The landscape of intelligent computing is rapidly evolving, driven by the deployment of large-scale AI workloads under strict performance, portability, and sustainability constraints. Emerging applications, from real-time large language models (LLM) inference and federated learning to multimodal media processing, demand low-latency execution, dynamic adaptability, and carbon-aware orchestration across heterogeneous compute nodes. Meeting these demands requires rethinking how we design “systems for AI”, modular, edge-native, runtime-adaptive infrastructures, and how we use “AI to optimize systems” through learned scheduling, drift detection, and failure anticipation. ScaleSys 2025 targets technically sound contributions that advance scalable, intelligent, and sustainable computing across the edge-cloud continuum. We seek works on serverless execution models tailored for distributed AI workflows, edge-native runtime systems supporting, e.g., WebAssembly or lightweight VMs, adaptive orchestration frameworks that leverage learning-based or multi-objective scheduling schemes, and carbon-aware placement strategies using real-time energy telemetry. We particularly encourage designs addressing hybrid LLM inference across compute layers, real-time reconfiguration, and AI-enhanced decision logic. Contributions should demonstrate evaluation on testbeds, simulators, open datasets, or reproducible benchmarks. The workshop will feature a full-day program of peer-reviewed papers, a keynote talk, and a concluding panel, aiming to serve as a convergence point for foundational research in systems-for-AI and AI-for-systems.