AI Meter tracks local token use, estimates power and water impact

AI Meter is a locally-run macOS application designed to measure token consumption across coding agents and translate that usage into estimates of electricity and water impact. Installation is simple—a single shell command downloads and installs the tool—and it operates entirely on your machine; no accounts are required and no usage data leaves your Mac.

The tool aggregates token counts from five major coding providers. It reads local provider files from Codex, Claude Code, Cursor, OpenCode, and Gemini CLI, deduplicates the counts, and combines them on your device. From that total, it derives two environmental estimates. For electricity, the default formula divides total tokens by 1 million and multiplies by 0.39 facility kWh; the tool notes an uncertainty range of 0.20–0.75 kWh per million tokens. For direct cooling water, the calculation takes the estimated energy, divides by a Power Usage Effectiveness (PUE) factor of 1.20, and multiplies by a Water Usage Effectiveness (WUE) of 0.45 liters per IT kilowatt-hour. Both the PUE and WUE values are adjustable inside the application, allowing users to tune estimates for different facility types.

The methodology is explicitly scoped. AI Meter excludes model training costs, embodied hardware impact, off-site electricity water, grid carbon, networking overhead, and user-device energy consumption. The estimates themselves are based on published research and represent local scenarios—not measurements from OpenAI, Anthropic, or other providers' actual data centers. This transparency is intentional: the tool follows the Sustainable Computing Initiative's framework for consumer-boundary analysis, making the assumptions explicit and editable rather than treating them as black-box calculations.

AI Meter addresses a gap in developer visibility into the environmental footprint of AI inference. As coding agents become embedded in developer workflows, the cumulative token usage across tools like Cursor and Claude Code is typically invisible; this tool surfaces that usage and attaches concrete resource estimates. The methodology is transparent: the tool follows a Sustainable Computing Initiative (SCI) framework aligned to consumer boundaries, meaning it measures only local inference activity and its direct electricity and water requirements, not the upstream costs of model training or hardware embodiment. The electricity estimate (0.39 kWh per 1M tokens) and water estimate (0.45 L per IT kWh at a 1.20 PUE) are drawn from published research rather than proprietary data-center measurements, which means they represent industry-typical scenarios rather than precise readings from OpenAI, Anthropic, or other providers. By making these estimates adjustable within the app, AI Meter invites users to explore sensitivity to different assumptions about facility efficiency and cooling practices.

No comments yet. Be the first to share your thoughts!