Methodology
Where the data comes from,
and how we use it.
What is reported, what is calculated, and what is only estimated — and where each comes from.
What is reported, what is calculated, and what is only estimated — and where each comes from.
Every figure on candorintelligence.ai comes from one of three places, and the site keeps them separate so you can always tell which you are looking at.
Data published by someone else, shown as published: hourly grid generation from the U.S. Energy Information Administration (EIA), and facility records from Epoch AI.
Arithmetic Candor does on reported data: carbon intensity, clean share, fuel shares, and the cleanest-hours comparison. Anyone can repeat it from the sources below.
Anything that depends on an assumption about how much electricity an AI workload uses. Candor does not measure any AI provider's electricity use. These figures are always labelled modeled estimates and never appear inside a facility page's factual sections.
The EIA publishes hourly electricity generation by fuel type for each U.S. balancing authority — the operator that balances supply and demand for a region. Candor reads it through the EIA's public API.
The mix of generation on the regional grid that serves a facility. It does not describe a facility's own power contracts, on-site generation or purchases. Where a facility generates a meaningful share of its own power on site, grid data describes the wrong thing for that site; the facility is then excluded and the reason stated (see section 5).
Hourly records are published with a delay of many hours — 16 or more in practice. "Most recent" on a facility page therefore means the most recent published hour, not the current hour, and the page shows its timestamp.
Numbers. The EIA returns values as text. Every value is converted to a number before any arithmetic.
Negative values. Hourly records sometimes include negative net generation (for example storage charging, interchange or reporting adjustments). Before any calculation, negatives are floored to zero so shares and intensity cannot become negative or exceed 100%. This is a simplification, applied the same way everywhere.
Missing hours. If an hour is not available it is shown as missing. It is never filled in or interpolated. A cleanest-hours comparison is shown only when at least 80% of the hours in the 30-day window were available; below that the page says it is not calculated and shows how many hours were.
Caching. Past hours do not change, so each hour is stored once and reused. Requests to the EIA are made in seven-day pieces for hours not yet stored, because the EIA returns at most 5,000 rows per request.
Energy storage. Stand-alone battery, pumped-storage and other storage discharge is not a fuel. Candor currently counts it under "other" (0.5 kg CO₂e/kWh). On a seven-day average this changes a grid's intensity by no more than about 0.004 kg CO₂e/kWh in October 2026 data, though storage-heavy grids such as Arizona Public Service (about 12% of generation) could differ more hour by hour. Solar and wind plants with integrated batteries are counted as solar and wind.
For each hour, intensity is the generation-weighted average of per-fuel lifecycle emission factors: each fuel's share of generation multiplied by that fuel's factor, summed.
| Fuel | kg CO₂e per kWh | Fuel | kg CO₂e per kWh |
|---|---|---|---|
| Coal | 0.820 | Solar | 0.048 |
| Oil / petroleum | 0.650 | Geothermal | 0.038 |
| Other / unknown | 0.500 | Hydro | 0.024 |
| Natural gas | 0.490 | Nuclear | 0.012 |
| Biomass | 0.230 | Wind | 0.011 |
Lifecycle medians from the IPCC Fifth Assessment Report (2014), harmonized by the NREL Life Cycle Assessment Harmonization Project. "Lifecycle" means they include more than combustion — fuel supply, construction and so on.
The share of generation that is not coal, natural gas or oil. Biomass is therefore counted as non-fossil in this figure even though it carries a non-zero factor above. That is a definition, not a judgment about biomass.
A location-based average intensity of the grid — not a marginal figure, and not a purchase-based (REC) one. It is consistent with the GHG Protocol's location-based approach and with the direction of 24/7 carbon-free-energy accounting. Candor makes no regulatory or compliance claim.
When hourly EIA data is not available, a facility page can show a regional annual average from EPA eGRID 2023 (Revision 2), using the CO₂-equivalent column of its Summary Tables. It is labelled "EPA eGRID 2023 regional emission factor used as fallback" and has no hourly spread.
Only eGRID subregions whose factor has been checked against the EPA spreadsheet are used: SRVC, RFCW, CAMX, NWPP, SRSO, SPSO, AZNM, MROW, ERCT and SPNO. A request for any other subregion is refused rather than guessed.
Each fuel's share of generation is its megawatt-hours divided by the hour's total, as a percentage. The facility page shows the mix for the latest published hour.
Where a daily view shows composition, a day's shares are the sum of each fuel's megawatt-hours across the hours observed that day, divided by the day's total. They are not an average of hourly percentages, which would over-weight quiet hours. Any fuel with a positive share is shown, even if it rounds to under 0.1%.
The charts use the last 30 days of published hourly data; a day shows only the hours that were observed. Candor compares the average intensity of all hours with the average of the cleanest 25% of hours and of the dirtiest 25%, and states the reduction a flat load would see if it ran only in the cleanest hours. When fewer than 80% of the window's hours were available, no comparison is shown, and the page says how many were.
This is a projection from recent patterns, not a record: "Based on recent grid patterns, a workload like this would project to roughly X kg CO₂." Projections are not recorded or saved.
Facility records — location, owner, users, capacity and construction timeline — come from Epoch AI's "AI Data Centers" dataset (CC BY 4.0).
An automated job checks the dataset monthly and proposes changes for review. Capacity and timeline updates to facilities already tracked are applied. New facilities, and any change to a tracked facility's owner or users, are never applied automatically: a person reviews each one, with the same checks every existing facility went through.
Each facility is tagged confirmed, likely or speculative, describing how certain the tenancy claim is — not how large or important the facility is. For sites a third party owns and leases to an AI company, the tag comes from Epoch's users field. For sites an AI company owns itself, confirmed ownership plus independent multi-source checks is treated as sufficient.
Each facility is matched to a balancing authority using the EIA's own utility data (Form 861, 2025). It lists the utilities that serve each county and each utility's balancing authority. A facility is assigned a grid when the utilities that serve its county agree on one, or when reporting names the serving utility. Where the county's utilities sit on different grids and the serving utility can't be confirmed, no grid is assigned. Each assignment records its basis, a confidence level (high or medium), and what was not checked, and it is shown on the facility page. A grid is a fact about a place, so a site is mapped even when its link to a provider is unconfirmed; mapping does not count it toward any estimate.
A facility is not counted toward any estimate if it has an unresolved tenant split with no published capacity split, is not yet operational, is a speculative-tier association, or runs on on-site generation that makes grid data inaccurate for it. Each exclusion states its reason.
Grid mapping currently covers the United States; facilities outside North America are not mapped. Capacity figures are Epoch's, in megawatts and in H100-equivalents, and are labelled as estimates. The two measures do not always imply the same relative size — recorded as an open question, not papered over.
A facility can draw from a clean regional grid on paper while running a meaningful share of its load off dedicated on-site generation. When that's true, grid fuel-mix data describes the wrong thing for that site. Facilities that were checked across multiple independent sources are recorded as such; where on-site generation was not researched, the facility page says so. A facility that fails the check is not assigned a grid, and the reason is stated.
Candor assigns a facility to a grid wherever the public record allows it. As of 22 October 2026, 33 of the 38 facilities on the tracker have one. For the other five, the facility page says why. The tracker covers the United States; facilities elsewhere in the dataset are not tracked. A grid is a fact about a place, so a site is mapped even when its link to a provider is unconfirmed; mapping it does not count it toward any provider's estimate.
One correction worth stating: five sites had no address in the facility dataset Candor uses. Their locations are in the public record, in developer and regulator announcements, and Candor now uses them. A missing address in a dataset is not the same as a location that isn't public.
A limit of the data. Three sites are reported to generate significant power on their own, so the regional grid's mix would describe the wrong thing for them. For one, the developer announced on-site fuel cells as the primary power source; for another, reporting describes a gas microgrid that has not yet been checked against primary documents.
Researched, and not settled. For two sites, the county is served by utilities on different grids, or the reporting describes power from two of them, and the dataset doesn't say which building is meant. Candor would rather show no grid than the wrong one.
A limit of the data feed. A cleanest-hours comparison is shown only when at least 80% of the hours in the 30-day window were available. Below that, the page says how many were. Missing hours are left out, never estimated.
Each mapped facility also shows how its grid was assigned, how confident that assignment is, and what was not checked (see how a facility is assigned to a grid).
No AI provider's actual electricity use per prompt or per token is available to Candor, and almost none is public. Anything that converts prompts or tokens into electricity is therefore a modeled estimate: built only from the published research below, shown with a range, and never mixed into a facility page's factual sections.
| Provider | What exists | How strong it is |
|---|---|---|
| Google (Gemini) | Median Gemini Apps text prompt: 0.24 Wh, measured by Google on its own production serving (accelerators, host, idle capacity and data center overhead). | Provider-measured — the only one found. Covers the Gemini Apps median prompt only. |
| OpenAI | An unexplained statement of about 0.34 Wh per average query; an independent estimate (Epoch AI) of about 0.3 Wh for GPT-4o with a 0.1–4 Wh range; 13 models in an independent benchmark. | A statement with no method, plus independent estimates. |
| Anthropic (Claude) | No disclosure. Three models in an independent benchmark: Claude-3.7 Sonnet, 3.5 Sonnet and 3.5 Haiku. | Independent estimate only. |
| Microsoft | No product-specific figure. Microsoft Research authors published a general model of frontier-scale inference (median 0.31 Wh per query). | A general model, not a measurement of a Microsoft product. |
| Meta, DeepSeek | Meta publishes training energy and emissions for each Llama model in its model cards, and nothing about inference. Ten Llama and four DeepSeek configurations appear in the independent benchmark; open models have GPU-only measurements. | Training disclosed by Meta; inference is independent estimate, GPU only. |
| Mistral | A lifecycle analysis: 1.14 g CO₂e per 400-token response. | A different metric — carbon, including manufacturing. |
The scenario calculator uses these published ranges, one per provider, as of 2026-10-22. These are published figures, not Candor measurements. Each source defines a 'typical prompt' differently and measures a different boundary; see each row. Most describe 2024-2025 models. Reasoning mode and long outputs multiply them (see the methodology). Never an average across sources.
| Provider | Reference range (Wh per typical text prompt) | Published figures behind it |
|---|---|---|
| Google (Gemini) | 0.24 Wh Single measured figure. Google states it does not represent every Gemini Apps prompt. |
|
| OpenAI | 0.30–0.42 Wh Non-reasoning, GPT-4o class. Published estimates for one model span 0.1 to 4 Wh. Newer and reasoning models are higher and mostly unmeasured. |
|
| Anthropic (Claude) | 0.95–0.98 Wh Short prompt, three 2024-2025 Claude models. No figure from Anthropic. Longer prompts: 3.0 to 4.5 Wh (1,000 tokens in and out) and 5.7 to 8.0 Wh (10,000 in, 1,500 out). No benchmark covers Opus or current Claude models. |
|
| Microsoft | 0.16–0.60 Wh A general model of frontier-scale inference, not a Copilot measurement. Median 0.31 Wh. Long reasoning-style responses are about 3.9 Wh (assumes about 5,000 output tokens). |
|
Boundaries differ. Google counts accelerators, host, idle capacity and facility overhead. The main independent benchmark counts only actively assigned GPUs and then applies a facility overhead factor. Open-model benchmarks count GPUs only. Mistral counts a full lifecycle. Figures from different sources are not the same quantity, even when the numbers look alike.
The direction of error is disputed. Peer-reviewed work from Microsoft Research (Joule, 2026) argues that widely cited estimates made outside production conditions are overstated by 4–20×. It does not name Claude or any specific commercial product. Other sources argue the opposite for totals. Any single figure could be too high or too low.
Jegham et al. (2025) is a preprint and is modeled, not measured. It combines public API timing data with published hardware specifications and runs 10,000 Monte Carlo draws; the hardware is inferred, not observed. It reports energy per query at three defined prompt sizes: short (100 input, 300 output tokens), medium (1,000 in, 1,000 out) and long (10,000 in, 1,500 out).
No source covers current-generation models. The benchmark covers 2025-era models.
Every claim above was re-read against its primary source, and the full survey, quotations and verification log are kept with the project's documentation.
Most of what is unresolved here is unresolved for one reason: the information that would settle it is not publicly disclosed. It is not that it hasn't been looked for. Where the reason is different, it says so.