The analysis explains why there is rarely a simple answer to the question of AI footprint. Impact depends on many variables at once, including model size, prompt type, data center location, cooling setup, infrastructure distance, and the balance between training and usage. At the same time, reliable reporting remains limited, as major model developers and hyperscalers are still selective in what they disclose.

To make the topic actionable, Cruz separates it into two layers: hardware and software. On the hardware side, the key factors are GPUs, cooling methods, and data center efficiency. NEMO Kennislink highlights differences in cooling approaches and the use of indicators such as PUE to estimate overhead beyond compute. On the software side, impact comes from model scale and from both high-cost phases: training and inference. Inference is growing quickly due to large-scale daily adoption.

The AI angle in the article focuses on task-level efficiency. Using the largest model for simple prompts is inefficient, while routing requests to smaller components can reduce compute substantially. Cruz and his team are developing an optimization layer that directs low-complexity requests to lighter model segments when possible. This approach reflects a wider industry direction, with similar strategies now appearing in publications from major AI and chip actors.

NEMO Kennislink’s core message is that technical innovation and governance need to move together. Without stronger transparency, meaningful comparison remains difficult, and without smarter model architectures, usage growth can outpace efficiency gains. In that sense, AI footprint is primarily a design and steering challenge, not just a capacity challenge.