Skip to content

Energy data licenses and attributions

Venturi’s energy, carbon, and water estimates are built on open methodologies and public datasets. This page credits those upstream sources and states exactly what Venturi reuses, reimplements, and does not redistribute.

The sections below describe the published methods, source-specific measurements, and attribution requirements used by the energy catalog. Each available catalog estimate or measurement records its basis in provenance, source_citations, and any source_url; entries without a matching source-backed energy value retain null energy.

The public catalog 2026-06-13.ap2 contains 51 entries: 25 modeled estimates from How Hungry is AI?, two AI Energy Score measurements, three Power Hungry Processing measurements, and 21 entries without an aggregate energy value. Catalog coverage and energy availability are separate measures.

AI Energy Score

The AI Energy Score is a standardized, multi-organization model-energy benchmark led by Hugging Face, with contributors from Salesforce, Neuralwatt, Cohere, Meta, and Carnegie Mellon University. Venturi bundles a versioned copy of its published coefficients and ratings as the primary energy tier (T1) and records the benchmark’s run-config provenance with every coefficient, as described in the energy methodology. The benchmark source is released under the MIT License, and the leaderboard Space declares Apache-2.0. The AI Energy Score name and label are used as reference material. Its documentation allows standalone star ratings in other applications provided they link back to the main leaderboard, so every standalone rating Venturi shows links back to it.

CodeCarbon: MIT License

CodeCarbon is an open-source emissions-tracking library released under the MIT License. Venturi uses CodeCarbon as the optional component-energy backend (tier T0) for self-hosted deployments (ingesting its emissions.csv, BoAmps JSON, and Prometheus metrics). From the CodeCarbon methodology, Venturi adopts four elements: carbon as grid carbon intensity multiplied by energy (C × E), the lookup of carbon intensity by cloud provider or by country, the per-fuel factor table for deriving intensity from a generation mix, and the world-average fallback. The explicit override, the optional Electricity Maps step, the five-step order of precedence, and deriving T0 power from energy-counter deltas (power_kw = Δenergy_kwh / Δhours) are Venturi’s own method, specified in the energy methodology. The CodeCarbon methodology credits its sources: the open-source energy-usage project for the fossil-fuel factors, the World Nuclear Association’s comparison of lifecycle emissions for the low-carbon factors, and the International Energy Agency’s Global Energy & CO2 Status Report 2019 for the 475 gCO2e/kWh world average. CodeCarbon’s own bundled datasets carry their own upstream terms (see Our World in Data and cloud-provider intensities below), which are preserved on any redistribution.

Our World in Data: grid carbon intensity

Per-country grid carbon-intensity and generation-mix figures originate from Our World in Data, as vendored through CodeCarbon’s grid-intensity datasets. Where Venturi redistributes these per-country intensities in its catalog, the Our World in Data attribution and the source dataset’s license are preserved. The current OWID Energy dataset includes the per-country electricity-intensity series and its codebook. Its license statement makes OWID-produced data available under CC BY 4.0; third-party data retains the original authors’ terms. Cloud-provider per-region intensities retain their own source links and terms.

How-Hungry-is-AI: methodology reimplemented, data not redistributed

The infra-aware energy method (tier T2), the two-term water model, and the eco-efficiency DEA and its AI-Index weights follow the methodology published in How Hungry is AI? Benchmarking Energy, Water, and Carbon Footprint of LLM Inference by Nidhal Jegham, Marwan Abdelatti, Chan Young Koh, Lassad Elmoubarki, and Abdeltawab Hendawi (arXiv:2505.09598, version 6, revised November 24, 2025).

In public catalog 2026-06-13.ap2, the 25 howhungry_infra_aware_medium_prompt rows use the paper’s Table 4 modeled per-query means for the named model and workload. Each row cites Table 4 and the Table 1 infrastructure assumptions. These figures are modeled estimates for the cited research workload. The rows preserve the workload, measurement unit, source date, and provenance.

The paper is released under CC BY 4.0. Reused formulas, constants, and published figures retain attribution to the paper and the cited table. The provider factors applied to these modeled rows are listed under provider sustainability inputs.

What Venturi does not redistribute

Venturi does not redistribute the How-Hungry-is-AI project’s CSV datasets or its Power BI dashboard. The project’s repository publishes no license; its code and data files are not the stated source for the current catalog’s paper-modeled estimates.

The provider-multiplier discrepancy described for catalog 2026-06-13.ap1 belongs to that historical version; see version history. In the public 2026-06-13.ap2 catalog, all 25 paper-modeled rows cite Table 1 and use the corresponding factor sets reproduced below. These factor sets apply to those modeled rows.

Artificial Analysis: performance and benchmark results

Artificial Analysis publishes first-token latency, throughput, and benchmark results relevant to the infra-aware and AI-Index methodologies. Each benchmark remains the work of its own authors. Artificial Analysis data, including its free data API, is governed by the Terms of Use and Data Platform Terms. Those terms do not permit redistributing raw Artificial Analysis data, whether as bulk JSON or CSV exports or embedded in a customer-facing product, and require attribution wherever the data, or data derived from it, is shown outside the recipient’s own systems.

  • Current public catalog. In 2026-06-13.ap2, all 51 rows have null latency_s, tps, ai_index, and eco_efficiency_score fields. The 25 modeled-energy rows cite How Hungry is AI? Table 4 and Table 1; the five measured-energy rows cite AI Energy Score or Power Hungry Processing benchmark data. The catalog records the cited source and accounting scope for each available energy value.
  • Derived figures require attribution. Wherever a published figure uses Artificial Analysis data, retain the label “Based on data from Artificial Analysis”, a statement that Venturi created the derived figure and that it does not represent Artificial Analysis’s views or endorsement, and an Artificial Analysis reference in the row’s provenance.
  • Catalog history. Source and attribution notes for earlier catalog versions remain in version history.

Power Hungry Processing: published measurements

Power Hungry Processing: Watts Driving the Cost of AI Deployment? by Alexandra Sasha Luccioni, Yacine Jernite, and Emma Strubell (arXiv:2311.16863, version 3, revised October 15, 2024) describes the inference-energy and carbon study.

The three Power Hungry rows in public catalog 2026-06-13.ap2 use task-specific measurements from the pinned benchmark dataset, with the paper cited for methods. Each is an arithmetic mean over 30 runs for the exact model and task. The catalog preserves the dataset revision, source hash, and measurement dates in its citation and provenance fields. Energy is machine energy per inference; carbon is retained from the source-reported CodeCarbon result. The accounting scope is the benchmark machine.

The paper is released under CC BY-SA 4.0. Its attribution and ShareAlike requirements apply to reused paper material. The raw benchmark dataset is a separate source, identified by its pinned revision and hash. Its pinned README declares CC BY-NC 4.0; the paper license does not replace the source dataset terms.

Mistral AI lifecycle analysis

Mistral AI published a lifecycle analysis of Mistral Large 2, carried out with Carbone 4 and the French ecological transition agency (ADEME), in Our contribution to a global environmental standard for AI (July 22, 2025). It reports 1.14 gCO2e and 45 mL of water for a 400-token response from its Le Chat assistant, including upstream impacts such as hardware manufacturing. Because that figure includes embodied impacts, which are outside Venturi’s operational-inference scope, Venturi cites it only as a validation reference in the provenance of the rows that name it. It never supplies a row’s energy, carbon, or water value, and Venturi does not redistribute the report.

Electricity Maps: customer-supplied licensed feed

Electricity Maps is an optional live carbon-intensity source. You supply an API token from your own Electricity Maps subscription under Settings / Admin / Credentials; Venturi stores it as an encrypted credential reference, and you rotate it in your Electricity Maps account. Venturi does not bundle, resell, or publicly redistribute Electricity Maps data: values resolved with your token are used only in your tenant’s carbon calculations, and the public AI Energy Index catalog does not use them. Every resolved value records Electricity Maps as its source, with the fetch time. Your use remains subject to the current Electricity Maps terms and your subscription, including its use type (for example internal or external commercial use), its rules on who may use your API token, and its attribution requirement when data is shown outside your organization. Confirm that your subscription covers use of the token by your Venturi instance before you enable the feed.

Eco-efficiency and provider sustainability inputs

The AI-Index weights (reasoning and knowledge 50%, math 25%, coding 25%) and the DEA cross-efficiency method follow Section 4.5 of the CC BY 4.0 How Hungry is AI? paper. Public catalog 2026-06-13.ap2 contains no computed ai_index or eco_efficiency_score values.

The 25 paper-modeled rows in that catalog use the following environment-specific PUE, WUE, and carbon-intensity factor (CIF) values and cite the paper’s Table 1:

Host PUE WUE_site (L/kWh) WUE_source (L/kWh) CIF (kgCO2e/kWh) Sources Table 1 cites
Microsoft Azure 1.12 0.30 4.35 0.35 PUE: Microsoft Azure blog (April 2022); WUE_site: Microsoft Cloud blog (December 2024); WUE_source: World Resources Institute; CIF: Microsoft 2024 Environmental Sustainability Report (data fact sheet)
AWS 1.14 0.18 5.11 0.287 PUE and WUE_site: Amazon 2023 Sustainability Report; WUE_source: World Resources Institute; CIF: the Electricity Maps live carbon-intensity map
DeepSeek datacenters in China 1.27 1.20 6.016 0.6 PUE and WUE_site: Hequan Wu, China Communications Enterprise Association (April 2024); WUE_source: World Resources Institute; CIF: Ni et al., “CO2 emission-mitigation pathways for China’s data centers”, Resources, Conservation and Recycling 202 (2024)

The World Resources Institute source is its Guidance for calculating water use embedded in purchased electricity (2024). The table retains the paper’s values and credits both the paper and its cited sources. The AWS CIF of 0.287 is the paper authors’ figure derived from the Electricity Maps live map; it is separate from the customer-supplied Electricity Maps feed described above.

These factor sets describe the paper-modeled rows. Other catalog entries use their benchmark’s accounting scope or have no aggregate energy estimate. Provider metadata alone does not supply missing model energy: entries with unavailable energy retain null energy, water, and carbon outputs.

Equivalence sources

The public viewer’s fleet-scale equivalence cards cite their own sources: the U.S. Energy Information Administration for household electricity use, the U.S. Environmental Protection Agency for passenger-vehicle emissions, and the World Aquatics Competition Regulations for Olympic pool dimensions. The divisors and links are listed on the Public AI Energy Index page.

Summary

Uses of the same sources:

Source Used for
AI Energy Score Two measured-energy rows in public catalog 2026-06-13.ap2
CodeCarbon T0 component-energy ingestion + carbon method and fuel factors
Our World in Data Per-country grid intensity
How-Hungry-is-AI 25 modeled-energy rows; Table 1 factors; water and eco-efficiency methodology
Artificial Analysis Performance and benchmark input attribution rules
Power Hungry Processing Three task-specific measured-energy and carbon rows
Mistral AI lifecycle analysis Validation reference in row provenance
Provider disclosures cited by Table 1 PUE, WUE, and CIF for the 25 paper-modeled rows
Electricity Maps Optional live grid intensity

Licenses and bases for the same sources:

Source License / basis
AI Energy Score MIT source; Apache-2.0 leaderboard Space
CodeCarbon MIT
Our World in Data Energy dataset license: OWID-produced data CC BY 4.0; upstream terms preserved
How-Hungry-is-AI Paper under CC BY 4.0; repository publishes no license
Artificial Analysis Artificial Analysis Terms of Use and Data Platform Terms
Power Hungry Processing Paper: CC BY-SA 4.0; pinned raw dataset README: CC BY-NC 4.0
Mistral AI lifecycle analysis Mistral AI publication
Provider disclosures cited by Table 1 Used as published in the CC BY 4.0 paper
Electricity Maps Your subscription and use type

Redistribution conditions for the same sources:

Source Redistributed?
AI Energy Score Pinned leaderboard measurements, exact evaluated identities, source links
CodeCarbon Method adopted; bundled data per its upstream terms
Our World in Data Yes, with attribution and source license preserved
How-Hungry-is-AI Paper Table 4 estimates and Table 1 factors, with citations; historical ap1 discrepancies remain in version history
Artificial Analysis Raw-input and eco-efficiency fields are null in ap2; any figure derived from Artificial Analysis data requires attribution and a non-endorsement statement
Power Hungry Processing Aggregates from the pinned benchmark dataset, with dataset and paper citations; retain the dataset’s applicable source terms
Mistral AI lifecycle analysis Single figure cited with attribution; never a row value
Provider disclosures cited by Table 1 Published factor values, with paper and source citations
Electricity Maps No; used only in your tenant

Where to go next

Give feedback on this page

Sign in to send feedback. Opens in a new tab.

Copy page details 
Docs feedback: Energy data licenses and attributions
Page: https://docs.venturi.systems/energy/licenses/