FUURAA AI Knowledge Library · Energy and carbon claim evaluation

How to evaluate AI energy and carbon claims

Use six evidence gates to turn an AI energy or carbon number into a comparable workload, functional unit, system boundary, measured energy, emission factors, hardware allocation, uncertainty and dated conclusion. Apply it to model reports, cloud disclosures, data-centre statements, procurement comparisons and public research.

Enter the six evidence gates ↓Open the failure matrix

Published24 August 2026Evidence statusMethod synthesis grounded in primary energy-system, software-carbon, data-centre efficiency, GHG-accounting and provenance sourcesScopeModel reports, cloud disclosures, data-centre statements, procurement comparisons and public research

Lower intensity is not automatically lower total impact

AI energy and carbon claims become comparable only when useful work, boundaries, measurement and accounting methods align.

The same AI system can show lower energy per task, higher total service demand and different location- or market-based emissions at once. Evaluation must start with workload and service quality, pass through energy measurement, facility overhead, electricity factors and hardware life cycle, then return to total impact and expiry.

Applicability boundaryThis is a public research method, not a FUURAA product-capability claim or an assessment of any model, provider, data centre, energy product, emission factor or environmental report. It does not replace energy or GHG expertise, meter calibration, life-cycle assessment, formal carbon accounting, assurance or audit, and is not legal, investment, procurement, certification or compliance advice.

Six rejectable evidence gates

Each gate must answer a decision question, produce minimum evidence and stop calculation or narrow the conclusion when material unknowns remain.

Freeze the exact claim, workload and service quality

Decision question: Is the claim about energy, power, carbon intensity, total emissions or avoided emissions; for which model, task, user outcome, quality target, scale, place and time?
Minimum evidence: Verbatim claim, claimant, date, workload definition, model and serving version, input and output profile, latency or quality target, volume, geography, period and intended decision.
Stop condition: Stop when a per-query figure omits task size or quality, a power reading is called energy, an intensity is presented as a total, or avoided emissions are mixed with inventory emissions.

Set the functional unit and system boundary

Decision question: What comparable unit of useful work is delivered, and which training, inference, storage, network, cooling, power, idle capacity and hardware life-cycle components are included?
Minimum evidence: Functional unit, service-quality constraints, process map, temporal and organisational boundary, included and excluded components, allocation rules, reserved capacity and boundary diagram.
Stop condition: Stop when unlike tasks share one denominator, facility overhead or networking disappears without disclosure, or operational and embodied emissions are silently exchanged.

Measure energy under a versioned operating state

Decision question: Where was energy measured, over what interval and load, with which meter, calibration, sampling, hardware, software, utilisation, PUE category and treatment of idle or failed work?
Minimum evidence: Meter location and identifier, calibration and accuracy, raw readings, timestamps, run logs, hardware and software versions, utilisation, repetitions, facility category, PUE inputs and excluded energy.
Stop condition: Stop when TDP, nameplate power or modelled estimates are presented as measured energy; measurement windows differ from workload windows; or the best performance and lowest energy come from different runs.

Convert energy to emissions with a named method

Decision question: Which electricity, fuel, refrigerant or other activity data was multiplied by which factor, for what grid region and time, using location-based, market-based, average or marginal accounting?
Minimum evidence: Activity data, factor publisher, identifier and version, geographic and temporal match, GWP basis, contractual instruments, residual mix, calculation code, units, rounding and separate results by method.
Stop condition: Stop when a global annual factor stands in for a time-sensitive local claim, market and location results are cherry-picked, offsets erase inventory emissions, or factor provenance is missing.

Include hardware allocation, uncertainty and sensitivity

Decision question: How are manufacturing and end-of-life emissions allocated, and how much do meter error, utilisation, lifetime, PUE, emission factors, geography, batching and service quality change the result?
Minimum evidence: Hardware inventory and life-cycle source, expected lifetime, time and resource shares, allocation formula, uncertainty interval, repeated measurements, sensitivity cases, quality results and dominant assumptions.
Stop condition: Stop when embodied emissions are declared zero because supplier data is absent, a single PUE or grid factor hides variation, or lower energy is achieved by degrading the useful result.

Separate efficiency, total impact, comparison and expiry

Decision question: Does the evidence support only a bounded intensity, a total inventory, a like-for-like comparison or a change over time, and what demand growth, system change or data revision triggers recheck?
Minimum evidence: Verified claim, comparison rule, workload volume, total-impact calculation, rebound or demand scenario, non-GHG limits, provenance graph, checked date, expiry and named update trigger.
Stop condition: Stop when lower intensity is advertised as lower total impact despite demand growth, unlike systems are ranked without shared product rules, or an old result survives hardware, software, grid or workload change.

Minimum energy-and-carbon claim failure matrix

Check these conditions deliberately to expose boundary, measurement, allocation and total-impact errors behind low-energy or low-carbon numbers.

01
A per-query number omits token count, output quality, latency target or failed requests
Record affected workloads, meters, boundaries, factors, allocations and decisions; preserve the narrowest statement that remains and specify whether to remeasure, recalculate, change factors, narrow comparison or withdraw the conclusion.
02
GPU TDP or nameplate power is presented as measured workload energy
Record affected workloads, meters, boundaries, factors, allocations and decisions; preserve the narrowest statement that remains and specify whether to remeasure, recalculate, change factors, narrow comparison or withdraw the conclusion.
03
IT energy excludes cooling, power conversion, network, storage or reserved idle capacity without disclosure
Record affected workloads, meters, boundaries, factors, allocations and decisions; preserve the narrowest statement that remains and specify whether to remeasure, recalculate, change factors, narrow comparison or withdraw the conclusion.
04
An annual national grid factor is applied to a time- and location-specific run
Record affected workloads, meters, boundaries, factors, allocations and decisions; preserve the narrowest statement that remains and specify whether to remeasure, recalculate, change factors, narrow comparison or withdraw the conclusion.
05
Location-based and market-based results are switched to show the lower number
Record affected workloads, meters, boundaries, factors, allocations and decisions; preserve the narrowest statement that remains and specify whether to remeasure, recalculate, change factors, narrow comparison or withdraw the conclusion.
06
Hardware manufacturing emissions disappear because supplier or allocation data is unavailable
Record affected workloads, meters, boundaries, factors, allocations and decisions; preserve the narrowest statement that remains and specify whether to remeasure, recalculate, change factors, narrow comparison or withdraw the conclusion.
07
Two systems are compared across different hardware, batching, utilisation or service quality
Record affected workloads, meters, boundaries, factors, allocations and decisions; preserve the narrowest statement that remains and specify whether to remeasure, recalculate, change factors, narrow comparison or withdraw the conclusion.
08
A lower per-unit intensity hides rising traffic, longer outputs or higher total electricity demand
Record affected workloads, meters, boundaries, factors, allocations and decisions; preserve the narrowest statement that remains and specify whether to remeasure, recalculate, change factors, narrow comparison or withdraw the conclusion.

Minimum energy-and-carbon claim evaluation record

Let the next reviewer rebuild intensity and total impact from the same workload and operating state, and understand when comparison must stop.

01verbatim claim, claimant, date, intended decision and evidence state
02model and service version, workload, input and output profile, latency and quality target
03functional unit, workload volume, comparison rule and included or excluded system boundary
04meter location, identifier, calibration, sampling, raw readings and timestamps
05hardware, software, utilisation, batching, repetitions, failures and reserved capacity
06facility category, PUE inputs, cooling, power, network and storage treatment
07emission-factor publisher, version, geography, time, GWP basis and accounting method
08hardware life-cycle source, expected lifetime, time share, resource share and allocation
09calculation code, uncertainty, sensitivity, alternative factors and quality results
10verified intensity and total, non-GHG limits, provenance, expiry and update trigger

Common evidence states

Bind conclusion status to workload, functional unit, boundary, measurement, accounting method and date—not green labels or one efficiency metric.

Supported

The workload, functional unit, boundary, energy measurement, factors, hardware allocation, uncertainty and dated interpretation are independently reconstructable.

Conditional

The calculation is reviewable but supports only the named model, workload, quality target, facility, geography, method and period.

Mixed

Reasonable boundaries, allocation choices, electricity methods or demand scenarios produce materially different intensity or total-impact conclusions.

Insufficient

Material workload, meter, boundary, factor, allocation or quality evidence is missing; PUE, TDP, offsets or polished dashboards do not repair the chain.

FUURAA analysisThe minimum decision unit for an AI energy or carbon claim is workload and quality target × functional unit × system boundary × measured energy × accounting method and factors × hardware, software, place and time × cut-off date. Lower per-task intensity can coexist with higher total demand; PUE, TDP, renewable contracts or offsets cannot answer total impact alone. FUURAA recommends preserving intensity and totals together and making every comparison reproducible, rejectable and expiring.

Primary sources and non-transfer boundaries

These methods constrain energy-system analysis, software carbon intensity, data-centre efficiency, product life cycle, purchased-energy accounting and provenance; none independently proves an AI energy or carbon claim correct.

Sources rechecked 24 August 2026. Each retains its publication timing, role in this method and non-transfer boundary.

Published 16 April 2026IEA — Key Questions on Energy and AI

Provides current system-level evidence on data-centre electricity demand, AI efficiency, infrastructure constraints, rebound effects and uncertainty across scenarios.

BoundarySector scenarios do not measure one model, workload, facility or supplier and cannot be used as a per-request carbon factor.

Open primary source ↗Version 1.1 released October 2024; ISO edition 1 published March 2024Green Software Foundation — Software Carbon Intensity Specification v1.1.0

Defines a rate-based method for software carbon intensity using a software boundary, functional unit, operational energy, regional carbon intensity and embodied hardware emissions.

BoundaryAn SCI score depends on declared boundaries and functional units; it is not total organisational emissions, a carbon-neutrality claim or proof that unlike services are comparable.

Open primary source ↗Current documentation for Portfolio Manager 26.0, released 24 February 2026U.S. EPA ENERGY STAR Portfolio Manager — Data Center IT Energy and PUE

Defines permitted IT-energy meter locations, recommends monthly readings and explains data-centre PUE as annual total source energy divided by annual IT source energy.

BoundaryPUE describes facility overhead relative to IT energy; it does not measure workload quality, server efficiency, embodied emissions, electricity carbon intensity or total GHG impact.

Open primary source ↗2011 standard; official materials dated October 2011GHG Protocol — Product Life Cycle Accounting and Reporting Standard

Provides requirements for goals, inventory boundaries, life-cycle stages, allocation, data quality, uncertainty, calculation, assurance and reporting for product-level GHG inventories.

BoundaryThe standard does not quantify avoided emissions or carbon neutrality, and extra product rules are needed for comparative claims; a GHG inventory is not overall environmental superiority.

Open primary source ↗Published January 2015; consultation on revision closed 31 January 2026GHG Protocol — Scope 2 Guidance

Standardises corporate accounting for purchased electricity, steam, heat and cooling, including location-based and market-based methods, contractual instruments and disclosure.

BoundaryScope 2 corporate inventory methods do not allocate emissions to one AI task or validate a supplier claim; location- and market-based results answer different accounting questions.

Open primary source ↗W3C Recommendation, 30 April 2013W3C PROV-DM — The PROV Data Model

Defines entities, activities, agents, derivations and responsibility links for tracing workload definitions, meter data, factors, allocation, calculations, revisions and publication.

BoundaryProvenance can expose how a result was produced, but it does not prove that meters, emission factors, allocation choices or conclusions are correct.

Open primary source ↗

Continue checking

Move from an energy or carbon claim back to quantitative verification, complete syntheses and evidence records.

Verify quantitative claims Review research syntheses Build an evidence record Enter AI Evidence Atlas