Transformer-based models, such as FlauBERT, deliver state-of-the-art performance in natural language processing (NLP) but are computationally expensive and energy-intensive. This study explores Kronecker-Inspired Embedding Compression, a structured compression method that reduces the embedding layer’s dimensionality while preserving model performance. Unlike traditional low-rank factorization or quantization, Kronecker-based embedding compression decomposes large embedding matrices into compact, structured representations, optimizing both storage and computation. We propose a mathematical formulation for Kronecker-based embedding compression and analyze its impact on computational complexity and memory efficiency. Through empirical evaluations on a French sentiment analysis dataset, we assess the trade-offs between compression, accuracy, and energy consumption. Experimental results show that Kronecker-based embedding compression significantly reduces model size and achieves up to a 36% decrease in energy consumption during the inference phase, while maintaining accuracy above 97.5%. Energy profiling, conducted using PowerAPI and the RAPL interface, validates the method’s efficiency across different compression levels. These findings highlight Kronecker-based embedding compression as a promising solution for deploying energy-efficient NLP models, particularly in resource-constrained environments. The proposed approach contributes to sustainable AI by reducing inference costs without significant loss in accuracy. Future work will explore extending Kronecker compression beyond embeddings to self-attention mechanisms and feed-forward networks in Transformers.