Signal Energy
The user side of the AI energy equation
What It Is
Signal Energy is the applied energy research division of Signal Literature. It studies an overlooked source of AI cost: unnecessary computation generated during the interaction itself.
Every AI response consumes energy. Longer outputs generally require more decoding work, which means that padding, repetition, drift, and unnecessary expansion can increase both token cost and energy use. Signal Energy asks whether better language and stronger session governance can reduce that waste without reducing answer quality.
The AI industry is improving chips, models, data centers, routing, and caching. Signal Energy works on another part of the system: the language entering the model and the structure governing the exchange.
The Demand Layer
The Demand Layer is our framework for studying how the form of a request influences the computational demand of the response.
Fragmented, ambiguous, or contradictory input can require clarification, qualification, repetition, or additional reasoning. Clearer input can sometimes allow a model to resolve the same task with fewer turns and fewer generated tokens.
This is not presented as a universal law. It is a measurable interaction-level hypothesis:
Can better-structured exchanges produce equal or better outcomes with less generated output?
From Human State to Observable Language
Signal begins on the human side, but Signal Energy measures what can actually be observed.
A person’s internal state is not directly measurable from a prompt. The language produced from that state is. We examine features such as clarity, contradiction, continuity, scope, correction memory, and task definition, then measure what follows:
total output tokens
number of turns required
repetition and drift
task completion
latency and cost
energy use where direct measurement is available
The focus is not simply shorter answers. It is less unnecessary computation for the same or better result.
AXIS
AXIS is a session-governance system developed under the Signal Literature research program. It is designed to preserve intent, constrain unnecessary expansion, retain corrections, and reduce conversational drift across extended interactions.
AXIS can be evaluated against an unmodified baseline using the same model and task set. Relevant outcomes include:
output-token reduction
turn reduction
task accuracy
correction retention
drift frequency
latency and estimated or directly measured energy use
Because AXIS operates at the interaction layer, it can complement model-level and infrastructure-level efficiency methods rather than replace them.
Why It Matters
Data-center electricity demand is rising rapidly. The International Energy Agency projects that global electricity consumption by data centers could exceed 1,000 terawatt-hours annually by 2030, with AI playing a major role in that growth.
Inference energy remains difficult to measure consistently across models, hardware, workloads, and deployment environments. New benchmarks are being developed precisely because the field still lacks standardized measurement.
Signal Energy addresses one narrow but practical question within that larger problem:
How much unnecessary computation can be removed by improving the structure of the interaction itself?