The Thermodynamic Efficiency Inversion: A Comparative Energy Lifecycle Assessment of Generative AI Inference versus Ad-Supported Web Search Sessions

Abstract

The dominant environmental narrative around generative AI treats each large language model (LLM) prompt as an energy-intensive event and compares it to a traditional web search query at the server level. This paper argues that such comparisons use the wrong functional unit. The relevant unit is not query-level computation, but task-level session energy: the total energy required to satisfy one user information need across the full retrieval stack. We therefore conduct a full-stack Comparative Energy Lifecycle Assessment (CELCA) of two modalities for complex information-seeking tasks: direct LLM synthesis and ad-supported web search.

The analysis includes server-side computation, mobile or fixed-network transmission, client-device energy during active use, webpage rendering, and the client-side burden of programmatic advertising — costs largely avoided by direct LLM responses. Using Google's peer-reviewed production benchmark for Gemini inference (arXiv:2508.15734, 0.24 Wh per median text prompt), HTTP Archive 2025 mobile page-weight data, Nokia mobile-network intensity estimates, CHI '25 experimental task-completion evidence (Spatharioti et al. (2025)), and a 10,000-draw Monte Carlo sensitivity analysis, we model matched user sessions rather than isolated queries.

The central finding is conditional but substantial: for complex synthesis tasks performed on mobile connections, a standard-LLM session consumes approximately 4–9× less energy than an equivalent ad-supported web-search session, with a central estimate of 5.4× and a worst-case floor of 1.6× across all modelled parameter combinations. This advantage narrows to parity for simple zero-click queries on fixed Wi-Fi, and reverses for reasoning models and agentic workflows utilising test-time compute. The paper's contribution is therefore methodological as well as empirical: when energy accounting shifts from queries to tasks, the conventional LLM-versus-search narrative is materially altered.

Keywords: Lifecycle Assessment, LLM Inference Energy, Programmatic Advertising, Mobile Network Energy, Sustainable AI, Information Retrieval, Generative AI, Large Language Models

JEL Classification: Q40, Q55, L86, O33