AI-Native RAN: How Artificial Intelligence Will Run Future Networks

For decades, radio access networks have been built on classical signal processing — mathematically precise algorithms designed by engineers and hard-coded into silicon. AI was added as an afterthought: analytics dashboards, optimization suggestions, anomaly detection running in a cloud server far from the antenna. This analysis is prepared by the 7G Network editorial team, drawing on primary sources from MWC 2026 demonstrations, 3GPP study items, and vendor disclosures.

That model is ending. At MWC 2026, the telecom industry demonstrated — not promised, demonstrated — that AI is moving into the radio itself. Neural networks replacing classical receiver chains. Machine learning models running on base station hardware in real time. The shift from AI-assisted to AI-native is no longer theoretical, according to NVIDIA (2026).

What "AI-Native" Actually Means

The distinction between AI-assisted and AI-native is architectural, not cosmetic:

AI-assisted (current 5G): Classical algorithms handle the radio functions — channel estimation, beamforming, scheduling, interference management. AI runs alongside, analyzing performance data and suggesting optimizations. The AI layer can be removed and the network still functions. AI makes the network better; the network does not depend on AI.

AI-native (6G target): Neural networks are the radio functions. Channel estimation is performed by a trained model, not a least-squares algorithm. Beamforming weights are predicted by inference, not computed by matrix operations. The AI cannot be removed because there is no classical fallback for the functions it performs. AI does not improve the network; AI is the network.

The practical difference: an AI-native air interface can handle radio conditions that no classical algorithm was designed for. Interference patterns too complex to model. Channel dynamics too rapid to track with traditional estimation. User behavior too unpredictable to schedule optimally with rule-based systems.

AI-native RAN replaces classical radio algorithms (channel estimation, beamforming, scheduling) with neural networks that cannot be removed. Unlike AI-assisted 5G, where AI is optional, AI-native 6G/7G networks depend on machine learning as the core protocol.

MWC 2026: The Evidence Arrives

MWC 2026 in Barcelona was the inflection point. For the first time, AI-native RAN moved from research papers to commercial product announcements and live demonstrations.

NVIDIA's Global Coalition

NVIDIA secured commitments from over a dozen global operators and technology companies to build 6G on open, secure, AI-native platforms. The coalition includes BT Group, Deutsche Telekom, Ericsson, Nokia, SK Telecom, SoftBank, T-Mobile, Cisco, and Booz Allen Hamilton.

The AI-RAN Alliance — the industry body driving AI-RAN standards — now has over 130 participating companies. NVIDIA's role is providing the GPU compute platform (NVIDIA Aerial) that runs neural network inference at the speeds required by radio processing — microseconds, not milliseconds.

NVIDIA also unveiled the All-American AI-RAN Stack with US telecom leaders — a domestic supply chain initiative for AI-native 6G infrastructure, reflecting growing emphasis on supply chain security in critical communications infrastructure.

Ericsson's Neural Radios

While Nokia bet on NVIDIA GPU acceleration, Ericsson took a different path: purpose-built silicon with embedded neural network accelerators. At MWC 2026, Ericsson unveiled ten new AI-ready radios built on its own custom chips, with AI capabilities integrated directly into the Massive MIMO hardware.

The portfolio includes:

• AI-managed beamforming: Neural networks predict optimal beam patterns faster than classical beam search
• AI-powered outdoor positioning: Sub-meter accuracy using base station signals alone, no GPS required
• Instant coverage prediction: Real-time radio propagation modeling using trained neural networks instead of ray-tracing simulations
• Latency-prioritized scheduler: Delivering up to 7x faster response times compared to rule-based scheduling

According to Ericsson (MWC 2026), this approach trades the flexibility of GPU-based inference for lower power consumption and tighter integration — the AI is in the radio, not running on a co-located server.

DeepSig: The AI-Native Air Interface

DeepSig demonstrated what may be the most forward-looking technology at MWC 2026: OmniPHY-5G — a pre-6G fully learned waveform running on top of a 3GPP Release 17 GPU-accelerated stack. This is not AI optimizing a classical waveform; it is a neural network generating the waveform itself.

Running on the compact NVIDIA DGX Spark, DeepSig showed a carrier-grade base station supporting both commercial 5G and emerging 6G devices — demonstrating that AI-native and classical signals can coexist in the same radio, enabling gradual transition rather than forklift upgrades.

Neural Receivers

A rapidly advancing area is the neural receiver — replacing parts of the classical receiver chain (demodulation, equalization, decoding) with neural network inference. Rohde & Schwarz and NVIDIA have done foundational work on digital twins for training and testing neural receivers, solving the chicken-and-egg problem of training a neural receiver for a signal that does not yet exist in the real world.

Nokia demonstrated AI-based digital post-distortion (DPoD) — using a neural network at the base station to reverse power amplifier distortion after reception, improving signal quality without modifying the transmit chain.

At MWC 2026, NVIDIA formed a coalition of 130+ companies for AI-native 6G, Ericsson unveiled 10 neural radios with 7x faster scheduling, and DeepSig demonstrated OmniPHY-5G — a fully AI-generated waveform running on NVIDIA DGX Spark hardware.

Base Stations as AI Data Centers

One of the most significant architectural shifts is the reconceptualization of the base station itself. In the AI-native model, a base station is not just a radio — it is a distributed AI compute node.

GPU-accelerated base stations can process AI workloads during periods of low radio traffic. Instead of sitting idle at 3 AM when few users are connected, the GPU runs inference tasks for enterprise customers — image recognition, natural language processing, anomaly detection — with the results served at the network edge, milliseconds from the end user.

This creates new revenue models for operators:

• Token-as-a-Service: Selling AI inference tokens from base station GPUs
• GPU-as-a-Service: Renting base station compute to enterprises during off-peak hours
• Edge AI hosting: Running customer AI models at the network edge for low-latency applications

The economic impact is significant: base station GPU utilization could increase from the current ~30% (radio-only) to 80–90% (radio + AI workloads), fundamentally changing the ROI equation for dense network deployments. This dual-use model aligns with 7G network architecture concepts where intelligence is distributed across every node.

GPU-accelerated base stations can run AI inference workloads during off-peak hours, increasing utilization from ~30% to 80–90%. New revenue models include Token-as-a-Service, GPU-as-a-Service, and edge AI hosting.

The Technical Challenges

Latency Budget

Radio processing operates on microsecond timescales. A 5G NR slot is 0.5 ms at its shortest. Channel estimation, beamforming, and scheduling must all complete within this window. Running neural network inference within this budget requires specialized hardware — general-purpose CPUs are too slow, and even many GPU architectures are not designed for the deterministic, bounded-latency execution that radio demands.

Training Data

Classical algorithms work from first principles — they do not need training data. Neural network replacements require massive datasets of radio conditions, channel measurements, and performance outcomes. Generating this data is expensive, and transferring a model trained in one environment to a different cell site (the "domain gap" problem) remains unsolved at scale.

Digital twins — high-fidelity simulations of radio environments — are the current best solution, but they introduce their own accuracy questions. A neural receiver trained on a simulated channel may underperform on the real channel if the simulation is not accurate enough.

Explainability and Certification

When a classical algorithm fails, engineers can trace the failure to a specific mathematical step. When a neural network fails, the failure mode is opaque. For safety-critical applications (V2X, emergency services), regulators may require explainable behavior that neural networks cannot currently provide.

The likely compromise: AI-native for performance-critical but non-safety-critical functions (throughput optimization, beam management), with classical fallback for safety-critical signaling (emergency calls, V2X collision avoidance).

Standardization Gap

According to 3GPP (2026), Release 20 study items are exploring AI for the air interface, but the consensus is converging on AI-assisted with classical fallback for the first 6G standard. Full AI-native — where no classical algorithm exists for a given function — is a Release 22+ aspiration, likely arriving in the 6G Advanced timeframe (2033–2035). The 6G standardization timeline provides further detail on these milestones.

Key technical barriers to AI-native RAN include microsecond-level inference latency, training data scarcity (the domain gap problem), neural network explainability for safety-critical functions, and a standardization gap — full AI-native is targeted for 3GPP Release 22+ (2033–2035).

From 6G to 7G: AI as the Protocol

If 6G makes AI native to specific radio functions, 7G is expected to make AI native to the entire protocol stack. Rather than training separate models for channel estimation, beamforming, and scheduling, a unified foundation model would manage the complete radio chain — sensing the environment, deciding what to transmit, choosing how to modulate it, and adapting in real time.

This is sometimes called "cognition-native networking" — the network does not just process data according to rules, it understands context, predicts outcomes, and makes decisions. The 7G network knows that a user walking toward a building entrance will need a handoff to an indoor small cell in three seconds, and begins the handoff before the user reaches the door.

The gap between today's AI-RAN demonstrations and this vision is enormous. But MWC 2026 proved that the first step — replacing individual classical functions with neural networks — works at commercial scale. The rest is engineering, funding, and time. For a broader perspective on how 6G and 7G compare, see our detailed analysis.

7G envisions "cognition-native networking" — a unified foundation model managing the entire protocol stack from channel estimation to scheduling. Unlike 6G's function-specific AI, 7G targets AI as the protocol itself, predicting user behavior and adapting in real time.

Sources

1. NVIDIA Telecommunications — AI-RAN coalition announcements and NVIDIA Aerial platform details, MWC 2026
2. Ericsson AI-RAN — Neural radio product portfolio and custom silicon AI accelerator specifications
3. DeepSig — OmniPHY-5G AI-native air interface demonstration and technical specifications
4. AI-RAN Alliance — Industry body for AI-native RAN standards, membership and technical roadmap
5. 3GPP — Release 20 AI/ML study items for the air interface, standardization timeline for AI-native functions
6. Rohde & Schwarz / NVIDIA — Digital twin training and testing methodology for neural receivers