Classifies the activity
Text, images, documents, video and audio are treated separately because their compute profiles can differ materially.
AImpronta Model v1.0
The engine behind AImpronta
A proprietary, versioned method whose evidence trail can be checked: how AI activity becomes transparent estimates of energy, water and CO₂e.
AImpronta Model v1.0 · Method 2.0.0 · Source review 23 September 2026
STATUSOperational and versionedAImpronta Model v1.0 · Method 2.0.0
REVIEW23 September 2026Sources and boundaries re-checked
EVIDENCEProviders + research + institutionsEach category has a declared technical basis
TRACEABILITYVersion tied to the resultSame structured input + same version = same output
AImpronta Model v1.0
A proprietary model built on verifiable evidence
AImpronta does not apply one number to every AI activity. The engine selects a technical basis appropriate to the type of use, separates provider disclosures, research benchmarks and engineering scenarios, normalises different measurement boundaries, and returns versioned estimates of energy, water and CO₂e.
The methodology is public in its principles, sources, boundaries and limitations. The implementation, processing coefficients and parameterisation logic remain proprietary parts of AImpronta Model.
AIMPRONTA LAB
How AImpronta Model works
An illustrated view of the model: data, benchmarks and usage parameters enter the engine and become comparable estimates of water, energy and CO₂e.
AIMPRONTA / MODEL v1.0
01 / YOUR AI USAGE02 / OUR MODEL03 / ESTIMATED IMPACT
MODEL v1.0
AImprontaFROM USAGE TO IMPACT01 Classify02 Normalise03 Estimate
ESTIMATEEnergykWh
ESTIMATEWaterL
ESTIMATECO₂eg CO₂e
DECLARED PARAMETERSActivity · provider · quantity · duration
SOURCES + BENCHMARKSPublic evidence and transparent assumptions
A READABLE RESULTThree indicators, with limits and ranges
WHY THE MODEL IS ROBUST
Not one formula, but an evidence chain
Selects the evidence
When a public provider measurement exists it is considered for that scope. Otherwise the model uses independent research or declared technical benchmarks.
Normalises boundaries
The engine distinguishes IT energy, data-centre overhead, direct water and electricity factors instead of combining quantities with different scopes.
Parameterises
Volume, duration, activity type and, when identifiable, provider change the reference used. Detailed coefficients remain in the server-side engine.
Shows result and uncertainty
The central reference is accompanied by sensitivity, equivalents and sources. Sensitivities are not labelled statistical confidence intervals when they are not.
EVIDENCE UPDATED TO 2026
The references that matter today
AImpronta Model is maintained against the best public evidence available. The current review incorporates institutional sources, peer-reviewed studies and provider technical disclosures available through 23 September 2026.
Electricity demand from AI-focused data centres grew 50% in 2025; total data-centre electricity demand grew 17%.
IEA source ↗Google estimates a median Gemini Apps text prompt at 0.24 Wh, 0.03 gCO₂e and 0.26 ml of water using a documented operational methodology.
Google methodology ↗Peer-reviewed research by de Vries-Gao estimates AI’s 2025 carbon footprint at 32.6–79.7 Mt CO₂ and water footprint at 312.5–764.6 billion litres.
Peer-reviewed study ↗On 21 September 2026 the Commission proposed a common data-centre rating scheme centred in part on transparency of energy and water use.
European Commission ↗These figures have different boundaries and are not automatically combined. AImpronta Model uses each source only in the context for which it is technically relevant.
SOURCE HIERARCHY
How we choose the technical basis
Provider measurement or disclosure
First choice when the data is public, the product is identifiable and the measurement boundary is sufficiently described.
Independent or peer-reviewed research
Used when provider measurements are unavailable or an external category benchmark is required.
Engineering scenario
Used only when no suitable direct measure exists. It is labelled as a scenario, not as a measurement.
Institutional conversion
IEA and other authoritative sources support contextual factors such as electricity mix while measurement boundaries remain separate.
TECHNICAL BASIS BY CATEGORY
Different activities use different evidence
We do not publish the engine’s internal coefficients. We do make the evidence trail and its treatment externally verifiable.
| Activity | Main basis | Evidence type | How it is used |
|---|---|---|---|
| Text | Google Gemini, OpenAI, Epoch AI | Provider disclosure + independent analysis | If the provider can be identified, a compatible provider-specific reference is used; otherwise the generic text reference applies. |
| Images | Luccioni et al., ACM FAccT | Independent research | Inference benchmark with declared hardware boundaries, used as the technical category reference. |
| Video | Delavande et al. | Technical benchmark | Duration enters the parameterisation; the benchmark is not presented as a measurement of Sora, Veo or other proprietary services. |
| Audio | Passoni et al. + NVIDIA | Research + hardware scenario | An engineering scenario explicitly distinguished from a direct measurement of commercial services. |
| Documents | Text/compute reference | Parametric scenario | The user sentence cannot reveal actual format, length and operation, so the result remains a scenario estimate. |
MODEL VERIFICATION
What “verified” means at AImpronta
Sources re-checked
Public references used by the current version were reviewed again as part of the 23 September 2026 release review.
Boundaries declared
The model separates provider disclosures, research, engineering scenarios, IT energy, overhead, water and electricity factors.
Versioned outputs
The engine returns the model version and links results to relevant source IDs, keeping updates traceable.
Deterministic consistency
Given the same structured input and model version, the engine returns the same output. Sensitive formulas execute server-side.
Important: this is an internal verification of source traceability and model consistency. AImpronta Model is not currently an environmental certification or a methodology independently validated by a third-party body.
Scope, data quality and limitations
The calculator estimates operational energy, water using the applicable reference, and electricity-related CO₂. It does not represent the full AI life cycle and does not read energy consumption from an individual account or data centre.
Model, context length, hardware, load, location and infrastructure can materially change consumption. A public reference does not measure every possible use.
Training, hardware manufacturing, data-centre construction, networks and user devices are excluded unless explicitly stated. The Global Counter has a separate boundary described below.
Calculator sensitivities explore variation around the central AImpronta reference. They are not statistical confidence intervals or provider-published ranges.
Equivalents and projections
Equivalents translate results into everyday examples such as bottles, LED hours or driving distance. They explain scale and do not imply that those resources were physically consumed in that form.
Collective and business projections retain the same reference scenario and do not automatically model economies of scale, infrastructure changes or specific energy contracts.
Main sources and documentation
Sources are exposed so the method can be checked externally without publishing the engine’s proprietary implementation.
| Source | Year | Role in the method | Boundary |
|---|---|---|---|
| IEA · Key Questions on Energy and AI ↗ IEA · Electricity 2026 ↗ | 2026 | Energy scenario and global context | Data centres, AI-focused data centres, electricity demand and global mix. Not a single-prompt measurement. |
| Google · Measuring AI inference at Google Scale ↗ Technical paper ↗ | 2025 | Direct provider disclosure | Median Gemini Apps text prompts with operational overhead. Not every prompt or provider. |
| de Vries-Gao · Patterns ↗ | 2026 | Global water/carbon cross-check | AI/data-centre-scale estimates; useful for global orders of magnitude, not single prompts. |
| Epoch AI · How much energy does ChatGPT use? ↗ | 2025 | Independent text reference | Engineering analysis, not a measured average across all models. |
| Sam Altman · The Gentle Singularity ↗ | 2025 | Public OpenAI disclosure | Reference for ChatGPT text requests; measurement boundaries are not documented to the same level as Google’s study. |
| Luccioni et al. · Power Hungry Processing ↗ | 2024 | Image benchmark | Research with defined hardware and tasks; not every proprietary image generator. |
| Delavande et al. · Text-to-video energy ↗ | 2025 | Video benchmark | Specific model and hardware; duration parameterisation is an AImpronta scenario. |
| Passoni et al. · Generative audio ↗ | 2025 | Audio benchmark | Research combined with hardware specifications; not a direct commercial-service measurement. |
| European Commission · Energy performance of data centres ↗ | 2026 | Transparency and reporting context | Data-centre energy and water rating/reporting. Not an AImpronta conversion factor. |
| EU Regulation 2024/1689 · AI Act ↗ | 2024 | Regulatory context | AI transparency and governance; not used to calculate consumption. |
EUROPEAN CONTEXT 2026
Data-centre transparency is becoming market infrastructure
On 21 September 2026 the European Commission proposed a common rating scheme for data centres to increase transparency around energy and water, alongside reporting under the Energy Efficiency Directive. AImpronta is not a legal compliance system, but it follows the same direction: making fragmented environmental data easier to understand and compare.
AImpronta Global Counter · v1.0.0
The model behind the global counter
Source review: 23 September 2026
The Global Counter is an AImpronta time-based estimate, not worldwide real-time telemetry.
Energy
Starts from IEA outlooks for AI-focused data centres and builds an annual scenario consistent with public 2025–2030 benchmarks.
Water
Combines direct cooling references and electricity-related water. Water consumption and withdrawal are not the same measure.
CO₂e
Uses a declared world-electricity-mix factor; it does not represent the actual contracts or generation mix of every AI facility.
Live
The number advances with time based on the annual scenario. It does not imply a worldwide sensor measuring AI second by second.
The IEA reports that electricity consumption by AI-focused data centres grew 50% in 2025. Because comprehensive statistics on worldwide AI usage do not exist, the counter should be read as a transparent, updateable scenario rather than observed telemetry.
Global Counter sources
MODELAImpronta Model v1.0
METHOD2.0.0
REVIEW23/09/2026
STATUSPublic-source estimate
© AImpronta 2026. AImpronta Model and its implementation are part of the AImpronta project. Public data and cited studies remain attributed to their original sources.