MWC 2026: Why the RAN Is Now an AI Compute Platform

The biggest signal from MWC Barcelona 2026 didn't come from a keynote. It came from three operators - NTT DOCOMO, SK Telecom, and a SoftBank-led consortium. Each releasing AI-native RAN demonstrations and white papers in the same 5 days window, before the show floor opened.

Read individually, they're interesting announcements. Read together, they're a coordinated industry position: the RAN is no longer purely a radio access system. It is a distributed AI compute platform, and the architecture decisions being made now will define network economics for the next decade.

The phrase "AI-native RAN" has circulated for two years. At MWC 2026, it finally has production-oriented specifics attached to it.

AI-native RAN means the radio access network is designed from the ground up to run AI inference workloads - predicting interference, optimising antenna configurations, managing handovers, detecting faults - not as an add-on layer, but as core network behaviour. The intelligence is embedded in the RAN architecture itself, not bolted on via an external platform.

The critical debate has been where that AI processing runs and on what hardware. The MWC 2026 pre-week answered that question - differently, across three distinct operator strategies.

DOCOMO's demonstration, confirmed February 24, 2026, is strategically significant precisely because of what it doesn't use: dedicated GPU infrastructure.

DOCOMO ran AI applications directly on general-purpose CPU resources inside a commercial vRAN network - HPE servers with Qualcomm accelerators, NEC vRAN software, managed on AWS infrastructure. AI processing runs in parallel with live network processing on COTS CPUs. No premium silicon required.

This challenges a widely-held assumption: that AI-in-RAN means AI-capex-in-RAN.

DOCOMO's architecture says otherwise. If AI inference can run on the same general-purpose compute already handling virtualised RAN workloads, the cost barrier to AI-native deployment drops substantially.

For operators watching AI infrastructure costs compound across thousands of sites, this is the most operationally important proof point at MWC 2026.

SK Telecom went further architecturally. Its ATHENA white paper released this week ahead of MWC calls for complete hardware/software decoupling and reintroduces the near-real-time RIC into a 6G-native architecture.

SKT's framing is deliberate: the RAN should simultaneously serve as communication infrastructure and AI inference infrastructure. xPU-based COTS servers replace proprietary base stations. Open interfaces eliminate vendor lock-in. The near-RT RIC handles real-time AI-driven optimisation decisions.

At MWC this week, SKT will demonstrate AI agents for network management, on-device AI-based antenna optimisation, and integrated communication-sensing a full-stack preview of AI-native operations in the field, not the lab. For context on how AI deployment has evolved across operators globally in February 2026, HYPHN's deep dive from February 24 covers the full picture including Telefónica's Level 4 automation progress and MTN's autonomous antenna operations.

SKT's ATHENA position is also a negotiating stance: complete HW/SW separation and open interfaces mean no single vendor controls the AI layer.

As AI becomes the differentiating capability in the RAN, that independence becomes commercially and strategically critical.

The third strand comes from a consortium: Northeastern University, SoftBank, Keysight, and zTouch Networks. Their AgentRAN demonstration at MWC 2026 (Keysight booth, Hall 2, March 2–5) introduces a hierarchy of AI agents that translates high-level operator intents directly into real-time 5G/6G network configurations - powered by a Large Telecom Model (LTM).

Announced February 26, 2026, AgentRAN represents the operational model that sits above the architecture debates: engineers define outcomes, AI agents execute configurations autonomously. This is not automation in the traditional scripted sense. It is intent-driven, context-aware orchestration the difference between programming a network and directing one.

Separately, NEC demonstrated Agentic AI-driven autonomous network operations in a verified multi-vendor environment, and significantly, the demo runs at the AWS booth, not NEC's own stand.

Hyperscaler-anchored telco AI is the infrastructure story of MWC 2026. The cloud is no longer adjacent to the RAN. It is inside it.

The vendor hardware layer has been quietly pre-positioned. Ericsson launched ten AI-ready radios with built-in neural network accelerators on February 16, 2026, ahead of MWC, timed precisely for this architecture conversation. These radios are designed to run the AI inference workloads that DOCOMO, SKT, and SoftBank are now defining at the software and orchestration layer.

The O-RAN Alliance Summit at MWC 2026 is hosting operator panels specifically addressing how AI is reshaping open RAN networks and how 6G should embed intelligence from Day 1 reinforcing that this is a standards-level movement, not individual operator experimentation.

Three operators. Three distinct architectures. One convergent thesis.

OperatorArchitecture PositionHardware StanceAI Layer NTT DOCOMOCPU-first, cost-flexible vRANCOTS CPUs, no GPUAI parallel to live RAN processing SK Telecom Full HW/SW separation + near-RT RICxPU-based COTS, vendor-agnosticAI agents for antenna, management, sensing SoftBank consortium Intent-driven agentic orchestration Hyperscaler-anchored (AWS)LTM-powered AgentRAN, autonomous config

They're not copying each other. They're each solving a different constraint - capex, vendor independence, operational agility, and arriving at the same destination: a RAN that runs AI natively, continuously, and autonomously.

When three tier-1 operators from three different regions announce coordinated architectural positions in the same week before MWC opens, they are not sharing news. They are setting industry direction.

The MWC 2026 show floor will be full of vendor AI demos polished, impressive, and optimised for best-case lab conditions. The real conversation is happening one architectural level below the demos.

The compute question is answered. NTT DOCOMO proved AI-in-RAN doesn't require GPU capex. The cost objection is now harder to sustain.

The independence question is being answered. SK Telecom's full HW/SW separation and open interfaces create a template for operators who want AI capability without vendor dependency.

The operations question is being asked. AgentRAN and NEC Corporation's agentic demos are the earliest production-oriented signals that intent-driven orchestration is moving from research to deployment.

Operators who engage MWC 2026 only at the device or services layer will miss the most consequential shift happening in the week of March 2–5. The RAN architecture decisions being validated at booths, summits, and operator panels this week will shape network investment and vendor relationships for the next five to seven years.

For ongoing analysis of AI deployment realities across global operators — not vendor roadmaps - subscribe to HYPHN's weekly intelligence brief. 1,200+ telecom leaders already do.

Follow HYPHN for AI in telecom intelligence. Subscribe free newsletter: getHYPHN.com