Candor Intelligence began with a question about environmental reporting. It became something broader. This is that story — including what we don't know yet.
Candor Intelligence began as an enterprise-focused environmental intelligence concept — helping organizations understand and report the environmental impact of their cloud infrastructure. The initial emphasis was on Scope 2 and Scope 3 emissions, cloud electricity consumption, and more timely, geographically specific environmental reporting than annual reconciliation allows.
The ambition was to make environmental reporting more transparent, accessible, and actionable. But building and researching that concept exposed a harder problem underneath it: the information available to independent researchers and cloud customers wasn't good enough to answer basic questions about where a workload's electricity actually came from, or when.
That wasn't a dead end. It was the starting point of a different, more fundamental investigation.
As the research extended into AI specifically, new questions piled on top of the old ones: the electricity required for training and inference, how that varies by model and hardware, how data centers draw power even while idle, and how little of any of this is independently verifiable from outside the company running it.
One distinction became central to everything built since: computational activity, electricity consumption, and greenhouse gas emissions are three different quantities, not interchangeable ones. A reduction in computation doesn't automatically imply an equivalent reduction in facility electricity draw, and a reduction in electricity draw doesn't automatically imply a proportional emissions reduction — it depends on what's actually happening on the grid at that moment. This isn't a new claim added for this page — it's already a structural rule in how this project's own engine is built, now simply stated in prose.
This section cites only sources that were directly checked, not recalled from memory. Where a claim below isn't backed by a link, it isn't published as fact.
What's still genuinely open, not resolved by the above: whether inference-side savings translate into facility-level electricity savings at realistic utilization (the idle-power research above complicates this, it doesn't settle it); per-request or per-facility attribution of a specific AI interaction (a stated non-goal of this project, not an unsolved engineering problem); and water or other non-carbon environmental costs, which this research pass didn't examine at all.
Independently evaluating AI's environmental impact remains genuinely difficult. Here's what that actually means, organized by the specific information gaps involved — not a blanket claim that every provider must disclose everything.
| Information | Why it matters | Where it stands |
|---|---|---|
| Model-specific inference energy | Understanding the electricity a given workload actually requires | Benchmarked for some open models (ML.ENERGY); not available for most proprietary models in production use |
| Facility-level electricity consumption | Distinguishing computational savings from actual electricity savings | Not independently published; this project uses grid-region data as a proxy, explicitly not facility metering |
| Workload-to-facility attribution | Connecting a specific AI interaction to physical infrastructure | Structurally unavailable from the outside — a stated non-goal of this project, not a gap it claims to close |
| Geographic & temporal electricity consumption | Evaluating emissions at the place and time electricity was actually used | Available at the facility level via EIA/eGRID, which this project uses directly |
| Marginal vs. average emissions | Evaluating what a specific load-shifting decision actually changes | This project currently uses average hourly fuel mix, not marginal factors — a real, stated limitation |
| Electricity procurement & matching claims | Understanding how "clean electricity" claims relate to actual consumption | Documented gap in the broader literature (RECs vs. 24/7 CFE); not something this project independently verifies per company |
| Water consumption | Environmental impact beyond carbon | Not researched by this project at all, as of this page |
Candor initially focused on improving environmental transparency. The research since has raised a harder question: is transparency the actual goal, or a mechanism for something more consequential?
Knowing that AI consumes electricity doesn't, by itself, reduce electricity consumption. Knowing that one model costs more to run than another doesn't automatically tell you which one to use. Knowing a data center sits on a carbon-intensive grid doesn't hand you a feasible alternative.
The question has shifted from how can we better measure AI's environmental impact to how can better measurement enable decisions that meaningfully reduce it. We don't assume more disclosure automatically leads to lower emissions — that relationship, between information and actual outcomes, is itself something to investigate, not assume.
None of the following is something Candor has built or demonstrated works. These are the questions worth investigating next.
Can choosing a smaller or more efficient model reduce electricity consumption while preserving the quality of a completed task? The evidence above says sometimes, not universally.
Can avoiding redundant inference — caching, deduplication, reducing repeated failed requests — produce real energy savings, and under what utilization conditions do those translate into facility-level reductions rather than just freed-up capacity for more requests?
Can the timing or location of flexible AI workloads reduce emissions? Which workloads are actually flexible enough for this to apply to?
How can organizations use more precise, temporally matched information to make better procurement decisions — and which disclosures would actually improve accountability without being infeasible to collect?
Understand — improve our understanding of AI infrastructure's environmental consequences through independent research, public datasets, and transparent methodology.
Evaluate — investigate which technological, operational, and organizational decisions can meaningfully reduce AI-related electricity consumption and emissions.
Inform — translate research findings into accessible evidence that can support technical decision-making, organizational accountability, and potential environmental policy.
These are ambitions, not achievements. Candor hasn't solved any of these problems — the point of publishing this page is to be honest about exactly where the line between "investigated" and "still open" actually sits.
The Infrastructure Tracker grew out of trying to connect AI infrastructure locations to the electricity grids supplying them. It can help investigate geographic carbon intensity, available infrastructure data, and public disclosure patterns.
Its limits are explicit, not hidden: geographic grid data alone cannot establish the electricity a specific AI request consumed, or the exact emissions tied to it. That's a structural limit of this approach, not a gap in effort.
Exploring more timely, actionable Scope 2 and Scope 3 reporting for enterprise cloud infrastructure.
Investigating what it actually takes to connect a cloud workload to physical infrastructure and the electricity grid behind it.
Extending that work into AI specifically — model-level energy, inference efficiency, and the real limits of current environmental estimates.
The current focus: investigating how better measurement could support operational changes that actually reduce environmental impact.
Every facility carries a confidence tier — confirmed, likely, or speculative — describing how certain the tenancy claim is, not how big or important the facility is.
For facilities a third party owns and leases to an AI company (most Anthropic and OpenAI sites), confidence comes from Epoch's own Users field. For facilities an AI company owns outright (Google's and Microsoft's own buildouts), confirmed Owner status plus independent multi-source verification is treated as sufficient — a different pathway for a genuinely different ownership structure, not a lowered bar.
A facility can draw from a clean regional grid on paper while actually 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, regardless of how accurately it's pulled.
Every weighted facility is checked for this across multiple independent sources — not assumed clean by default. When a facility fails this check, it's excluded and the reason is stated.
Confidence tier — how certain the tenancy is: confirmed (directly tagged by the source), likely (probable, less firmly sourced), speculative (an association appears in reporting but isn't confirmed).
Exclusion reason — why a facility isn't counted toward an estimate: shared site with no published capacity split; not yet operational; speculative-tier association; or on-site generation that makes grid data inaccurate for that site.
Megawatt ratings and estimated GPU-equivalent capacity sometimes imply different relative facility sizes for the same two sites. The cause hasn't been established — recorded as open, not papered over.
Regulatory filings and orders tied to a specific facility are a planned category, but nothing publishes until it clears the same primary-source standard as everything else here. Empty means nothing has cleared that bar yet, not that nothing exists.
New facility data is checked by diffing against what's already published, never auto-applied beyond a capacity or timeline update — but the refresh itself runs on demand, not a fixed schedule yet.
If a fact on this site is wrong, outdated, or missing its source — that's a real problem worth fixing quickly, not a footnote.