The LLM Footprint site introduces an interactive calculator for estimating the electricity use and carbon emissions associated with AI workloads. Users enter input and output token totals, set the share of cached input, and optionally provide request counts. The calculator offers three illustrative inference classes: Flash, represented by Qwen3.8-Flash-Next; Frontier, represented by DeepSeek V4 Pro; and SOTA, represented by Kimi K3. These are model proxies rather than measurements of the user’s actual model. Its energy equation separately accounts for fresh input, cached input, and output tokens, with cached input modeled at about 56% of fresh-input energy and output at about 10 times the per-token cost. Serving profiles adjust the estimate for throughput-first, balanced, or latency-first operation, while the displayed range reflects those profiles rather than a statistical confidence interval. Carbon is calculated separately by multiplying estimated electricity by a selected grid-intensity scenario, including U.S. average, California, West Virginia, or a custom value. Presets cover illustrative chat and agent workloads as well as reported token totals from Anthropic’s compiler experiment and Bun’s Zig-to-Rust rewrite, but the site warns that these examples are not typical workloads and that the coefficients are modeled, provisional, and subject to uncertainty from architecture, hardware, context length, routing, and other serving details. The tool also provides everyday energy and emissions comparisons, which it labels as alternative yardsticks rather than additive quantities.
