Nokia has reported more than 20% efficiency gains from existing AI-driven radio innovations and sees advanced AI models operating at radio timescales pushing spectral efficiency toward more than 2x conventional architectures.

Those gains will not arrive uniformly. Ed emphasized that different algorithms apply to different radio configurations, traffic patterns and network conditions. AI-based multi-user MIMO pairing, channel estimation and link adaptation, for instance, can produce different results depending on whether a site uses TDD massive MIMO, FDD or other configurations. The highest returns may also come at congested sites where additional capacity has the greatest economic value.

But the more important architectural argument is what happens after the first efficiency gains are realized. Traditional RAN innovation remains tied partly to silicon refresh cycles. Nokia’s AI-RAN proposition moves more of that innovation into software, allowing new algorithms and models to change network performance without corresponding hardware replacements.

The longer-term goal of Nokia’s AI-RAN strategy is more ambitious. The RAN is among the most geographically distributed compute footprints in existence. If operators can combine radio processing with AI inference on common accelerated infrastructure, that footprint potentially becomes an edge AI platform rather than infrastructure dedicated solely to connectivity.

Nokia describes that end state as a programmable, AI-native RAN in which RAN functions and AI workloads operate on a shared computing foundation. Generative, agentic and physical AI could also change traffic itself — increasing uplink requirements, demanding lower latency and creating more need for distributed processing.

That direction aligns with a broader shift in telecom AI. The central challenge is no longer demonstrating that AI can perform useful tasks; it is introducing probabilistic systems into networks engineered around deterministic behavior, reliability and accountability. The practical industry response is to phase AI into bounded domains with appropriate guardrails rather than wait for some future moment when autonomy becomes risk-free.

Ed makes a similar distinction between demonstration and deployment. “Having a demonstratable capability is one thing versus deploying this capability in a commercial grade scale,” he said, noting that models may require continuous fine-tuning and even adaptation to site-specific conditions. And accelerated computing has to satisfy basic operator economics. Ed reduced the test to “the power, the price, and the performance”: power and price need to remain comparable while performance improves.

The next problem is orchestration. Shared AI-RAN infrastructure adds another resource-management layer: operators must coordinate onboarding, lifecycle management and allocation of both AI and RAN workloads across heterogeneous distributed infrastructure. Ed acknowledged that this creates additional complexity even as Nokia works with partners on AI-RAN orchestration.

That tension captures Nokia’s larger AI-RAN bet. Spectral efficiency provides the business case to begin. Software-defined infrastructure creates a mechanism for continuous improvement. Accelerated computing expands what can run in the RAN. And orchestration ultimately determines whether operators can turn those capabilities into a manageable platform.

As the broader AI-RAN market develops, Nokia is effectively arguing that the prize is bigger than a smarter scheduler. The end state is a RAN that continuously improves through software and increasingly participates in the AI economy — an architectural transition that begins in 5G, accelerates through 5G-Advanced and becomes foundational to AI-native 6G.

TELUS sees Open RAN as a key enabler for AI

For TELUS, Open RAN is increasingly about more than disaggregation or supplier diversity. The Canadian operator’s brownfield transformation illustrates how openness, virtualization and programmability can create the foundations for a more intelligent network — and, ultimately, for AI-RAN.

The starting point was unusually pragmatic. TELUS was already facing a major equipment transition driven by hardware lifecycle requirements, changes to Canada’s telecom supply-chain rules and its evolving vendor strategy. Rather than simply replace existing equipment, the operator used that refresh cycle to change the RAN architecture itself.

“We were in our hardware refresh cycle, and we were anyway swapping our network,” TELUS Director of RAN Strategy Sushil Rawat said. “So it helps when you’re riding a ongoing hardware refresh cycle and just change the architecture because then the technology itself does not cost more to do that truck rolls and and make those changes in the network.”

That alignment speaks to the economics of brownfield Open RAN. Replacing equipment with significant residual value can undermine otherwise compelling TCO assumptions. For TELUS, architectural transformation became part of work that already had to happen. As Rawat put it, “It’s it’s out of necessity. It’s not just out of choice of a technology that you make.”

TELUS’ resulting architecture is deliberately multivendor. Samsung provides key elements including virtualized RAN software, radios, orchestration capabilities and the RIC platform, while the deployment also incorporates third-party radios, COTS servers, an independent cloud layer and applications from multiple suppliers.

The important point is that TELUS sees openness as something that has to work operationally, not merely exist in a specification. “We have at least kept this principle where we will have openness in the architecture, and not only on paper, but also exercise having interoperability with third party,” Rawat said.

That architecture is now scaling quickly. Rawat said roughly 25% of the network was running Open RAN as of July 2026, with TELUS targeting approximately 40% by year-end, 50% by the end of 2027 and 100% by the end of 2029.

But the more consequential shift may be operational. TELUS’ original Open RAN business case included the SMO, RIC, tool consolidation and simplification of its automation environment. Those same architectural elements create natural insertion points for AI.

Rawat described current AI-for-RAN work as an evolution of automation. TELUS is applying smaller AI agents to repeatable operational tasks while treating less deterministic activities more cautiously. “I would call it an extension of automation, which is more self-defined,” he said. “Now, with the help of AI, I think it is becoming more reusable and easier to configure an AI agent to to develop certain use case.”

For TELUS, however, AI-RAN ultimately has to translate into measurable network performance. “When it comes to from my perspective, AI RAN is how does AI increase the efficiency of the RAN stack that is deployed, right?” Rawat said. “How do we deliver more data? How do we deliver more data using the same spectrum asset that you have?”

That outcome-driven approach extends to accelerated compute. Rawat said he does not currently see a need to deploy GPUs at cell sites over the next 12 months, although TELUS is already using accelerated computing centrally for model training and anomaly detection. If new technology delivers performance unavailable today, however, “There might be an opportunity to do that.”

The harder problem is governance. Once AI agents move from analysis to making changes in a production network, TELUS sees identity management, access control, conflict management and established change-management processes as essential.

“You can build a use case. You can demonstrate it in lab, quick and easy,” Rawat said. “Taking it to the production, scaling it for day-to-day operation. This is a very important aspect of it.”

TELUS’ journey points to a broader relationship between Open RAN and AI-RAN. Openness does not automatically make a network intelligent. But by separating hardware and software, introducing programmable control layers and enabling multivendor applications, Open RAN creates an architecture in which intelligence can be introduced more rapidly — and progressively moved from recommendation toward autonomous action.