3. Definitions

AI. Artificial intelligence; in this context, large language models (LLMs) and embedding models.

Aleatoric Uncertainty. The inherent variation between comparable workloads, which does not reduce with better modelling but does average out when values are aggregated across many billing line items.

Batching. The process of grouping multiple, separate prompts and running them through an AI model simultaneously in a single operation.

Cloud AI. Artificial intelligence hosted in the cloud and accessed remotely via an API, rather than run locally.

CO₂e. Carbon dioxide equivalent. A metric that expresses the global warming potential of various greenhouse gases as the equivalent amount of carbon dioxide.

Continuous Batching. The production standard for serving LLMs, in which requests are admitted to and retired from a running batch as they arrive and complete, rather than being grouped into fixed batches in advance.

Cradle-to-Gate. The scope of an embodied emissions assessment covering raw material extraction, manufacture, and transport up to the point of sale, excluding the use phase.

Decode. The final, autoregressive phase of AI inference where the LLM generates output tokens sequentially. This phase is bound by memory bandwidth rather than compute.

Dense Model. A model architecture in which every parameter participates in each forward pass, as distinct from a mixture-of-experts architecture.

Electricity Generation Water Consumption. The water consumed during the generation of the electricity that inference draws, reflecting the regional generation mix. Classified as Scope 2.

Embedding. A dense numerical vector that captures the meaning of a token, word, or passage. Embedding models convert text into these vectors for use in search, retrieval, and similarity tasks.

Embodied Emissions. The greenhouse gas emissions associated with the manufacture, transport, and end-of-life processing of server hardware, expressed in CO₂-equivalent terms and amortised over the hardware's assumed operational lifespan.

Epistemic Uncertainty. The uncertainty in the predictive model itself, arising from how densely the benchmark data covers the region being predicted. It reduces as more models are benchmarked.

Grid Carbon Intensity. The average emissions produced per unit of electricity consumed from a given electricity grid, varying by region and time. Abbreviated as GCI.

Inference. The process by which a trained AI model applies its learned knowledge to new, unseen data to make predictions, decisions, or generate outputs.

Input Token. A token fed during the prefill phase into an LLM via API calls or local inference engines, converted from documents, images, code, and other inputs.

Interquartile Range. The spread between the 25th and 75th percentiles of a distribution, used in this methodology to express the range around a reported median. Abbreviated as IQR.

IT Cooling Water Consumption. The water consumed on site in cooling the hardware that serves inference, for example through cooling loops, evaporative cooling, and humidification. Classified as Scope 1.

KV Cache. The key-value cache, being the intermediate attention states stored during prefill and reused for each token generated during decode. Its memory footprint scales with the number and length of concurrent requests.

kWh per Token. The primary unit of energy measurement used by this methodology, representing the kilowatt-hours of electricity consumed per token processed or generated during inference.

LLM Parameters. Numerical values within the neural network, learned during training, that determine the connection strengths between neurons and encode the model's learned knowledge. The active parameter count refers to the portion of the model used per request and is relevant for mixture-of-experts architectures.

Mixture of Experts. A model architecture in which a routing network selects a subset of the model's parameters for each token, so that the parameters active per request are fewer than the model's total. Abbreviated as MoE.

Output Token. A token generated during the decode phase by the LLM in its response, converted into text.

Overlap. The phase of inference in which one request's prefill runs in the same forward passes as another request's decode, which occurs when an inference engine interleaves concurrent requests.

Parallel Requests. The number of requests a model processes simultaneously during a batched forward pass. The average achieved in production can sit well below the configured maximum, and energy per output token falls as it rises.

Parameter Band. A bounded range of total parameter counts to which a proprietary model is assigned, in place of a single estimated parameter count, as described in Section 6.4.

Power Usage Effectiveness. A ratio expressing the energy efficiency of a data centre, defined as total facility energy divided by the energy consumed by its IT equipment. Abbreviated as PUE.

Prefill. The initial, compute-intensive phase of AI inference where a model processes the entire input prompt simultaneously to understand context and token relationships.

Prefix Caching. An LLM inference optimisation that stores and reuses the computed context of common initial prompt segments to reduce redundant computation.

Quantisation. A compression technique that reduces the numerical precision of an LLM's parameters and activations, using less GPU memory and enabling faster inference at lower cost. FP8, INT8, and NVFP4 are the precisions covered by this methodology's benchmarking.

Statistical Uncertainty. The uncertainty arising from variation between comparable workloads and from the predictive models themselves, comprising the aleatoric and epistemic components.

Systematic Uncertainty. The movement of a reported value across the range of assumptions the methodology samples over, where a deployment characteristic cannot be observed.

Token. A discrete subword unit produced by splitting text, code, audio, or other input into chunks for processing by an LLM. Each token is mapped to a numerical identifier and then to a high-dimensional vector during model processing.

Total Facility Energy. The total energy consumed by a data centre, comprising both the energy used by IT equipment and facility overhead such as cooling, lighting, and power distribution. Abbreviated as TFE.

Training Energy Uplift. A multiplier applied to the operational energy of inference to account for the energy consumed in training the model, amortising that training cost across the tokens the model serves.

VRAM. Video random access memory, being the memory available on a GPU. The memory footprint a model's weights occupy in VRAM determines how many accelerators are required to serve it.

Water Usage Effectiveness. The ratio between the water consumed by data centre systems and the energy consumed by the IT equipment, expressed in litres per kilowatt-hour. Abbreviated as WUE.