5. Inputs and Evidence Sources

The methodology draws on the following categories of inputs.

5.1 Hardware Benchmarking Data

Energy measurements collected from physical AI inference workloads running on real GPU hardware. GPU-board power is captured via the NVIDIA Management Library (NVML), sampled at an interval of 10 milliseconds, with non-GPU server energy estimated from the inference-phase wall time and the power specification of the host server.

5.2 Server Specification Data

Published hyperscaler AI instance specifications, used as the reference configuration for estimating the non-GPU server contribution, covering processor generation, core count, memory capacity, and the resulting power draw at a given accelerator count. Processor utilisation follows the default assumption applied in the core Greenpixie compute methodology.

5.3 Model and Architecture Data

Publicly available and vendor-published information about AI model characteristics, including parameter counts, active parameter counts for mixture-of-experts (MoE) architectures, quantisation levels, and VRAM requirements. For open-weights models, this information is typically available directly from model documentation and hosting platforms.

For proprietary models where architecture details are not publicly disclosed, the methodology establishes a plausible range for each characteristic that cannot be observed. These ranges draw on provider tier naming conventions, the published distribution of frontier open-weights releases, independent datasets of notable model parameter counts, published throughput measurements, and the architectural characteristics of comparable open-weights models. The banding and sampling process is described in Sections 6.4 and 7.6.

5.4 Cloud Provider Billing and Pricing Data

Publicly available pricing structures and billing configurations from AWS, Azure, and GCP, used to identify token-type distinctions, processing tiers, and caching discounts, and used in the price-ratio validation described in Section 9.5.

5.5 Embodied Emissions Data

Hardware manufacturing and lifecycle emissions data adapted from the Boavizta methodology and associated formulae for server and GPU cradle-to-gate embodied emissions calculations.

5.6 Carbon Intensity Data

Regional and hourly electricity grid carbon intensity factors, as described in the core Greenpixie Methodology for Cloud Emission Measurement.

5.7 Water Consumption Data

Cloud provider Water Usage Effectiveness values and regional water intensity factors, as set out in the Greenpixie Methodology for Cloud Emission Measurement.

5.8 Published Research and Industry Sources

Publicly available research from institutions and organisations including the International Energy Agency, Google, Mistral AI, Epoch AI, Artificial Analysis, MIT Technology Review, and academic inference-energy benchmarking literature, used to inform contextual understanding, methodological design, and the external comparisons described in Section 9.4. A full bibliography is provided in Section 12.