Skip to content

Energy & Carbon Tracking

Every LLM call has an energy cost. BIJOTEL estimates Wh and gCO₂ per call based on model and token count, then aggregates per region's grid intensity.

Why

  • Carbon reporting for ESG / sustainability claims
  • Cost-of-energy as a routing signal (some models are 8× more efficient per token)
  • Provider migration justification — show the real CO₂ delta before committing

Backfill existing chain

If you already have a chain.db from before energy tracking was wired, run a one-shot backfill:

bijotel energy backfill --db chain.db --region us-east

This populates an energy_log table linked to chain.seq. Idempotent — re-running doesn't double-count.

REST API

curl http://localhost:8080/api/energy/summary

{
  "total_wh": 19.95,
  "total_co2_grams": 7.58,
  "total_calls": 5459,
  "by_model": {
    "claude-haiku-4-5-20251001": {"wh": 8.21, "co2_grams": 3.12, "calls": 4102},
    "claude-sonnet-4-20250514":  {"wh": 11.74, "co2_grams": 4.46, "calls": 1357}
  },
  "by_region": {
    "us-east": {"wh": 19.95, "co2_grams": 7.58, "calls": 5459}
  }
}

Estimate energy for a hypothetical call:

curl -X POST http://localhost:8080/api/energy/estimate \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-haiku-4-5-20251001","tokens":1500,"region":"us-east"}'
# → {"wh": 0.0015, "co2_grams": 0.00057}

Real production example (GENA)

15 days of operation, 5,459 LLM-call rows in energy_log:

  • Total: 19.95 Wh / 7.58 g CO₂
  • Per call avg: 3.66 mWh / 1.39 mg CO₂

The migration from Sonnet to Haiku on 2026-05-21 cut daily CO₂ ≈ 8× — captured retroactively, not designed in. The signal lives in the chain forever and is auditable.

Energy model

The estimate is intentionally simple — Wh per 1K tokens, looked up per model, then converted to gCO₂ via region intensity:

wh = (tokens / 1000) * rate_wh_per_1k_tokens[model]
co2_grams = (wh / 1000) * grid_intensity_g_per_kwh[region]

The rate dictionary DEFAULT_ENERGY_RATES and the regional grid intensity DEFAULT_GRID_INTENSITY are public constants in bijotel.layers.energy:

from bijotel.layers.energy import DEFAULT_ENERGY_RATES, DEFAULT_GRID_INTENSITY

DEFAULT_ENERGY_RATES["claude-haiku-4-5-20251001"]  # 0.001  (Wh per 1K tokens)
DEFAULT_GRID_INTENSITY["us-east"]                  # 380.0  (gCO₂ per kWh)

Override with your own dicts for higher fidelity.

Supported regions

Region gCO₂/kWh Notes
us-east 380 Virginia (Anthropic, AWS us-east-1)
us-west 250 Oregon
eu-west 300 Ireland
eu-central 350 Germany / Frankfurt
eu-north 30 Sweden / Norway (hydropower)

Custom regions: pass intensity_overrides={"my-region": 120.0} to CarbonCalculator.

Accuracy

R3-E4 production test: 20/20 = 100% exact match between energy_log rows and manual recompute from DEFAULT_ENERGY_RATES × DEFAULT_GRID_INTENSITY.

The rates themselves are estimates (we don't measure your data center). What this validates is that the engine's storage matches its config — i.e. no silent rate drift between live calculation and persisted records.

Next