AI-RAN: Bringing software-speed innovation into the radio network

In our recent blogs, we described how AI is reshaping traffic, why autonomy must be governed, and how networks are beginning to act more like distributed nervous systems. Those shifts now converge in the Radio Access Network.

Nokia’s AI-RAN initiative begins with a simple observation: AI is changing not only how networks are operated, but also the nature of the traffic they carry. AI workloads have already reached massive scale, with mobile devices accounting for more than half of AI interactions. Large language model interactions introduce richer uplink flows and burstier patterns as devices continuously send context to models.

Agentic systems take this further by breaking the traditional rhythm of human-driven network usage. Agents don’t sleep. Neither can the networks that serve them.

Physical AI raises the bar again. When AI begins guiding robots, vehicles, and machines interacting with the physical world, connectivity must support flows that are not only latency-sensitive but also jitter-sensitive and deterministic.

AI is not simply increasing traffic volume. It is changing the shape of network demand.

Networks must become far more adaptive, capable of improving and updating themselves continuously. And that pressure lands first on the RAN. Even before AI-driven applications emerged, the radio system had already become one of the most complex parts of the network. Its control loops operate within strict real-time windows while balancing constantly changing radio conditions. AI-shaped traffic and AI-driven applications add further demands. The next stage of radio evolution therefore cannot rely solely on more sophisticated heuristics. The RAN must begin incorporating learned intelligence directly into its decision making. To enable this, it is imperative that software and algorithmic innovations are decoupled from hardware cycles.

Innovation at the speed of software

Decoupling innovation from hardware

Historically, many gains in radio systems came from tightly integrated hardware and software design. That model delivered efficiency, but it also tied innovation to hardware refresh cycles.

AI models evolve differently. They improve through experimentation, validation, deployment, rollback, and retraining. Their lifecycle runs at a much more rapid software-oriented cadence.

AI-RAN therefore begins with hardware and software decoupling. Nokia enables this through a hardware abstraction layer (HAL) that separates software innovation from the details of the underlying compute implementation. HAL allows algorithms, models, and optimization logic to evolve continuously in software while preserving a stable execution foundation across different hardware architecture. Hardware defines the available computational horsepower. Software drives the evolution. AI-RAN is not simply a feature added to the RAN; it fundamentally and irreversibly changes how the RAN will evolve.

The RAN as a programmable platform

Once intelligence becomes part of the execution path, the RAN can no longer be treated as a fixed-function pipeline. It must become a programmable platform capable of hosting, governing, and evolving intelligent workloads safely and deterministically. This platform spans two tightly coupled domains.

The first is the execution domain, where radio decisions such as scheduling, link adaptation, and interference management occur within deterministic timing windows measured in microseconds.

The second is the learning domain, where models are trained, validated, deployed, monitored, refined, and, when necessary, withdrawn.

This separation allows Nokia and the ecosystem to innovate continuously without compromising runtime stability. It enables faster progress in areas such as channel estimation, link adaptation, scheduling, beamforming, and spectral-efficiency optimization, where better algorithms translate directly into measurable network gains at software speed. Bringing AI into the RAN is therefore not only about inference. It is equally about lifecycle, governance, and control.

Injected, not fused

Historically, radio innovation often came from fusing new algorithms directly into the execution path. That produced highly optimized systems, but also rigid ones.

Learned intelligence evolves differently. If intelligence is fused into the base platform, its evolution becomes constrained by the lifecycle of the underlying system. Nokia’s AI-RAN therefore adopts an injection model rather than a fusion model. Models can be introduced, validated, replaced, or removed independently of hardware generations and base platform software. This preserves baseline stability while allowing intelligence to evolve rapidly. 

Injection is not only about flexibility. It is also about trust. In our earlier blogs, we argued that AI-native networks must remain glass boxes rather than black boxes. The same principle applies here. Injectable intelligence enters the RAN through governed interfaces with clear boundaries around inputs, outputs, and operational impact. Operators can introduce intelligence incrementally while maintaining observability, security, and control. 

An open ecosystem for innovation

The RAN was always parallel. Now the platform is too.

To bring learned intelligence into the radio network, the platform must perform significantly more computation within the same deterministic timing windows. Many radio signal-processing pipelines are inherently parallel. Channel estimation, beamforming, MIMO detection, and forward-error correction all involve large vector and matrix operations executed concurrently across many data elements, such as antennas, subcarriers, OFDM symbols, spatial streams, and active users.

Nokia’s AI-RAN architecture builds on this by introducing a software-driven parallel compute substrate into the radio system. Given Nokia’s focus on innovation at the speed of software, on a programmable platform architecture, and on alignment with the global AI ecosystem, our partnership with NVIDIA and our joint work around ARC-Pro were natural ones.

Today, the vast majority of AI innovation occurs on GPU-accelerated platforms built around modern AI frameworks and execution stacks. These execution stacks now support wireless systems by providing runtime foundations and kernel libraries that bring telco-grade, latency-bounded processing to programmable compute platforms. Nokia builds on this broader ecosystem with its own carrier-grade RAN software implementation and system expertise. By aligning AI-RAN with the AI ecosystem, Nokia makes the radio network accessible to one of the largest developer communities in computing while preserving carrier-grade RAN software boundaries and control.

On top of these shared foundations, two software paths can coexist. An open-source reference Layer-1 environment is maintained that allows developers and researchers to explore new algorithms and AI-native techniques on programmable platforms. In parallel, Nokia delivers its own carrier-grade RAN software stack for commercial deployment. Because both paths build on the same programmable substrate, innovations developed in the reference environment can transition naturally into Nokia’s telco-grade implementation, where they are integrated, optimized, and hardened for production networks. This enables Nokia’s AI-RAN platforms to accelerate both classical signal processing and AI models within the same programmable infrastructure.

This is how Nokia expands the innovation surface of the radio network. Developers can introduce new models through governed interfaces, and any inference workload that fits within the available compute envelope can potentially run alongside radio functions.

In this way, we are creating an open innovation platform for AI-RAN: one that aligns with the AI ecosystem, preserves telecom-grade guardrails, and allows innovation from across the ecosystem to enter the radio network.

Layers of gain

In Nokia’s AI-RAN architecture, this programmable substrate enables multiple layers of improvement.

First, it expands the computation available within real-time radio windows. More candidate decisions can be evaluated and richer context incorporated without missing the deadline. Scheduling illustrates this clearly. Decisions that once relied on heuristics to balance channel conditions, interference, mobility, and latency sensitivity can now incorporate richer context and more sophisticated algorithms while still meeting deterministic timing constraints.

Second, it enables advanced algorithms previously impractical under strict timing constraints. Techniques such as RKHS-based channel estimation already demonstrate spectral-efficiency improvements of 5–10% in challenging propagation environments, becoming practically deployable at scale thanks to the massive parallelism available in GPU-accelerated platforms.

Third, it enables learned models to participate directly in radio decision loops. Because the substrate is programmable, these models can be deployed, measured, and improved at software cadence rather than deferred to the next generation.

Early field trials already demonstrate the compounding effect: up to 25% user-throughput gains from learned channel estimation, up to 11% improvements from optimized multi-user MIMO pairing, and up to 12% improvements in secondary-cell selection for carrier aggregation. These are not theoretical projections; they are gains already demonstrated in real-world networks and available through software on AI-RAN infrastructure. Once the RAN becomes a programmable platform, these gains begin to compound.

Under telco guardrails

Nokia’s AI-RAN strikes the right balance at both the hardware and software layers. At the hardware layer, our substrate remains compatible with the broader AI ecosystem while respecting telco realities: power consumption, thermal envelope, and total cost of ownership. At the software layer, we align with mainstream AI frameworks while extending them with the real-time kernels, controls, and guardrails required for telecom-grade operation. This evolution preserves openness, aligning with 3GPP standards and open interface principles advanced by O-RAN.

Nokia AI-RAN realization

Nokia is bringing this architecture to life through its AI-RAN initiative. In partnership with NVIDIA, Nokia combines NVIDIA’s AI-RAN software foundations with Nokia’s radio expertise, carrier-grade RAN software, and system integration capabilities. Nokia’s ARC-Pro-based platform provides GPU-accelerated infrastructure optimized for telecom environments, meeting the power, thermal, performance, and TCO constraints that operators demand. These capabilities integrate across Nokia’s portfolio through both COTS-based deployments and AirScale-based solutions, allowing operators to introduce AI-RAN within existing network architectures.

This is how Nokia brings software-speed innovation into the radio network while preserving the performance, openness, and determinism operators require.

Beyond optimization

AI-RAN is not only about improving radio performance. It also enables new forms of intelligence in and around the network.

Integrated Sensing and Communications (ISAC) illustrates this potential. At MWC, Nokia demonstrated scenarios where radio infrastructure becomes a sensing platform, detecting human presence, tracking drones, and deriving weather conditions directly from RF signals. The RAN is no longer just connecting intelligence. It is becoming part of it.

The same substrate also enables operators to host AI inference workloads alongside radio functions at the network edge, opening the door to new monetization opportunities as AI services move closer to users and devices.

Only the beginning

The gains available from AI-RAN will not come from one single algorithm or one single loop. They will come from multiple layers of improvement across scheduling, signal processing, control, and inference, compounding across the system.

This is where the next levels of spectral efficiency, energy efficiency, and adaptability will come from. And we have only scratched the surface. At Nokia, we are building our AI-RAN platform so that improvement can compound.

A software path to 6G

AI-RAN should not be treated as something that begins with 6G. The industry needs a software-led path where intelligence can be introduced continuously, governed safely, and improved without waiting for generational resets. In that sense, 6G is not a reset. It is a software upgrade within an architectural transition already underway.

That transition is now becoming concrete. Nokia launched its AI-RAN initiative in October 2025, demonstrated it with T-Mobile, moves into commercial trials in 2026, and targets commercial release in 2027. This is how AI-RAN moves from concept to platform, and from platform to deployment.

Voice made mobile essential. Data made it universal. AI will make it adaptive. And that next chapter begins now.