Three telecom giants dropped AI RAN announcements in the same week before MWC 2026 kicked off. NTT DOCOMO, SK Telecom, and a SoftBank-led consortium all revealed their AI-native RAN solutions within days of each other. This wasn't coincidence. It signals a fundamental shift: RAN networks are evolving from radio communication systems into distributed artificial intelligence platforms.
đ Read more: Nokia AI RAN: GPU-Powered Networks Ready for Q4 2026 Trials
đ§ From Reactive to Predictive: The AI RAN Evolution
By 2026, "AI-native RAN" has gained concrete meaning. Instead of external systems monitoring and analyzing networks, we're seeing artificial intelligence baked into the core RAN architecture itself. The shift is dramatic. Traditional RAN networks reacted to problems after they appeared. Now they predict interference, autonomously optimize antennas, and manage handovers before users even notice issues.What AI-Native RAN Actually Means:
Architecture designed from the ground up to run AI inference workloads â not as an add-on layer, but as fundamental network behavior.
Google Cloud at MWC 2026 unveiled its Autonomous Network Operations framework, moving carriers from "siloed automation" to self-healing networks. Companies like Deutsche Telekom and Vodafone are already testing systems that reduce operational complexity and transform connectivity from basic service into value creation engines.Architecture designed from the ground up to run AI inference workloads â not as an add-on layer, but as fundamental network behavior.
The Computing Power Dilemma
The critical question was where exactly AI processing would run. Do you need dedicated GPU clusters or can it execute on general-purpose CPUs? NTT DOCOMO delivered a surprising answer.⥠Three Different Approaches, One Common Goal
NTT DOCOMO: AI Without GPU Overhead
The Japanese carrier upended expectations. On February 24, 2026, they confirmed running AI applications directly on general-purpose CPU resources inside a commercial vRAN network â HPE servers with Qualcomm accelerators, NEC vRAN software, AWS management. This changes everything. AI processing runs parallel to normal network processing on the same COTS CPUs. No premium silicon required. For carriers watching AI infrastructure costs multiply across thousands of sites, this was MWC 2026's most significant proof point. If AI inference can run on the same general-purpose hardware already handling virtualized RAN workloads, the entry cost drops dramatically.SK Telecom: Complete Hardware/Software Decoupling
South Korean SK Telecom went further architecturally with their ATHENA white paper. They're demanding complete hardware/software decoupling and bringing the near-real-time RIC into 6G-native architecture. SKT's position is deliberate: RAN must function simultaneously as communication infrastructure and AI inference infrastructure. xPU-based COTS servers replace proprietary base stations. Open interfaces eliminate vendor lock-in.Level 4-5 Autonomy Target
Zero-touch Network Operations
SoftBank Consortium: Intent-Driven AI Orchestration
The third thread comes from a consortium: Northeastern University, SoftBank, Keysight, and zTouch Networks. Their AgentRAN introduces hierarchical AI agents that translate high-level operator intents directly into real-time 5G/6G network configurations â powered by Large Telecom Models (LTM). This isn't automation in the traditional scripted sense. It's intent-driven, context-aware orchestration â the difference between programming a network and directing it.đ The Underground Revolution: Hyperscaler Integration
A parallel theme unfolding at MWC 2026 is deep hyperscaler integration into telco infrastructure. Cloud isn't beside the RAN anymore â it's inside it. NEC demonstrated Agentic AI-driven autonomous network operations in a verified multi-vendor environment. Significantly, the demo ran at the AWS booth, not their own stand. Google Cloud presented a unified graph data layer that dissolves silos between operational and analytical data through Spanner Graph for digital twins.Network Digital Twin Evolution:
From static map to dynamic, temporal graph representing live physical and logical network state, with ability to query historical states for instant root-cause analysis.
Ericsson prepared the hardware layer: on February 16, 2026, they launched ten AI-ready radios with built-in neural network accelerators, timing aligned with the conversation operators are opening.From static map to dynamic, temporal graph representing live physical and logical network state, with ability to query historical states for instant root-cause analysis.
Graph Neural Networks in Practice
Operators can now train Graph Neural Networks (GNNs) on their network digital twin data in Vertex AI. Using the trained GNN models with Spanner's ML.PREDICT capability and real-time data, they move from monitoring to predicting â mathematically tracking how a failure might propagate and solving it before it affects subscribers.đ Read more: Ericsson AI RAN Delivers 10% Spectrum Efficiency Boost
đź Agentic AI: The Next Phase of Autonomy
MWC 2026 revealed the future isn't about "autonomous telco" but about delivering fundamentally new products and services aligned with business outcomes.Data Steward Agent
Automates data governance for accurate digital twins
Autonomous Network Agents
Manage voice core and OSS networks with active execution â like independent traffic rerouting
Nokia "Network as Code": Programming with Natural Language
Nokia is converting complex technical code into AI agents that understand everyday language. This allows telcos to simply ask the network to execute complex tasks â like prioritizing network resources for critical services such as emergency response or remote healthcare â without manual engineering."The shift to agentic AI requires changing how we measure success. As agentic AI coordinates actions across workflows, KPIs must move beyond productivity metrics toward predictive accuracy, resolution effectiveness, and revenue impact."
HFS Research, MWC 2026
đ Sovereignty vs. Efficiency: The Geopolitical Dilemma
Alongside technological evolution, MWC 2026 highlighted digital sovereignty as a recurring theme. Geopolitical tensions and regulatory pressure are pushing operators to position themselves as credible providers of sovereign digital infrastructure. The problem? There's no universal definition of digital sovereignty. For some, it's data residency and jurisdictional control. For others, it extends to technology supply chains, infrastructure ownership, and AI ecosystem control. Many "sovereign" claims still rely on global cloud and hardware stacks. This ambiguity gives operators, hyperscalers, and vendors opportunity to shape sovereignty around their own strategic positioning.The Agentic AI Connection
That's why agentic AI sits at the center of the sovereignty debate. Scaling agents transforms "control" from narrative to operating model choice. Those who decide where AI will exist, how it connects across domains, and what business outcomes it produces will be the real winners. Instead of enterprise-wide rollout, successful operators follow a staged adoption playbook. They start with domains where feedback loops are strong and risk is contained â customer experience, financial operations, IT workflows. As governance and trust mature, orchestration expands into BSS and OSS environments.đŻ Frequently Asked Questions
What's the difference between AI-native RAN and traditional systems?
Instead of reacting to problems after they appear, AI-native RAN predicts and prevents them. Artificial intelligence is embedded in the architecture, not an external add-on.
Do you need expensive GPUs for AI RAN deployment?
Not necessarily. NTT DOCOMO proved AI inference can run on general-purpose CPUs parallel to normal RAN tasks, dramatically reducing entry costs.
How is security ensured in autonomous networks?
Through staged adoption with bounded risk, governance frameworks, and continuous monitoring. Autonomy develops gradually from low-risk domains toward critical functions.
MWC 2026 didn't just show demos. It revealed a coordinated industry movement toward a future where networks think, predict, and act autonomously. The architectural decisions being made now will shape network economics for the next decade. And it appears the major players have already made their choices.