CLEER Dashboard Estimates the Energy and Carbon Impact of AI Inference

Summary

The CLEER dashboard presents a method for estimating the environmental impact of AI inference, including energy, carbon, water, and embodied emissions, when proprietary models cannot be measured directly. CLEER combines peer-reviewed research, large-scale empirical testing, and corporate-emissions use cases. Its underlying evaluation includes more than 2,400 test runs and over 425 million tokens, described as the largest production-representative open-model energy test to date. The method benchmarks open models on known hardware, matches closed models to proxies using observed performance, and projects their energy use onto measured power curves. It converts per-token estimates into representative chat and agentic sessions based on public ShareChat conversations, Qwen serving traces, and AgentX coding-agent sessions. Models receive identical token workloads, so comparisons primarily reflect energy intensity rather than differences in verbosity, although real-world results still depend on how many tokens a model generates. Reported impacts include accelerator and host-server energy, idle capacity, cooling, power distribution, and embodied emissions from hardware and data-center construction. The default emissions scenario assumes behind-the-meter gas generation at 640 gCO₂e/kWh and US-average data-center conditions with a PUE of 1.45; both settings can be changed. The dashboard offers a CLEER data feed for integration, while API access is listed as forthcoming. Its figures may be reused with attribution for non-commercial purposes under the stated licenses.