How much energy⇄ does AI inference use?

This dashboard presents a new methodology for understanding the environmental impact of closed AI models. It combines the latest peer-reviewed science, new empirical testing across 1,700+ tests and 230 million generated tokens, and real-world application to corporate carbon inventories.

CLEER at the core

CLEER (Closed-model Latent Energy Estimation Range) estimates per-token energy for proprietary models, where direct measurement is not possible. We benchmark open models on known hardware, match each closed model to the closest open proxies based on observed speed and behaviour, then project onto those measured power curves. Learn more → CLEER-Text-0926 Technical Report

CLEER estimates energy per token (input/output), the unit customers see in a real inventory. → Reach out for per-token factors.

Use case

Because tokens are difficult to interpret, this dashboard reports whole sessions. The 2 profiles draw on public production data: chat uses ShareChat conversations and Qwen serving traces; agentic uses AgentX, a corpus of real coding-agent sessions.

Note: In reality, verbosity (number of output tokens for the same query) differs between models. This dashboard holds token counts constant across every model so the comparison isolates per-token energy.

From chip to facility

A model's accelerator energy is only part of a data center's energy use. Below figures account for the rest: non-accelerator components, idle capacity, and cooling / power distribution. In reality, these factors vary by configuration, operator, and site. This dashboard shows 1 documented scenario, though the methodology can accommodate different assumptions for each. → Emissions methodology

Carbon

Carbon emissions depend on how and where electricity is generated, with results varying by more than 10× across possible grids. The methodology is designed to work across different emissions assumptions; the dashboard shows 1 scenario: behind-the-meter gas generation at 730 gCO₂e/kWh, used in place of location-specific grid data, together with US-average facility efficiency and embodied emissions from hardware and data center construction. Different grid assumptions can produce substantially different results. → Emissions methodology