What does "100% renewable electricity" mean for a data centre?

Google, Microsoft, Meta and Amazon all state that they cover 100% of their electricity consumption with renewable electricity: Google since 2017, Amazon for the 3rd year in 2025, Microsoft in its financial year 2025. Their data centres nonetheless remain connected to grids that burn gas or coal at certain hours. The 2 claims count different things.

The figure is a ratio between 2 annual volumes. The numerator is the renewable electricity bought by contract, produced on site or present on the grids where the company consumes; the denominator is its total consumption. Google says so in its 2026 report: the numerator counts all the renewable electricity bought, "regardless of the market" in which the electricity was consumed, and wind purchases in Europe offset the lack of purchases in Asia. Amazon and Microsoft describe the same method.

Purchases take 2 forms. A power purchase agreement, often called a PPA, commits the company to buy, over several years, the output of a specific wind or solar farm. A guarantee of origin, in Europe, is an electronic document proving that a quantity of electricity was produced from renewable sources; in France, each one is worth 1 MWh produced during a given month. It is sold separately from the electricity: at the French auctions of July 2026, where the State sells the guarantees of the installations it supports, a wind guarantee of origin was worth €1.76 per megawatt-hour. In 2025, 1,091 TWh of guarantees of origin were issued in the European AIB system, of which 82% for renewable electricity.

A data centre consumes day and night; a solar farm produces during the day. Our calculation for an actual day, 21 June 2025 in France: a 100 MW data centre, with constant consumption, buys a solar contract that produces over the day exactly as much as it consumes. The count for the day is 100%. Hour by hour, the solar power bought covers 51 % of its consumption: almost nothing from 21 h to 7 h, more than 2 times its needs at 14 h. The other half comes from the grid, and the midday surplus goes onto the grid for other consumers.

Google publishes this count for its own data centres, with the contracts it has signed and the carbon-free electricity present on each grid. In 2025, it covered 65% of its consumption worldwide hour by hour, 66% in 2024. The gap between regions depends on the grid and on contracts. In Finland, Google reaches 98%: the grid there is 95% carbon-free, and its contracts alone cover 75% of its consumption, hour by hour. In Singapore, it reaches 5%: the grid there is 5% carbon-free, and no contract delivered carbon-free electricity to it in 2025. In the MISO grid, in the central United States, it reaches 88% on a grid that is 36% carbon-free, thanks to its contracts. Google itself writes that companies at 100% over the year continue to depend on fossil plants at the hours and in the places where carbon-free electricity is lacking.

The GHG Protocol, the standard companies follow for their emissions inventory, has since 2015 required the emissions of electricity to be published in 2 ways: according to the average intensity of the grid where the electricity is consumed, and according to the contracts and certificates bought. The gap reaches a factor of 4,400 at Meta, which reports 1,358 tonnes of CO2e for its 2024 electricity according to its contracts, and 5.97 million tonnes according to the grid. Our note on Meta's 1,358 tonnes details this gap, and the analysis of AI decarbonisation gives Google's hourly figure.

The 2015 rules only require a certificate to be issued "as close as possible" to the consumption period, in practice the year. The GHG Protocol put out for consultation, from October 2025 to January 2026, a revision that would require each certificate to correspond to the same hour as the consumption. It received nearly 1,100 responses. In July 2026, its standards board said it was in majority favourable to 2 market-based methods: one with hourly matching, on the same grid and with new plants, the other compliant with the 2015 rules, allowed until a date still to be set. In October 2026, the GHG Protocol announced a new consultation in the 2nd quarter of 2027 and a common standard with ISO in the 4th quarter of 2028. Until then, the annual count remains allowed.

The remaining question is whether a company's purchases lower grid emissions. 2 modelling studies answer it. Jesse Jenkins's team at Princeton concludes that, under current US policies, annual matching has little or no effect on the long-term emissions of the power system, and that hourly matching lowers them markedly, at an extra cost that can exceed 20 dollars per megawatt-hour. Riepin and Brown, at the Technical University of Berlin, model buyers who have new plants built in their area, in Ireland: in 2030, if they represent 10% of corporate and services demand, they would reduce grid emissions by up to 0.4 MtCO2 a year with hourly matching, against 0.2 Mt with annual matching. None of the 2 studies measures the effect of certificates bought from existing plants.

In the 2 studies, hourly matching, on the grid where the electricity is consumed, reduces emissions more than annual matching. In February 2025, Microsoft stopped buying the short-term certificates from existing plants that had supported its annual "carbon neutrality" claim.

For us at Projet Celsius, a data centre cannot be carbon neutral, and neither can a company. ADEME has said so since 2022: carbon neutrality only makes sense at the scale of the planet, and an organisation can only contribute to collective neutrality; Carbone 4's Net Zero Initiative adopts the same rule. Contracts for carbon-free electricity have their place. The 2 studies cited above find them to have more effect on grid emissions when they fund a new plant on the data centre's grid and their output is matched hour by hour. An annual "100%" adds up purchases and the renewable share of grids over the year; Google's hourly rate, 65%, says what share of each hour of consumption is covered by carbon-free electricity. In the hub, we calculate the electricity of data centres with the intensity of the grid; in an emissions inventory, we ask for both of the 2 figures.