Modeling AI-Driven IoT Energy Consumption: From Device-Level Forecasts to System-Level Dynamics

The rapid expansion of artificial intelligence workloads—particularly at the network edge—has led to a pressing need for accurate energy consumption forecasts in AI-driven Internet of Things (AIoT) ecosystems. In this paper, we develop and calibrate two complementary forecasting frameworks: (1) a bottom-up device-level model that aggregates per-device power draws P<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">j</inf>, active device populations N<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">j,t</inf>, and overhead functions f(D<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">t</inf>,M<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">t</inf>) to estimate total energy consumption across heterogeneous device classes; and (2) a top-down workload-level model that leverages aggregate inference counts N(t) and time-varying energy per inference e(t) (following Koomey’s Law) to project overall energy usage. By fitting both exponential and logistic growth formulations to 2025 baseline parameters (e.g., 2 × 10<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">9</sup> edge AI-capable devices, average P<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">avg</inf> = 7.5W, e<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">inf</inf> = 0.154mWh/inference), we obtain convergent 2025 energy forecasts of approximately 112–131 TWh / year, representing 0.4–0.45% of global electricity generation. We perform sensitivity analyses to quantify how small errors in growth-rate estimates amplify forecast uncertainty and demonstrate that integrating logistic constraints (device saturation K<inf xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">j</inf> and efficiency improvement rate ρ ≈ 30%/year) yields more realistic long-term projections. Finally, we outline an agent-based modeling framework to capture spatiotemporal heterogeneity and discuss policy levers—such as efficiency incentives and deployment caps—to ensure sustainable AIoT expansion.