1
Operators can switch on Ericsson’s AI models on existing AI-ready hardware
In sum – what we know:
- Software-only delivery – AI in RAN activates on existing Ericsson hardware, including previous-gen AI-ready radios, with no capex-heavy box swap required.
- AI at the radio’s timescale – “Telco-grade” models run on constrained edge hardware in basebands and radios, no cloud GPUs needed, optimizing scheduling, beamforming, and coding in near-real time.
- Trial-backed numbers – Live trials on T-Mobile’s 5G Advanced network showed roughly 10% better spectral efficiency and up to ~15% higher downlink throughput, with full deployment targeted for Q3 2026.
Ericsson has launched a new software suite it calls “AI in RAN,” a subscription package that embeds AI models directly into basebands and radios. Rather than treating AI as a layer that sits in the cloud and reaches down into the network, Ericsson wants the intelligence living inside the radio access network itself, operating at the timescales the radio actually runs on. The company frames this as a step toward “AI-native” RAN, with the first wave of six features focused on performance, efficiency, observability, and automation.
The most consequential detail is the delivery model. AI in RAN is software only, which means operators can activate it on existing Ericsson hardware, including previous-generation “AI-ready” radios, without a capex-heavy box swap. That matters in a mature 5G market where operators are wary of writing big checks for incremental gains. Ericsson is essentially offering an upgrade path for 5G and 5G Advanced networks that doesn’t require ripping anything out.
It’s not a distant roadmap item, either. Some features are available now, with the rest arriving later in the year, and the whole thing builds on live commercial trials, most notably with T-Mobile US, where Ericsson’s AI-native scheduler produced measurable spectrum and throughput gains. Whether those gains justify the subscription cost is the question operators will actually be asking, but the trial data gives Ericsson something concrete to point to.
Core capabilities
At the heart of the package are what Ericsson calls “telco-grade” AI models, embedded in both the RAN data plane and control plane and tuned for microsecond-level radio timescales. “Telco-grade” is a marketing term, but the underlying claim is reasonable enough. These models are meant to run with the reliability and determinism a live carrier network demands, not the best-effort behavior of a lab demo. The practical upshot is near-instant optimization of things like modulation, coding, beam direction, and resource scheduling.
The more interesting engineering choice is where these models run. Ericsson stresses that AI in RAN does not require custom cloud data-center GPUs. The models are designed for constrained edge hardware inside basebands and radios, including the AI-ready radios and RAN Compute processors Ericsson rolled out over the past couple of years. That’s a deliberate contrast with the GPU-heavy approach favored elsewhere in the AI world, and it’s what makes the software-only deployment story credible. The compute headroom is already in the field.
Ericsson also leans on continuous learning, with models retrained or fine-tuned on data the network generates itself rather than frozen at deployment. The idea is that a model can adapt to traffic patterns like rush hours and stadium events, to radio environments ranging from urban canyons to suburbs, and to new services such as AR/VR and industrial IoT.
Then there’s “agentic AI,” which Ericsson is keen to highlight. The pitch is multi-agent systems that observe the network, interpret telemetry, and take or recommend actions, handling troubleshooting, anomaly detection, and self-healing behaviors. This moves past a single ML model bolted onto a single function toward agents that coordinate across features and layers. It’s also the part of the announcement that raises the most questions about oversight, rollback, and explainability when an AI agent is allowed to act on a live network. Ericsson doesn’t say much about those guardrails, which is worth keeping in mind.
Expanded AI feature wave
The marquee feature is the AI-Native Scheduler for Link Adaptation, a neural-network scheduler that predicts rapidly changing radio conditions in real time and adapts modulation, coding, and resource allocation accordingly. This is the feature with the hard numbers behind it. In live trials with T-Mobile’s 5G Advanced network, Ericsson reports roughly 10% better spectral efficiency and up to about 15% higher downlink throughput versus traditional rule-based schedulers. Those are vendor figures from a friendly trial, so treat the upper bound with some caution, but even the conservative end of that range is material. In a mature 5G market, a 10% spectral efficiency gain means serving more traffic from the same spectrum holdings, which translates fairly directly into deferred spectrum purchases and fewer new sites.
AI-Powered Macro Positioning uses radio measurements to pinpoint user location more accurately within macro coverage. The payoff is better handover decisions and load balancing, more precise capacity planning, and the door it opens to location-aware services. It’s a quieter feature than the scheduler, but the location data it produces feeds into the rest of the suite.
AI-Managed Beamforming steers and shapes the beams from massive MIMO antennas, learning the patterns that work best in complex environments full of reflections and blockages. The goal is higher throughput, fewer dropped calls, and less interference and signaling overhead, particularly in dense urban areas where beamforming has the most to gain. This isn’t entirely new ground for Ericsson, which introduced AI-managed beamforming in its earlier AI-ready radio launch. AI in RAN folds it into a coherent package alongside everything else.
AI-Powered Multi-Layer Coordination handles traffic across spectrum layers, predicting where users will be and what they need, then shifting them to the best layer for the job. High-band and mid-band carry the heavy data; low-band handles coverage and mobility. The aim is better spectral efficiency and a smoother experience with less manual parameter tuning, which is the kind of work operators currently throw a lot of engineering hours at.
The last two features are less glamorous but arguably foundational. Performance Management Event Schema Files restructure raw network data so it’s natively machine-readable by AI agents rather than formatted for human engineers. That makes it easier for AI systems, including rApps and xApps, to ingest performance data, correlate events across RAN elements, and automate root-cause analysis. Augmented Observability complements it by generating richer, more granular data on user experience KPIs, interference, and timing, giving the models better input to work from. Neither feature does anything visible on its own. Both are the plumbing that makes the rest of agentic AI plausible.
Rollout timelines
The first AI in RAN features are available immediately, in the June 2026 window, with the remainder shipping later in 2026 as part of the same package. That’s a near-term commercial product, not a roadmap promise, and the timing reflects work that’s already been validated in the field rather than a fresh start.
Ericsson is selling all of this as a software subscription, which cuts both ways. For Ericsson, it means predictable recurring revenue and fits the company’s broader pivot toward software- and services-led income from RIC platforms, rApps, and managed services. For operators, it lowers the upfront cost and softens the business-case friction compared with a hardware refresh. The open question Ericsson hasn’t answered publicly is how the pricing actually works, per site, per sector, or per feature, and that detail will determine how quickly the efficiency gains pay back the subscription.
The proof points come from T-Mobile US, which ran extensive commercial trials of the AI-native scheduler on live 5G Advanced traffic. The trials started in early Q2 2025 and expanded across several U.S. markets including Los Angeles, New York, and Salt Lake City. T-Mobile is now targeting full commercial deployment in Q3 2026, which gives Ericsson a clear trial-to-rollout path with a flagship operator attached to it. The two also showed Cloud RAN software running on NVIDIA AI infrastructure, hinting at how the cloud and edge pieces are meant to fit together.
The longer-term sell to operators is efficiency. Ericsson argues that the same capacity targets can be met with fewer radios over time, and that AI can selectively power down radios and tune power levels based on traffic, cutting energy consumption. Both claims feed directly into operator narratives around opex reduction and sustainability. They’re plausible given the spectral efficiency numbers, though the “fewer radios” benefit accrues gradually rather than on day one.
