New model estimates per-request HTTP carbon emissions without power telemetry

In a new paper presented at LIMITS 2026, authors Geerd-Dietger Hoffmann and Verena Majuntke tackle the practical challenge of generating per-request environmental impact data for HTTP services. While existing sustainability proposals focus on what information (e.g., carbon headers) should be transmitted, they leave open how to actually compute per-request values in production without fine-grained power telemetry. The team's solution: benchmark endpoints offline under controlled conditions, derive compact endpoint-specific energy models from observable request features, then evaluate those models online at the HTTP server boundary.

The mechanism is implemented as an nginx extension that loads a JSON model registry and emits per-request metadata for energy, grid intensity, embodied emissions, and total request-level impact. The approach supports constant, linear, and piecewise models for heterogeneous request classes, and extends to endpoints where the dominant cost driver is only visible at the application layer (e.g., token counts in AI inference). Evaluation confirms operational feasibility with low runtime overhead, making it practical for real-world deployment at scale.