Results and methodology
Aitra Meter produces neutral, reproducible measurements of AI infrastructure efficiency, from model to facility. This page explains what it measures, how the measurement is kept trustworthy, and how results will be published. It is not a leaderboard. Aitra does not rank, certify, or endorse vendors, models, or products.
What Aitra Meter measures Shipping
The core primitive is joules per output token, measured continuously across every workload, model, and GPU. Energy is read from GPU hardware and attributed to AI output volume at the token level, then combined with site configuration to derive cost and carbon.
| Metric | What it measures |
|---|---|
aitra_j_per_token | Joules per output token, per workload, model, and hardware. |
aitra_cluster_j_per_token | Cluster-wide J/token (total energy divided by total tokens). |
aitra_co2_per_token_grams | Grams of CO2 per token (J/token times grid intensity). |
aitra_cost_per_million_tokens_usd | Dollars per million tokens (J/token times electricity cost). |
aitra_idle_power_watts | GPU power draw when no inference is running. |
aitra_tokens_per_joule | Inverse efficiency, output produced per joule. |
aitra_measurement_cv | Rolling coefficient of variation, a stability signal for the measurement itself. |
The full list, with types and labels, is in the metrics reference.
How the measurement stays trustworthy
A result is only as good as the measurement behind it. Aitra Meter is built so that each number can be trusted and checked:
- CV gating. A rolling coefficient of variation is tracked over recent measurement windows. Unstable measurements are flagged rather than reported as clean results. See the measurement-unstable runbook.
- Calibration and attribution. The aggregation service calibrates energy and attributes it to the workloads and models that consumed it, rather than assuming a flat split.
- Pluggable energy sources. NVML, AMD, DCGM, and Zeus providers measure the same primitive from different hardware paths, documented in the energy providers guide.
- Open specification. The measurement model is written down in the technical specification and refined through Architecture Decision Records.
How to reproduce a measurement
Everything needed to reproduce the measurement on your own hardware is open source today:
- Install Aitra Meter with one Helm command. See Try Aitra Meter or the getting started guide.
- Use a published example site configuration, for example site-config-singapore.yaml or site-config-airgapped.yaml.
- Follow the build, deploy, and test guide to run the stack end to end.
- Read the raw metrics from Prometheus, or use the pre-built Grafana dashboards shipped with the chart.
How results will be published
When a public results set is released, it will follow the same principles that govern the project:
- Neutral and reproducible. Every figure will link to the exact hardware, inference server, model, and site configuration used, plus the commands to reproduce it.
- No ranking. Results describe energy and efficiency behavior, not a comparison that places one vendor, model, or product above another.
- Measurement-gated. Only measurements that pass the stability gating described above will be published as results.
Follow the roadmap and changelog for when results land. You can also subscribe for updates on the home page.