Closed model and API providers do not disclose the data needed to compute a real footprint

Why this is its own facet

The parent measurement sub-issue can be closed for compute you run or bill for. It cannot be fully closed for compute you consume through a closed API, because the provider withholds the inputs. This facet isolates that structural opacity, which caps the accuracy of estimation approaches like model-based LCA no matter how good the method is.

Observable evidence

Providers generally treat model size, hardware, and data-centre efficiency as competitive secrets. Independent benchmarking shows the effect directly: on one public energy leaderboard, closed providers have largely declined to submit their production models even through a secure sandbox, while by one count 15 of the 20 most-used models are closed (Hugging Face, 2025). The ranges that make cross-model comparison meaningful therefore cannot be completed.

Separating fact from interpretation

The observed fact is non-disclosure and the resulting absence of closed models from comparable datasets. The interpretation—that this is a deliberate competitive choice rather than a technical barrier—is supported by the existence of secure evaluation sandboxes that providers decline to use.

Impact if unaddressed

Buyers cannot choose on efficiency, regulators cannot verify claims, and the fastest-growing slice of AI usage—closed frontier models in production—stays the least measurable. Estimation tools remain necessary but permanently approximate.