Large language models (LLMs) are increasingly power-hungry, and their rapid adoption has led to a surge in AI datacenter energy consumption, raising environmental concerns. However, most existing work on reducing LLM power use focuses on model-level or training-time optimizations, which are inaccessible to practitioners using closed-source “frontier” models such as GPT-5 or Gemini. To address this limitation, we propose two lightweight, input-level chat history reduction strategy—TF-IDF filtering and semantic search via sentence embeddings—that prune conversational context before inference without changing the model itself. We evaluate these strategies using a novel, production-scale, inference test framework built with vLLM and CodeCarbon for precise energy measurement, simulating real-world datacenter inference workloads. Experi- mental evaluation with Mistral-Small-24B-Instruct demonstrates that our methods achieve up to 66% energy savings per request for chat conversations with more than 19 messages, with only a modest (4–8%) drop in accuracy, and up to 93% savings for conversations longer than 200 messages. These results show that content-aware context pruning is a practical and highly effective approach to reducing inference energy consumption, enabling more sustainable and energy-efficient LLM deployment.