Energy and carbon attribution¶
The attribution engine does more than tell you who and how much; it also estimates how much energy each AI invocation consumed and how much carbon it emitted. Energy and carbon are attributed across the same attribution graph as cost, so you can see per-team, per-model, and per-cost-center energy and emissions alongside spend.
Reference methodology and scope
Energy and carbon estimates provide sustainability context; they are not carbon-accounting or compliance outputs. Unknown terms return null, never zero, and energy is never restated as cost. See the release-state authority.
This page explains how those estimates are produced, where they come from, and how Venturi keeps them honest.
What Venturi estimates¶
For each AI request the engine observes, it attaches an energy and carbon block to the invocation:
| Estimate | What it is |
|---|---|
| Estimated energy (kWh) | The GPU energy consumed by the request. |
| Energy Score rating (1–5 stars) | The model’s relative energy efficiency for its task type. |
| Estimated carbon (gCO2e) | The emissions from that energy, given the deployment region’s grid intensity. |
These flow into your dashboards, your FOCUS exports, and the optimization recommendations Venturi makes.
How energy is estimated¶
Venturi estimates GPU energy from three inputs:
- Request characteristics: request count. The current public-catalog field
is named
energy_wh_per_1k_output_tokensfor compatibility, but its published value is Wh/query and is not multiplied by token count. - Model identity: which model served it.
- Per-model energy coefficients: drawn from the AI Energy Score leaderboard.
The AI Energy Score is a standardized, multi-organization benchmark maintained by Hugging Face, Salesforce, Cohere, and Carnegie Mellon University. It measures AI model inference energy on NVIDIA H100 GPUs across 10 task types: text generation, summarization, extractive QA, text classification, sentence similarity, image classification, object detection, speech recognition, image generation, and image captioning. It assigns each model a 1-to-5 star rating for relative efficiency within its task category. Venturi bundles a versioned copy of the catalog and looks up each invocation’s model to derive its energy estimate.
The result records how it was derived (the energy derivation method), so the provenance of every energy number is auditable. Benchmark coefficients are measured at batch size 1 and are not throughput-representative of production serving; see the methodology for the run configuration and its limits.
How carbon is estimated¶
Each invocation with a non-null energy estimate carries an estimated carbon emission, computed as:
Venturi resolves the grid carbon intensity in a fixed 5-step order of precedence, the same chain the energy methodology defines, so the most specific value available is always used:
- An explicit override supplied for the calculation.
- A live grid-intensity feed (optional ElectricityMaps integration).
- The cloud provider and region (vendored per-region intensities).
- The country ISO code, via the generation-mix derivation.
- The world-average fallback of 475 gCO2e/kWh from the current CodeCarbon methodology, always flagged as estimated, never a silent value.
Set your region for accurate carbon
A value you configure resolves higher up the chain. Configuring your actual deployment region, or supplying a region-specific intensity, makes carbon estimates materially more accurate; configuration changes take effect without a restart.
The honest-unknown rule: null, never zero¶
Venturi only estimates energy and carbon for models it can find in the catalog. For any model that is not in the catalog, both estimates are returned as null, never zero.
Null is not zero
A model absent from the energy catalog produces a null energy estimate and
a null carbon estimate. Venturi never substitutes zero, because zero would
falsely imply no energy was consumed. This honest-unknown discipline is the
same principle that governs attribution confidence:
Venturi reports what it does not know rather than fabricating a number.
In dashboards and exports, these nulls are shown with explicit honest-uncertainty treatment so an uncatalogued model is visibly unrated, not silently counted as emission-free.
Where you see energy and carbon¶
| Surface | What it shows |
|---|---|
| Dashboards | Per-team total energy, per-model efficiency on the catalog’s per-query basis, and per-cost-center allocation alongside cost. |
| FOCUS export | Energy and carbon columns: estimated energy (kWh), Energy Score rating, and carbon (gCO2e), added to the standard FOCUS export. Carbon is non-null for catalogued models and null for uncatalogued ones. |
| Energy budgets | Per-team and per-cost-center energy caps (kWh per period), evaluated with the same advisory and gate semantics as cost budgets. |
| Optimization recommendations | When a more efficient model exists for a task, recommendations include the star ratings of the current and proposed models, the energy differential in kWh per 1,000 requests, and the annualized carbon impact of switching. |
Energy is not cost¶
Never read an energy multiplier as a cost multiplier
Energy efficiency varies enormously between models and modes: the same model can vary by a large factor depending on configuration, and reasoning-enabled operation is far more energy-intensive than standard operation. These figures are energy-transparency context only. Energy is not cost. Never restate an energy multiplier as a cost range: cost comes from the billing and pricing path, energy comes from this engine, and the two are reported separately.
Where to go next¶
For the full formulas, constants, and provenance behind the numbers above:
- Energy & carbon methodology: the T0–T4 tier ladder, the watt-hour primitive and star bands, the AI Energy Score run-config, the reasoning premium, the 5-step carbon resolution chain, and the fuel factors.
- Water methodology: the two-term per-query water model.
- Eco-efficiency methodology: capability per environmental cost, distinct from the energy star rating.
- Energy catalog version history: why a star can change without the energy changing.
- Public AI Energy Index: the public viewer and its explicitly unsigned current seed dataset.
- Energy & carbon FAQ: the common questions, answered.
And the surrounding engine docs:
- The attribution engine: how energy attaches to the same records as cost.
- Confidence and evidence: the honest-unknown discipline applied to attribution.
- The RAIL pipeline: how each invocation is resolved before energy is attached.