Skip to main content
Infrastructure-Level Observability

Carbon Observabilityfor AI Workloads

You monitor latency and errors. Now monitor your carbon footprint. Get real-time CO₂e metrics for every request, automatically calculated based on model architecture and grid intensity.

Start Monitoring Read the docs

The Missing Metric in AI Ops

Invisible Costs
AI compute is energy-intensive, but providers hide the impact. You can't manage what you can't measure.
Scale Multiplier
A single agent run might be negligible, but millions of autonomous runs create a massive footprint.
Compliance Risk
New regulations (CSRD, SEC) require reporting Scope 3 emissions from digital supply chains.

What we track

Note: These are estimates. Actual impact varies based on hardware, load, and real-time grid conditions. See docs for methodology details.

Model parameters

Energy scales with model size and architecture:

  • Parameter count (7B vs 70B vs 400B)
  • Architecture type (dense vs MoE)
  • Inference optimization level

GPT-4o-mini: ~0.001g CO₂e/1K tokens

Provider infrastructure

Carbon intensity varies by data center location:

  • Grid carbon intensity (gCO₂/kWh)
  • Renewable energy percentage
  • PUE (Power Usage Effectiveness)

Iceland (hydro): ~20g/kWh vs US avg: ~400g/kWh

API response data

Every response includes carbon metrics:

  • Estimated CO₂e for the request
  • Model used and token count
  • Provider and region info

x-agentlify.co2e: 0.0023g

Energy per token

Estimated from model parameter count and architecture type (dense, MoE, reasoning)

Dashboard analytics

Aggregate data in your dashboard:

  • Total CO₂e by time period
  • Breakdown by model and router
  • Export CSV/JSON for reports

Monthly report: 2.3kg CO₂e from 15M tokens

Why Environmental Optimization Matters

50%+
Potential carbon reduction by routing to efficient models
10x
Carbon difference between energy sources by region
100%
Transparency into every request's environmental cost

Calculation methodology

1

Final calculation

energy_per_token × tokens × PUE × grid_intensity

2

Model Energy Consumption

We estimate energy usage based on model parameters, architecture efficiency (dense vs. sparse/MoE), and token throughput. Larger models consume more energy per token.

3

Provider data

Datacenter PUE and regional grid carbon intensity from public sources

4

Regional Carbon Intensity

Using real-time carbon intensity data from electricity grids, we convert energy consumption to CO₂e emissions. A model running in Iceland (geothermal) has far lower emissions than the same model in coal-heavy regions.

Start tracking carbon today

Free tier includes carbon tracking. No credit card required.

Get API keyView docs

No credit card required • Track impact from day one