The AI Environmental Ledger

← Back to Insights

Executive summary

Enterprises can now account for the energy, carbon and water of their AI use at the level of the individual agent action, but only if every figure carries its evidence, its boundary and its uncertainty. This paper sets out a method that does exactly that, and shows how it runs as one ledger shared between an AI governance platform and an ESG accounting platform.

The evidence has matured quickly. For a short text prompt to a large production model, the best 2025 and 2026 measurements and estimates agree on 0.24 to 0.43 Wh per query. Beyond that simple case the spread widens by two orders of magnitude or more, because output length, reasoning, context length, serving conditions and the measurement boundary each move the answer by multiples. Reasoning alone raises energy per request by roughly 13 to 30 times on average. No provider yet publishes energy per input and per output token for a closed frontier model at a full facility boundary, and that single gap accounts for most of the uncertainty in any enterprise estimate.

The AI Environmental Ledger answers this with five design choices. It meters fresh, cached, output and reasoning tokens separately from each model response. It draws every factor from a five-tier evidence hierarchy and gives every result an A to E data-quality grade and a 95% range. It reports location-based carbon as the headline, with market-based carbon beside it and a flag for the quality of any renewable claim. It counts water consumed on site and in electricity generation, with scarcity weighting as a second metric. It converts each physical quantity to currency at a declared price before anything is added, which replaces the common but meaningless practice of adding kilowatt-hours, kilograms and litres.

The worked examples make the stakes concrete. A trillion short-chat tokens served for an enterprise on the Singapore grid use about 520 MWh (95% range 133 to 2,040 MWh) and emit about 209 tCO2 on a location basis. That range is roughly fourfold either side of the central value, and it does not narrow as volume grows, because the uncertainty sits in shared factors rather than in individual calls. At a social carbon price of US$190 per tonne (the US EPA's 2023 estimate), the external carbon and water cost adds about 84% to the electricity bill; at a typical corporate internal price it adds about 10%. Looking forward, valuing each year's emissions at the rising social cost of carbon and adding the air-pollution damage of the electricity raises today's valuation by about half, and a trillion-token portfolio carries a central present value of about US$6.5 million to 2050, of which carbon prices would recover less than half.

Per-token carbon estimation is no longer new, and this paper builds openly on the frameworks and products that already provide it. What is distinctive, to our knowledge, is that the ledger is designed to sit inside AI governance rather than beside it. A board can see the environmental cost of an agent, a business process or a successful outcome, not only of an abstract prompt, and can treat that cost as a governance control graded for evidence like any other. Three narrower contributions accompany it: a single five-class token equation, a grade and propagated range on every figure, and a FinOps-keyed environmental cost priced unit by unit.

Leaders should act on three fronts now. Instrument AI usage at the gateway so token data by type exists before disclosure rules ask for it. Write per-token energy, facility efficiency and renewable instrument quality into AI procurement contracts. Pilot the ledger on one governed business process, reconcile it against provider invoices and carbon reports, and test the evidence file with an assurer before limited assurance becomes mandatory. The companion paper, Net AI Impact, extends the same method to workforce, customer and fairness effects, and tests whether a claim that AI made a company more efficient, greener or fairer can be defended.

This is the public edition. The full whitepaper, with all 18 formulas, worked examples, exhibits, the evidence hierarchy and the complete source list, is available as a PDF. Download the full whitepaper (PDF), or read the sections below.

01Section 1Why AI's footprint is now a board question

AI has moved from an IT line item to a material and fast-growing energy load, and the companies that buy it cannot yet see what they are buying. The International Energy Agency puts data-centre electricity at about 415 TWh in 2024, roughly 1.5% of global demand, rising to about 945 TWh by 2030 (IEA). The hyperscalers' own reports show the pressure: Google's total electricity use rose 27% in 2024 and 37% in 2025, and Microsoft's total emissions rose 25% in its last financial year, to 20.3 million tonnes, even as it contracts record volumes of clean power (Google 2026 Environmental Report; DCD).

For the enterprise buyer, three forces turn this into a board matter rather than a technical curiosity. The first is disclosure. AI consumed through an API sits in Scope 3 Category 1, purchased goods and services, and Scope 3 reporting is arriving in the markets where CorpStage's clients operate: Singapore requires it of Straits Times Index constituents from FY2026, and the revised ESRS, adopted by the European Commission in July 2026, are due to apply to the largest EU groups from FY2027 once in force (ACRA; EFRAG). The second is claims risk. Companies that promote AI-enabled efficiency while ignoring the footprint of the AI itself invite the same scrutiny that greenwashing enforcement now applies to any unsupported environmental claim. The third is local impact: a Bloomberg News analysis found that about two-thirds of US data centres built since 2022 sit in water-stressed regions, which makes water a community and licence-to-operate issue, not only a metric (ELI).

The measurement gap is wide. Capgemini found that only 12% of executives at large organisations using generative AI say their organisation measures its footprint, and nearly three-quarters find it hard to measure because providers disclose too little (Capgemini 2025). On one large routing platform, the models carrying 84% of usage disclosed nothing about their energy (Luccioni et al. 2025). Regulation has not closed the gap for buyers: the EU AI Act requires general-purpose model providers to document energy for regulators, but gives downstream deployers no right to that data.

A sense of proportion matters here. For most enterprises, AI inference will be a small share of total emissions today, and for a bank it is dwarfed by financed emissions. The case for a ledger does not rest on size alone. It rests on growth rates, on the credibility of every AI-related sustainability claim a company makes, and on the fact that agentic AI is turning token consumption into an operating cost that multiplies without anyone deciding it should. Governing that consumption requires measuring it at the level where decisions are made: the agent, the process and the outcome.

02Section 2What the evidence says

The literature now agrees on the simple case and disagrees, for good reasons, on everything else. Our review covered more than 250 sources, from provider telemetry and peer-reviewed measurement studies to standards texts and vendor documentation, published mainly between 2023 and September 2026.

The boundary is the first source of disagreement. Google's production study is the most complete disclosure to date: its median Gemini text prompt uses 0.24 Wh, of which active accelerators account for 58%, host CPU and memory 25%, idle reserved machines 10% and data-centre overhead 8% (Elsworth et al. 2025). The narrow method common in published estimates, which counts active chips only and samples the most efficient sites, gives 0.10 Wh, so it understates the full figure by 2.4 times. Many published estimates and open tools draw a similarly narrow boundary, which is why the ledger treats the boundary as a mandatory attribute of every number.

Output tokens, not input tokens, carry most of the energy. Generating each output token is sequential and limited by memory bandwidth, while processing the prompt runs in parallel. Every study that separates the two finds output dominant; one fitted model explains more than 96% of the variation in per-request energy from input length, output length and their interaction (Wilkins et al. 2024). Long contexts still matter at scale, because attention cost grows faster than linearly with context length.

Reasoning is the largest single multiplier. Hugging Face's AI Energy Score found reasoning models use about 30 times more energy on average, and 150 to 700 times more when reasoning is switched on for the same model (Luccioni and Gamazaychikov 2025). ML.ENERGY measured an average of 4,625 J for a problem-solving response against 184 J for a conversational one, about 25 times more (ML.ENERGY v3.0). Reasoning-token counts also varied by up to 9.7 times across repeated runs of identical queries in one benchmark study, so they must be metered from actual usage rather than inferred from the task (Chen et al. 2026).

Serving conditions matter as much as the model. Batch size alone moves energy per token by three to five times, an optimised serving stack cut energy by up to 73% against a naive baseline, and the same open reasoning model used about four times as much energy for long prompts on its developer's own servers as on Microsoft Azure (ML.ENERGY v3.0; Fernandez et al. 2025; Jegham et al. 2025). Agentic workflows compound all of this: charging every retry and failure to the goal, one study found agentic designs used 4.33 times more energy per successful goal than matched linear baselines (Panigrahy and Tyagi 2026).

Water figures differ by more than a hundredfold because they measure different things. Google's 0.26 mL per prompt counts only on-site cooling water; Mistral's 45 mL covers a 400-token response, assessed within a full life-cycle analysis (Mistral 2025). The water consumed off site, in generating the electricity, is often the larger share: with US average factors it is about 87% of operational water, on our reading of their figures (Li et al.). Adding it to Google's disclosure raises the median Gemini prompt's water by 179% (Sharma et al. 2026).

Location and accounting method move results more than most efficiency measures. Grid emission factors across CorpStage's markets range from about 0.40 kgCO2 per kWh in Singapore to about 0.71 in India, and the global data-centre PUE has held near 1.54 for six years while hyperscalers report 1.10 to 1.15 (EMA; CEA; Uptime via mGrid). Market-based and location-based figures can differ by an order of magnitude: a Guardian analysis found that Apple, Google, Microsoft and Meta's location-based data-centre emissions for 2020 to 2022 were almost eight times their market-based figures (TechInformed, secondary).

Uncertainty is the field's weakest point. Published figures for inference span four to five orders of magnitude across models, and 75% of media articles relayed an energy estimate for a ChatGPT query without mentioning any uncertainty (Luccioni et al. 2025). The statistical tools to fix this are mature and sit in IPCC, GHG Protocol and EU life-cycle guidance. They have simply not been applied systematically to AI.

03Section 3Where current approaches stop

Per-token carbon estimation now exists in several forms, but none of them connects the footprint to the governed decision that caused it. The landscape falls into three groups.

Cloud providers report by service, not by model or token. AWS, Google Cloud and Microsoft Azure all offer carbon reporting to customers (AWS moved its reporting to a standalone Sustainability console in March 2026), but the finest grain is account, project or resource by service and region. None reports emissions per model, per token or per request for their managed AI services (AWS; Google Cloud; Azure). Google states that its customer-report data are not assured, and allocates some emissions by list price, which ties the carbon figure partly to cost.

Open frameworks and specialist products now estimate per token and per request. This paper builds on and cites the closest prior work, including Watershed's open framework and the paper by Bistline et al. (arXiv 2608.06733), Llopis's four-tier Scope 3 Category 1 method (arXiv 2606.10660), the Green Software Foundation's SCI for AI, EcoLogits, Sharma et al. on water, Lu's GreenPassport (arXiv 2609.06784) and Greenpixie. Other services, among them Scope3, Antarctica, Neuralwatt and GreenGate, also report per inference or per team.

What remains missing is the governance link. Our prior-art check covered the major carbon-accounting vendors, AI governance platforms, LLM observability and gateway tools, the primary texts of the closest frameworks, indexed patents and the literature to 25 September 2026. Governance platforms record what agents did and under which policy, but not what it cost the environment. Carbon platforms hold factors and disclosures, but not which agent action consumed which tokens. Observability tools can tag energy estimates by agent, but their figures are ungraded, and they neither feed a disclosure nor return to a control framework as evidence. We found no product or paper that closes this gap. Three further gaps run across the field: only one commercial product separates cached tokens, and none combines cached, reasoning and context terms in one equation; no method gives every figure both a data-quality grade and a propagated uncertainty range; and no method joins AI footprints to the FinOps FOCUS billing standard or prices each environmental quantity before summing.

04Section 4The AI Environmental Ledger: design principles

The ledger treats every environmental figure as a distribution with a pedigree, attached to a governed business activity, rather than as a point estimate attached to a prompt. Six principles follow from the evidence.

Meter, do not infer. Token counts by type come from the usage fields each model response already returns: fresh input, cached input, cache writes, visible output and reasoning. Reasoning volumes vary too much to estimate from task type, and cached tokens cost a fraction of fresh ones, so the ledger reads them rather than assuming them.

Make the boundary explicit. Every figure states what it includes: accelerators, host, idle capacity, facility overhead, embodied hardware, training, on-site water, off-site water. Figures computed at different boundaries are never mixed without a conversion factor that carries its own uncertainty.

Grade the evidence and carry the range. Each factor comes from the best available tier of evidence, from metered data down to spend-based estimates, and each result reports a central value, a 95% range and an A to E grade. The grade reflects the factors that actually account for the uncertainty.

Report physics first and claims second. Location-based carbon is the headline, because it describes the grid the electricity actually came from. Market-based carbon sits beside it, with a flag for the quality of the renewable instrument. Embodied hardware and training emissions are reported on separate lines, never buried in the total.

Price each unit before adding anything. Energy, carbon and water are converted to currency at declared prices before they are summed. Internal energy cost is reported apart from the external carbon and water cost, and the price vector used is disclosed with every result.

Restate, never overwrite. A new factor or corrected activity record creates a new ledger entry that supersedes the old one, so the history an assurer needs is always intact. The full system boundary, with what is reported in the mandatory total, what is reported separately, and what is declared out of scope, is set out in the PDF.

05Section 5How the ledger calculates

The method runs in five steps, from the tokens a model call consumed to an environmental cost in currency, and every step carries its uncertainty forward. The full specification, with 18 numbered formulas, sits in the PDF; this section gives the logic a CFO or sustainability lead needs to follow it.

Step 1: energy per model call. Each call's active energy is the sum of what each class of token costs, plus a term for long contexts. In the default proxy set, an output token costs about four times an input token, a cached read about a tenth of a fresh one, and a reasoning token about 1.8 times an ordinary output token.

Step 2: from chip to facility. Active energy is raised to the facility boundary by explicit uplifts for host servers, idle and reserve capacity, and data-centre overhead through PUE. Where only a chip-level benchmark exists, a single measured conversion factor of about 2.4 replaces the three uplifts, carrying its own uncertainty. Calls then roll up to the governed action, the agent, the business process and the successful goal, with failed attempts and retries charged to the goal they served.

Step 3: carbon, twice. Location-based carbon multiplies energy by the average emission factor of the grid zone, hourly where available and annual otherwise. Market-based carbon applies contractual instruments by their quality class, and a high-integrity view credits only hourly-matched, deliverable supply.

Step 4: water, on site and off site. On-site cooling water follows the facility's water usage effectiveness; off-site water follows the water intensity of the electricity consumed. Both are counted as consumption, not withdrawal, and a scarcity-weighted figure using WULCA's AWARE factors is reported alongside.

Step 5: environmental cost in currency. Each quantity is priced before anything is added, with internal energy cost kept apart from the external shadow cost of carbon and water. For API use, the energy term is already inside the invoice, so the ledger reports it for information and never adds it to spend.

Uncertainty runs through every step. Each factor carries a log-normal distribution whose width reflects its evidence tier and its data-quality scores. Period totals are computed by Monte Carlo simulation in which shared factors, such as a model's energy coefficients, the facility PUE and the grid factor, are drawn once for the whole period rather than once per call. That is why the relative range does not shrink as volumes grow: a ledger that reported tighter ranges for larger volumes would be making a statistical error. The evidence hierarchy, from Tier 1a metered data to Tier 4 spend-based estimates with its typical ranges, and all 18 formulas, are set out in full in the PDF.

06Section 6What the numbers look like

A trillion tokens of enterprise AI use a few hundred megawatt-hours and emit tens to hundreds of tonnes of CO2, depending far more on the grid than on the workload, and every figure carries a range of about four times either side. The examples below were computed by Monte Carlo simulation with 200,000 draws, using the Tier 3 proxy factors that most enterprises will start from today.

We modelled three workload profiles. Short chat sends 400 input and 300 output tokens per request. Long-context retrieval sends 10,000 input tokens, of which 2,000 are cached instructions, and returns 1,500 output tokens. Reasoning agent makes eight model calls per goal, each with 6,000 input tokens (4,000 cached from earlier turns), 500 visible output and 2,500 reasoning tokens, and succeeds on 75% of goals. Each ran on three grids: Singapore (0.402 kgCO2 per kWh), US East (0.282 kgCO2e per kWh) and an illustrative low-carbon grid (0.057 kgCO2e per kWh).

Results for one trillion tokens

Central values with 95% ranges in brackets. Carbon is location-based; water is consumption on site plus off site. Monetised cost prices electricity at a placeholder US$0.10 per kWh, carbon at US$190 per tonne and water at US$6 per cubic metre. Results scale linearly: divide by one million for a million tokens and by one thousand for a billion.

Selected rows. The full nine-row table, with long-context and reasoning-agent results on all three grids, is in the PDF.

Five findings for implementation

The range does not narrow with scale. The ratio of the upper to the lower bound of energy is 15.3 at a million, a billion and a trillion tokens alike. The uncertainty lives in shared factors.

One missing number accounts for most of the uncertainty. Energy per output token explains about 94% of the variance in the short-chat result. The single most valuable request an enterprise can make of an AI provider is a model-specific energy per output token at a stated boundary.

Per-token intensity depends on workload shape. Long-context retrieval runs at 0.30 Wh per thousand tokens against 0.52 for short chat, while the reasoning agent reaches 54 Wh per successful goal. A single blended per-token factor would misstate all three.

Carbon and water can move in opposite directions. The low-carbon grid cuts short-chat carbon about seven times against Singapore but, under a nuclear cooling assumption, consumes about 2.9 times more water. In the US case, off-site electricity water is 91% of the total.

The carbon price sets the story. At US$190 per tonne, the external shadow cost equals 84% of the declared electricity cost in Singapore; at a US$25 corporate internal price, carbon adds about 10%. The price vector must therefore be declared and disclosed, never left implicit. The forward valuation to 2050, including a trillion-token portfolio's central present value of about US$6.5 million, is set out in the PDF.

07Section 7Where the ledger lands in reporting and assurance

No disclosure framework yet asks for AI by name; AI enters through energy, Scope 2, Scope 3 Category 1 and water, and the ledger maps onto those datapoints directly. The deployment mode decides the treatment. AI consumed through an API is a purchased service, reported in Scope 3 Category 1 and in ISO 14064-1 Category 4. Self-hosted inference puts the electricity in Scope 2 and the hardware in Scope 3 Category 2.

The ledger maps, with the strength of each link stated honestly as direct, indirect or none, onto the GHG Protocol (Scope 2 and Scope 3 Category 1, and the current revisions), IFRS S2, ESRS E1 and E3, Singapore's ACRA and SGX regime, the EU AI Act and the EU data-centre rating scheme, Germany's Energy Efficiency Act, ISO/IEC 21031 and the SCI for AI, and ISSA 5000. The full mapping table is in the PDF.

Assurance readiness rests on three properties. The ledger exposes, for every figure, the method, assumptions, data sources and change history that ISSA 5000 directs assurers to examine. It separates data the company generates itself, the token counts captured at the point of governance, from external information such as provider factors, which is exactly the distinction assurers test. It quantifies uncertainty in the way ISO 14064-1 expects.

The honest limit is that estimate-heavy Category 1 figures can reach only limited assurance until providers supply attested per-token factors. Singapore's deferral of limited assurance of Scope 1 and 2 to FY2029 and the EU's decision to keep assurance at the limited level give companies time to pilot the evidence file with an assurer before it becomes mandatory.

08Section 8From ledger to governance: one record across two platforms

The ledger splits responsibility along the line between what happened and what it means. CorpStage's AI governance platform, AIVARA, records what happened: which agent took which action, under which policy, consuming which tokens. CorpStage's ESG accounting platform, ESG 360, determines what it means: the factors, the calculation, the scope classification and the disclosure. Both work from one shared record, so the activity a governance team approves and the figure a sustainability team discloses can never drift apart.

The governance layer meters. When an agent takes a consequential action, the governance gateway reads the usage block of every model response: fresh input, cache-read, cache-write, output and reasoning tokens. It records the model, provider, region and timestamps, and attributes each call to the governed action, the agent, the accountable person and the business process, at the point of governance and before anyone has an incentive to adjust it.

The accounting layer calculates and discloses. ESG 360 selects factors by evidence tier, computes the distributions, classifies scope and category by deployment mode, and keeps every factor, method version and restatement on record so an assurer can rebuild any figure. Each month it reconciles the bottom-up totals against provider invoices and carbon reports, and flags any variance outside the propagated range.

The result becomes evidence. The environmental figure returns to the governance layer, where it is graded like any other control: well evidenced where activity is fully captured and factors come from measured or hardware-level data, estimated where factors are proxies, and weak where figures are spend-based or reasoning is unmetered. The footprint of an agent becomes a governed property of that agent.

This design gives a board three things that, to our knowledge, no other approach offers today. It can see the environmental cost of an agent, a business process or a successful outcome, not only of an abstract prompt. It can trace that figure to the same governance record that shows the agent acted within policy. It can treat environmental performance as a governance control, graded for evidence like any other. The ledger is also the environmental pillar of CorpStage's companion paper, Net AI Impact.

09Section 9What leaders should do now

The first year is about getting honest data flowing, not about precision: instrument usage, report ranges, then move the largest uncertainties up the evidence hierarchy through procurement. The sequence runs in four phases, each with a gate to the next, and lets a company start disclosing defensible, graded figures within six months and reach assurance readiness inside two years.

Phase 1, instrument (months 0 to 3). Route model traffic through a governed gateway; capture tokens by class, model, provider and region; inventory every AI provider and deployment mode; produce a spend-based baseline. Gate: token data by class covers at least 90% of AI spend.

Phase 2, calculate (months 3 to 6). Apply proxy factors with ranges and grades; report location-based and market-based carbon; join results to billing data in FOCUS format; declare the carbon and water price vector. Gate: first graded quarterly figures reviewed by finance and sustainability.

Phase 3, improve the evidence (months 6 to 12). Add disclosure clauses to AI contracts; move the largest factors to provider or hardware tiers; meter self-hosted inference at node and facility level. Gate: energy per output token at Tier 1b or 2 for the top three models by volume.

Phase 4, assure and govern (months 12 to 24). Pilot the evidence file with an assurer under ISSA 5000; include AI in Scope 3 Category 1 disclosure; set intensity targets per successful goal; add footprint to model routing and agent approval policies.

Procurement is the fastest lever. Because energy per output token accounts for most of the uncertainty, a handful of contract clauses does more for accuracy than any modelling refinement. Request, for each model used: energy per thousand input and output tokens at a stated boundary; facility PUE and consumptive WUE by serving region; the quality class of any renewable instrument claimed, preferably hourly-matched; and notice of material changes in model, hardware or region.

Start with one governed process. A single agentic process, such as invoice handling or client onboarding, is enough to run the ledger end to end: governed actions, activity records, calculated footprints, reconciliation against the provider's invoice and carbon report, and energy per successful outcome. That pilot produces the evidence file an assurer will ask for and the internal benchmark every later process can be measured against.

10Section 10Contribution, limitations and research agenda

Per-token and per-request estimation of the energy, carbon and water of AI use is now established, and we build on that work rather than claiming it as new. To our knowledge, no previously published method or product makes these four contributions together: attribution of each figure inside governance, returned as graded evidence; a single five-class token equation covering fresh, cached-read, cache-write, output and reasoning tokens with a context term; a data-quality grade and a propagated Monte Carlo range on every figure; and a unit-correct, FinOps-keyed environmental cost that prices each quantity before summing.

Limitations. The method is only as good as its weakest factor, and today that factor is estimated. No provider publishes energy per input and per output token for a closed frontier model at a full facility boundary, so most enterprise figures will sit in the proxy tier with ranges of about four times either side. The energy saved by a cache hit rests on the ratio of provider prices, not on measurement. The idle share is measured only for Google. Several water scarcity factors and a defensible rule for hydro-reservoir evaporation are still open. No per-query metric captures rebound effects.

Research agenda. Seven workstreams will close these gaps and turn this paper into a validated standard: backtesting against published provider figures; a metering study on current GPUs at a partner facility; empirical calibration of the uncertainty ranges; a design-partner pilot on one governed process; provider engagement on the recommended contract clauses; an assurance-readiness review under ISSA 5000; and peer review with publication of the factor library and simulation code. The full contribution statement, limitations and agenda are in the PDF.