AI's Hidden Carbon Bill: Arabic Supermodel Noor's Full Footprint Revealed

Most of us picture AI's environmental cost as a one-time burst of electricity — a massive computer farm training a model for weeks. A new study says that picture is incomplete. Researchers tracked the entire carbon bill of Noor, an Arabic-language AI with up to 13 billion parameters (settings the model learns), designed to follow plain instructions and handle many tasks at once.

They counted everything, not just training: gathering and storing the text data, research and development budgets, the pretraining run itself, future electricity for serving users, and outside costs like travel and coordination for an international team.

The headline result is uncomfortable for the industry. The energy burned answering questions — called inference, or simply running the AI — plus those outside factors can add up to more than the training itself. In other words, a model's pollution doesn't stop when it's finished. It grows every time someone uses it, and popular AI tools are used millions of times a day.

The authors suggest ways to shrink the footprint, such as smarter scheduling and more efficient serving. For everyday readers, the takeaway is simple: as AI spreads into search, email, and apps, the environmental cost is paid continuously, not once. Choosing when you really need a giant AI — and when a smaller tool will do — is becoming a real climate decision, not just a technical one.