Qualcomm has announced a new Hexagon NPU architecture featuring an Element Accelerator and 50% larger shared memory to support on-device agentic AI. The platform is designed to improve performance for Mixture-of-Experts models and accelerate INT4 prefill tasks for mobile devices.
This hardware evolution signals a transition from monolithic AI models to specialized, routed systems that operate locally on mobile chipsets. By integrating the Element Accelerator and expanding on-chip memory, Qualcomm reduces the latency and bandwidth bottlenecks that previously limited complex agentic reasoning in mobile environments. For the streaming and media ecosystem, this enables more sophisticated on-device metadata processing and personalized content discovery without relying on cloud compute. Watch for the first commercial devices featuring this architecture to debut at the Snapdragon Summit 2026 to see real-world token generation benchmarks.
Nokia has built its agentic AI strategy around a tightly integrated hardware and software stack anchored by Nvidia GPUs. In June 2026, Nokia announced an agentic AI framework built into its Network Services Platform for IP network operations, with a root-cause troubleshooting agent as the first commercial deployment expected by end of 2026. The framework operates within operator-defined safety boundaries, allowing AI agents to act on real-time network data without human intervention for routine tasks. This positions Nokia as a direct competitor to Qualcomm's on-device agentic approach, but at the network infrastructure layer rather than the endpoint.
The competitive landscape for agentic AI in telecoms intensified sharply in June 2026. Ericsson launched its AI in RAN commercial software subscription on June 11, claiming up to 20% higher downlink throughput across more than 15 live deployments, while Verizon disclosed that its 60,000-site vRAN network is now applying agentic AI network operations to configuration changes and service assurance. Verizon publicly called for industry-wide interoperability standards for agentic systems, highlighting a gap that mirrors early Open RAN interoperability challenges. The TM Forum's Autonomous Networks L4/5 roadmap and 3GPP 6G standardization process will need to incorporate agentic AI interoperability as a core requirement to prevent vendor lock-in.
Nokia's broader architecture play extends beyond individual agents into a full-stack autonomous network vision. Nokia announced partnerships with AWS and Databricks at DTW Ignite to build data, cloud, and control layers for its Autonomous Network Fabric, claiming operators are already achieving automation rates above 90%, service delivery times under four hours, and up to 85% reduction in slice rollout time. Meanwhile, the fundamental hardware divergence between Nokia and Ericsson on AI-RAN is widening. Light Reading reported that Nokia is designing its entire Layer 1 RAN to run on Nvidia CUDA platforms and GPUs, while Ericsson uses GPUs only for forward error correction and runs all other L1 software on CPUs. This architectural split means the two vendors are building incompatible acceleration paths, a dynamic that Arm CSS for Mobile 2 approach sidesteps by targeting endpoint inference rather than network-side processing.
Qualcomm has unveiled a new Hexagon NPU architecture featuring a specialized Element Accelerator and 50% more shared memory. This update optimizes transformer and Mixture-of-Experts models for mobile devices. By reducing data transfer bottlenecks, it enables sophisticated on-device agentic AI, allowing for faster, more responsive local processing without relying on cloud compute.
The Element Accelerator is purpose-built to optimize transformer operations and multimodal reasoning tasks, improving responsiveness for agentic AI on mobile devices.
The architecture increases shared memory capacity by 50%, which minimizes the need for data transfers to external DDR and improves overall system efficiency.
The NPU provides up to 50% faster prefill performance for INT4-based models and supports Mixture-of-Experts models, enabling 30B-parameter class experiences using only 3B active parameters.
The first commercial devices featuring this new Hexagon NPU architecture are expected to debut at the Snapdragon Summit 2026.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source