SYNAPSE-AI aims to develop new strategies for the sustainable and energy-efficient management of computing infrastructures dedicated to Artificial Intelligence, with particular reference to HPC systems and the serving of large-scale linguistic models (LLMs). The project combines load and power forecasting techniques, causal inference, dynamic KV-cache management, and GPU resource control, with the goal of reducing energy consumption while maintaining high levels of performance and quality of service. Another innovative element is the use of time synchronization between workload data, GPU telemetry, and electrical measurements, thus improving the reliability of energy analyses and orchestration decisions.