Agentic artificial intelligence is moving from lab demos into operator trials and trade‑show headlines, with vendors and carriers pointing to double‑digit throughput gains and a potential shift in compute strategy for radio access networks. New demonstrations at Mobile World Congress 2026 and field tests from multiple operators show Ericsson’s reinforcement‑learning based “agentic” Cloud RAN can raise downlink throughput by more than 20 percent in challenged cells while Intel says its latest CPU extensions let that inference run on standard servers — potentially delaying the need for widespread GPU rollouts.
At MWC 2026 Ericsson highlighted an end‑to‑end call with AT&T on March 3 that ran through an AI‑native link adaptation model. Unlike traditional Cloud RAN implementations where link adaptation follows static rule tables, Ericsson’s approach uses reinforcement learning agents that observe channel conditions, plan modulation and coding actions, and update policies every transmission time interval. Company engineers say the model reduced block error rates and enabled higher modulation orders, producing higher average user throughput without adding spectrum or radios.
Operator trials provide early performance signals. Optus reported more than 20 percent average downlink gains in three days of suburban Sydney testing during medium and poor RF periods. AT&T’s lab measurements recorded throughput uplifts approaching 20 percent at the cell edge, and Bell Canada’s pilots measured roughly a 10 percent improvement in spectral efficiency in early tests. Vendors caution that sustained, network‑wide gains will require longer and broader live deployments across seasons and varied load conditions.
A second hardware story runs in parallel. Intel’s Granite Rapids‑D Xeon platform, the company says, incorporates AVX‑512 and AMX vector engines that let many inference layers execute on conventional CPU cores. Intel demonstrated co‑hosting distributed unit (DU), central unit (CU), user plane function (UPF) and AI inference on a single server, claiming sub‑millisecond model latency, maintained carrier‑grade determinism under mixed traffic, and up to 20 percent lower rack power compared with configurations using discrete GPUs. Company executives stressed their software remains hardware agnostic, positioning CPUs as a capital‑efficient starting point for early AI radio tasks.
Related Recipes
Solidion unveils Gen-ECB battery platform with graphene thermal management for space, LEO data centers and lunar missions
800-Year-Old Wax-Coated Medieval Notebook Unearthed in Paderborn Latrine Reveals Late Medieval Literacy and Merchant Life
Marvell surges after Nvidia CEO calls it the next trillion-dollar company, fueling bets on AI silicon and a $10 billion custom‑chip revenue target by 2029
AI-driven S&P 500 rally signals narrow breadth as AI stocks lead gains
That positioning has sparked a compute strategy debate. Nokia and NVIDIA are pushing GPU‑heavy designs that anticipate larger, more complex models and multi‑agent coordination, while Ericsson’s parties emphasize flexibility across Intel, AMD or ARM server ecosystems. Operators weighing a uniform server fleet and simpler procurement against potential future accelerator needs face trade‑offs: immediate savings and integration simplicity versus avoiding possible silicon lock‑in and ensuring headroom for higher‑order AI workloads.
Operational and governance risks also loom. Agentic systems replace human‑readable rules with learned policies, raising regulator and operator demands for explainability, rollback controls and robust telemetry. Researchers and vendors propose hierarchical agent architectures with safety layers and continuous assurance hooks; Ericsson has indicated telemetry features to support oversight. Analysts say disciplined monitoring, staged rollouts and clear model‑ownership agreements are prerequisites before nationwide adoption.
Trending Now
Andreeva Faces Chwalińska in Roland Garros Final for Maiden Grand Slam Title
Uncertain AMOC weakening highlights need for intensified ocean monitoring and emissions cuts after 2020 acidification risk finding
D.C. Circuit panel questions Trump administration argument that projects underway are untouchable in White House East Wing ballroom dispute
Paige Bueckers’ 2025 WNBA debut jersey sells for $64,720, a record for women’s game-worn memorabilia
Industry roadmaps envision agentic control expanding beyond link adaptation to spectrum allocation, energy management and slice orchestration by 2030, effectively forming autonomous cell clusters. Intel plans further vector width increases and multi‑agent features to support that evolution. For now, the combination of Ericsson’s reinforcement‑learning link adaptation and Intel’s CPU acceleration presents a capital‑efficient path to early AI‑native 6G capabilities — but operators will demand broader field evidence, clear economics and tight governance before committing to large‑scale deployments.
Continue Reading
