Summary
- Google reports 100% for FY2025 on an annual, global, purchase-based match and approximately 65% on an hourly, location-specific carbon-free-energy share across data centres and offices. The 35-point gap is the averaging window, not arithmetic failure.
- The narrower problem is that the approximately 65% is an average whose weight vector — electricity by grid, or normalized load shares — I could not verify in any document I read. A 44-row Cloud-region table that looks like the audit trail describes a different population from the 22 grid rows behind the global figure.
Google had its largest load growth in history in 2025, "a 37% annual increase in electricity demand," and still matched "100% of our electricity consumption with renewable energy purchases (on a global and annual basis) for the ninth consecutive year" (2026 Environmental Report). The word that does the work is "annual." A volumetric match is satisfied by the sum over twelve months, wherever and whenever the clean electrons were produced. No single hour is being claimed clean.
Google says so itself, in the same document: "Going further than a 100% renewable energy match, 24/7 CFE requires an hourly and local match" (2026 report). Two metrics, two clocks, one page apart.
What the hundred percent actually measures
The ambition is dated: "In 2012, we set out on a pursuit to match 100% of our annual global electricity consumption with renewable energy purchases" (2026 report). The prior-year edition is more precise about the series: Google has "maintained a 100% renewable energy match on a global basis every year since 2017" (2025 Environmental Report).
That is a purchasing statement. It says contracts and certificates were bought in sufficient volume, inside the year, to cover the year's consumption. It does not say the power flowed at the moment of use, and nothing in the documents I read specifies the contract or certificate mix used to satisfy FY2025. The annual match is a balance, not a schedule.
What the sixty-five percent measures
The hourly figure runs on the other clock. Google Cloud's region page defines its metric as "the average percentage of carbon free energy consumed in a particular location on an hourly basis, while taking into account the investments we have made in carbon-free energy in that location" (Cloud region-carbon page). The prior-year report puts the same test in load terms: "every kilowatt-hour of electricity we use is matched with clean power produced on the same grid during the same hour" (2025 report).
Google has been explicit for years about how the two relate. "Of course, all regions are matched 100% with renewable energy on an annual basis, so the CFE% tells you how well matched the carbon-free energy supply is with our demand. A lower-scoring region has more hours in the year without a matching, local amount of carbon-free energy" (2021 Cloud post). And the 2022 post defines the target as "matching our electricity demand with carbon-free energy supply every hour of every day, in every region where we operate" (2022 Cloud post).
The FY2025 global result is one sentence long: "In 2025, we achieved a global average of approximately 65% CFE across our data centers and offices, remaining relatively flat compared to our 66% in 2024" (sustainability page). The word "approximately" is the publisher's, not mine. A1 (mine): the published global average and the annual match sit 35 percentage points apart for the same year — the size of the window, not a quantity of dirty electricity, and not a gap that can be closed without the hourly series.
The average is where the receipts stop
Here is the editorial problem, and it is small and specific. A consumption-weighted average needs a weight vector: how much electricity each included grid, or each included hour, contributed. Without it, an outsider can read the result but cannot rebuild it, and cannot separate two very different stories.
First story: the grids got cleaner, so the share rose. Second story: the load moved to cleaner grids, so the share rose even though no grid improved. Google's 37% demand growth makes the second story live rather than hypothetical — new consumption lands somewhere, and where it lands moves a consumption-weighted average. The documents I read let me state the mechanism. They do not let me decide which story FY2025 tells.
What Google prints instead is granular but differently shaped. The Cloud region page lists 44 region rows for 2025, from 0% (me-central1) and 1% (me-central2) to 100% (europe-north2), with entries such as 15% for africa-south1, 2% for asia-east2, 88% for us-central1 and 98% for europe-north1 (Cloud region-carbon page). A machine-readable dataset publishes the same values as fractions beside a grid carbon-intensity column in gCO2eq/kWh — africa-south1 at 0.15 beside 641.42; asia-east2 at 0.02 beside 488.52 (2025 dataset).
That table is easy to mistake for the audit trail of the 65%. It is not. A1 (mine): those 44 percentages sum to 2,475, an unweighted mean of 56.25% — a figure Google does not publish, and one that is not load-weighted, so it is not a counter-estimate of anything. A2 (mine): 44 Cloud-region rows against the 22 data-centre grid rows attributed to the report is exactly 2, and no document I read gives a mapping that would convert one into the other.
Separately, a third-party compilation prints a 22-row FY2025 per-grid table attributed to the 2026 report, with Singapore at 5%, Taiwan 15%, Japan (TEPCO) 23%, Finland 98%, Denmark 92%, and regional averages of 13% Asia Pacific, 68% United States, 83% Europe and 90% Latin America (AI Safety Tracker compilation). It also sets Google's share against the grid's own: MISO at 88% Google CFE against 36% grid CFE — a reminder that a high operator figure can sit on a mostly fossil-fired grid. That table is an attributed transcription. I could not open the report to confirm its rows, and the compilation gives no page or table label, so it is quoted as a compilation, not as a Google-published table. A4 (mine): the "nine of" ratio moved from nine of 20 grid regions in FY2024 (2025 report) to nine of 22 as attributed for FY2025 — the numerator held, the denominator grew by two, and the share fell from 45% to 40.9% without any region losing 80% status.
The Cloud page also says the hourly grid-mix and carbon-intensity data behind its metric come from Electricity Maps and "have not been assured" (Cloud region-carbon page). Read that narrowly: it is a statement about an input class, not a claim that Google's global aggregate or its own calculation is unassured.
What would make the number checkable
An analyst with the right inputs should be able to reproduce approximately 65% inside a stated rounding tolerance. That takes five things: the list of included locations, the aggregation rule and order of operations, the component CFE values preferably unrounded, the electricity volume or normalized load share per included location, and a crosswalk reconciling the 44 Cloud regions with the physical grid population behind the global data-centre-and-office average.
Normalized load shares, rather than absolute volumes, would close most of the gap at little commercial cost. A single global electricity total would not: a denominator without a distribution cannot weight anything.
The documents I read on 19 November 2026 did not contain that weight table. The 2026 report PDF exceeded my route's size limit, so its passages reached me through search-index renderings rather than a page-by-page read, and page and table labels were not verified; two methodology documents — the free-standing 24/7 CFE white paper and the 2021 methodologies-and-metrics paper — returned no readable text in either attempt. I therefore cannot say Google publishes no weights anywhere. I can say the metric is defined and its global weighted result is not independently reproducible from the documents I reviewed.
That is the narrower, and more useful, sentence.
The definitional work here is honest and unusually clear. The company separated its two windows in print, told readers which one describes delivered electrons, and published per-region detail that most operators never disclose. The remaining ask is not rhetorical. Publish the weight vector, the included population and the formula, and a stranger can check whether approximately flat in 2025 was grid progress, load migration, or some of each. Until then the headline travels and the arithmetic behind it does not, and the more checkable-looking table is the one whose denominator is least visible.
For the directory record behind this briefing, see Google LLC.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






