Skip to content

Energy catalog version history

Venturi distributes energy coefficients, carbon factors, and water-intensity reference data as a versioned catalog. This page identifies the active public version and distinguishes new measurements, updated estimates, source corrections, and changes to relative rating cohorts.

Published versions

Catalog version Generated Schema Rows Rows with energy Catalog coverage Public signature status Eco-efficiency scores
2026-06-13.ap2 2026-06-13 1.1.0 51 30 24.9% against the snapshot’s estimated 205-model reference population Unsigned marker; not an Ed25519 signature None published; eco_efficiency_score is null on all 51 rows

Signature status and score availability belong to each version. A later version states its own and leaves an earlier version’s row unchanged.

The active public catalog viewer manifest names aimodels.2026-06-13.json. That file contains:

  • 25 modeled coefficients, with derivation_method = howhungry_infra_aware_medium_prompt, from the published How Hungry is AI? v6 Table 4. They use Wh/query for the medium prompt fixture and include facility PUE.
  • Two measured AES coefficients, with derivation_method = aes_measured_sample, from the pinned February 2025 leaderboard. They use Wh/query, preserve GPU-only scope, and retain the exact evaluated checkpoints and official rating cohort.
  • Three measured Power-Hungry raw-dataset coefficients, with derivation_method = powerhungry_measured_per_1000_inferences. They use Wh/inference for their stated tasks and retain machine scope and separately source-reported carbon.
  • 21 rows without an energy coefficient. Their measurement kind is unavailable; pending, unestimated, or life-cycle-anchor provenance is not a numeric operational-inference estimate.

All 30 rows with energy have source_citations. energy_wh_per_unit, water_ml_per_unit, and carbon_g_per_unit are the native-unit fields; the legacy token-energy, per-query water, per-query carbon, and energy_score_rating fields are null. Two rows carry official_rating = 5 with official_rating_cohort = "AES February 2025". No row has an eco-efficiency score.

The populated modeled rows use the study’s Table 1 environments: Azure (PUE 1.12, WUE_site 0.30, WUE_source 4.35, CIF 0.35), AWS (1.14, 0.18, 5.11, 0.287), and DeepSeek-native (1.27, 1.20, 6.016, 0.60). WUE is in L/kWh and CIF in kgCO2e/kWh. The listed hosting environments and factors are study assumptions. Some unavailable rows still carry other reference multipliers; their energy, water, and carbon remain null, so those placeholders must not be presented as populated estimates.

The source_signature contains an UNSIGNED-SEED marker rather than an Ed25519 signature. Verify a signed distribution against its actual payload and signature. The manifest selects the active publication; its generated date and the per-row source_date provide separate freshness information. This version’s generated date is June 13, 2026, and its populated source periods are older. An active publication can contain older source measurements.

Historical ap1 documentation notes (June 13, 2026 catalog)

The earlier documentation for 2026-06-13.ap1 recorded these attribution and multiplier exceptions:

  • It described 24 T2 rows whose provenance did not name Artificial Analysis and noted that the public viewer did not display the attribution and non-endorsement statement discussed in licenses.
  • It reported Azure CIF values of 0.34 kgCO2e/kWh against the study’s 0.35, and AWS CIF values of 0.30 against 0.287. It also recorded Meta PUE of 1.1 against 1.14 and Amazon WUE_source of 5.12 L/kWh against 5.11.
  • It noted provider multipliers for Google, xAI, and Alibaba that those documentation references did not source.

These are the prior documentation’s recorded ap1 exceptions. Use the ap2 record above for the active public catalog’s populated values, source citations, and remaining unavailable-row reference inputs.

Recalibrate, not re-benchmark

Catalog changes have distinct meanings:

  • New measurement: a source measures the stated model and workload again. Record the source date, checkpoint, hardware, scope, and protocol; a new measurement need not change the numerical mean.
  • Estimate update: modeled inputs or assumptions change, producing an updated estimate rather than a new observed measurement.
  • Source correction: the source value, exact identity, denominator, or transcription is corrected. The current catalog explicitly marks its How-Hungry source corrections as not new measurements, and records replacement of unsupported earlier AES sample values with pinned source aggregates.
  • Rating recalibration: a separately derived comparison population changes while the underlying coefficients are preserved. Keep the population and rating method explicit.

A coefficient is versioned evidence with a stated unit and scope; measured and modeled values retain their different meanings. An upstream official rating remains tied to its source cohort. The current two AES ratings are from AES February 2025, not a new quintile ranking of this catalog.

Compare evidence

Compare energy only after checking the native unit, task, workload, checkpoint, hardware, scope, and protocol. Compare ratings only within their stated source or derived cohort. A changed value or star alone does not establish a change in a model’s real-world efficiency.

What every catalog version records

The current public catalog records:

  • schema_version, catalog_version, and generated_at.
  • source_signature, whose actual value must be interpreted. The active public file carries an unsigned marker. An Ed25519-signed distribution requires a real signature over canonical bytes and successful verification; see the signed-catalog procedure.
  • A coverage block with catalog-row totals, the snapshot’s reference population estimate, and models_with_energy. Measurement-kind and derivation-method counts can be computed from the model rows; they are not separate breakdowns in that block.
  • Per-row measurement_unit, measurement_kind, energy_scope, source citations, workload, protocol, exact benchmark identity, and provenance.
  • An optional official_rating and official_rating_cohort. A source cohort identifies an upstream rating; there is no per-star issue-date field in this public schema.
  • snapshot_date for the catalog snapshot and source_date for the source period. In this version every snapshot date is June 13, 2026, while populated source dates refer to 2023, February 2025, or November 24, 2025. The snapshot date must not be described as the date every coefficient was measured.

Retain the versioned artifact, source coordinates, native unit, and scope when reproducing a published figure. Keep the source period distinct from the date the catalog was generated.

The signed-distribution contract requires consumers to verify and pin a version, retain the last-good version when verification fails, and preserve previously published versions without overwriting their bytes.

Freshness

Venturi’s catalog watchdog checks provider release sources every week, on Mondays at 09:00 UTC. Net-new models are added as a new catalog version after review and after the catalog’s validation gates pass.

The freshness threshold equals that weekly poll interval. When the served catalog’s generated date is more than seven days old, the in-product status surface and the degradation API (GET /api/v1/system/degradation) report the catalog as stale, with its catalog_version, generated date, age, and the ratings it affects. Venturi never presents a stale catalog as current. For the public viewer, the generated date in the live manifest is the freshness authority.

Keep an upstream official rating tied to its recorded source cohort. If a derived population is recalibrated, record that population and method separately. Flag changed coefficients for provenance review: a numerical change may be a new measurement, an estimate update, or a source correction; it is not by itself proof of re-benchmarking.

Reading a version diff

When you compare two catalog versions:

  • Coefficient changed → inspect the unit, source, workload, protocol, and correction note to distinguish a new measurement from an updated estimate or source correction.
  • Coefficient unchanged, rating changed → inspect the rating method and source or derived cohort before attributing it to recalibration.
  • New rows → newly cataloged models; inspect availability before counting them as measured or estimated coverage.
  • derivation_method changed → the selected source or method changed. Check measurement_kind and provenance before calling it an upgrade to measured evidence.

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 catalog version history
Page: https://docs.venturi.systems/energy/version-history/