The demand comes directly from the United Nations headquarters, with an amplification that leaves no room for misunderstanding. UN Secretary-General Antonio Guterres has put artificial intelligence companies in front of a clear choice: stop treating the environmental cost as someone else’s problem, and start accounting for how much carbon, water, and land their networks consume.

For those developing and deploying Large Language Models, the impact has never been a negligible detail, but it has often remained confined to vague estimates. Training a large model involves electricity consumption that can rival the annual needs of dozens of households, not to mention the liters of water used to cool servers and the land occupied by data centers. The United Nations’ call aims to break the veil of opacity: organizations should make these numbers public, enabling independent assessments.

AI’s invisible footprint

Attention has so far focused on computing power, model quality, or cost per token during inference. But behind every request to a chatbot or an image generation system lies a physical infrastructure that demands resources. Specialized chips – from GPUs to custom processors – run at full capacity in facilities that, unless powered by renewable sources, worsen the climate footprint. Moreover, the production of hardware itself requires rare earth mining and energy-intensive industrial processes. Without clear accounting, the sector risks becoming a giant with feet of clay, where innovation outpaces responsibility.

On-premise: control and sustainability

For organizations evaluating on-premise deployment, the environmental question intertwines with data sovereignty and Total Cost of Ownership. Those who manage their own servers running LLMs or training pipelines can choose the energy source, optimize workloads, and reduce waste. Of course, a self-hosted infrastructure is not automatically green: it requires investments in efficient hardware and clean power purchase agreements. But it offers a degree of transparency that multi-tenant cloud services do not always guarantee. This is exactly the kind of disclosure the UN is urging: knowing how much CO₂ is emitted for every activity, from model quantization to inference serving.

Beyond marketing: a governance choice

International pressure shifts the debate from corporate communications to governance. It is no longer enough to declare generic commitments to carbon neutrality: verifiable metrics, extended to the entire supply chain, are needed. Development banks and institutional investors may soon demand them as a condition for financing. In this scenario, organizations that already measure the energy consumption of their machines – for instance through energy monitoring frameworks integrated with container orchestration – start with a competitive advantage. Sustainability, from a negligible variable, becomes a trust asset.