What do we really know about the environmental costs of the infrastructures behind the AI models we use daily? To answer this simple question, there is first what we currently know, from energy consumption and related carbon emissions, water usage for cooling systems in data centres, or material needs to produce GPUs. There are then a lot of blind spots, like the so-called “enabled emissions” that occur when AI is used to accelerate polluting activities, or the “rebound effect” where efficiency gains lead to an increase in a technology usage, which eventually increases overall emissions.

In this landscape, the ecological footprint of data centres remains however largely concealed behind high-level corporate pledges and selective transparency. Every year, the publication by tech companies of their annual environmental reports offers a unique opportunity to get environmental information on the development of AI and its related infrastructures. In this opinion piece, I review the most recent environmental sustainability reports published by Microsoft and Google, offering a glimpse into the ecological realities of two of the world’s most influential tech giants. In their communications campaigns, they emphasise carbon removals, water replenishment, and renewable energy purchases, but are these all their environmental reports have to say?

By this short piece, I want to make sure that all citizens have access to the full depth of the environmental information (or lack thereof) that can be found in these sustainability reports. Citizens and policymakers are of course free to adhere to big tech communication narratives, but it should be based on full information and not only on marketing campaigns conducted by the companies concerned.

A CONTINUOUS DRIVE IN ENERGY USE AND RELATED CARBON EMISSIONS FROM DATA CENTRES

Microsoft’s electricity consumption almost tripled between 2020 and 2024, climbing from 10.8 million MWh to 29.8 million MWh. Google’s trajectory mirrors this, with usage more than doubling in the same timeframe, from 15.2 million MWh in 2020 to 32.2 million MWh in 2024. These figures are alarming on their own, but what’s more concerning is how they sit alongside declining market-based carbon emissions.

Under market-based accounting, which incorporates corporate renewable energy certificates and power purchase agreements, both companies show decreasing emissions. Microsoft’s market-based scope 2 emissions¹fell from 456,119 metric tons CO₂ in 2020 to 259,090 in 2024. Google’s market-based emissions dropped 10% over the past year. But these figures are largely decoupled from physical reality.

The more accurate location-based emissions, which reflect the carbon intensity of local grids where electricity is actually consumed, tell a different story. Microsoft’s location-based scope 2 emissions more than doubled in four years, rising from 4.3 million metric tons CO₂ in 2020 to nearly 10 million in 2024. Google’s climbed from 5.8 million to over 11.2 million metric tons CO₂ over the same period. These numbers reflect the true planetary burden of AI development and data centre operations:

Scope 3 emissions, which include the carbon costs of a company’s entire value chain², are also downplayed. Microsoft reported 15.1 million metric tons in 2024, down slightly from the previous year but still forming most of their total footprint. Google reported 12 million metric tons, but neither company discloses “enabled emissions”, i.e. carbon released through downstream use of AI tools, including those applied in oil exploration or logistics optimisation. These emissions may be substantial, as Karen Hao argues, yet remain completely absent from disclosures. They should, however, be included in Scope 3 emissions as this category includes “the use of products and services sold”.

If one replaces the market-based scope 2 figures with the location-based ones and add scope 3, Microsoft’s actual emissions in 2024 would be around 25.2 million metric tons CO₂, not the 15.5 million they report. Google’s total emissions would reach 23.4 million, not the marketed 15.2 million. Such gaps deserve scrutiny not because companies are breaking the rules (they are not) but because the rules allow them to tell only part of the story.

MISLEADING PROGRESS NARRATIVES

Google highlights in its sustainability report a 12% reduction in “data centre energy emissions”, a claim made plausible only through market-based accounting. In reality, data centre electricity use increased by 27% between 2023 and 2024 alone, and location-based emissions are up from 22%, making the reduction in emissions a product of credit accounting.

Moreover, Google’s claim that its products “enabled others to reduce 26 million tons of CO₂ emissions” warrants scepticism. More than half of this figure (15 million tons) is attributed to renewable project siting facilitated by Google Earth Pro, based on interviews with corporate partners and absent third-party validation. These are speculative savings, yet they are reported as factual environmental benefits.

The narrative disconnect is stark. Enabled emissions are not reported, while speculative avoided emissions are highlighted. As a result, environmental benefits are inflated while environmental costs are hidden.

COMMUNICATING ON REPLENISHMENT WHILE DEPLETING RESOURCES IN WATER-STRESSED AREAS

Water is another area of concerns when it comes to the environmental costs of AI. Google reports that it replenished 64% of its freshwater consumption in 2024, up from 18% the year prior. Yet this comes as water consumption rose 28% in just one year, from 24 to 30 billion liters. Withdrawals, including from water-stressed areas, surpassed 41 billion liters, over three-quarters of which was potable water.

Google’s data centres alone consumed 27 billion liters of potable water in 2024. Furthermore, 28% of total withdrawals occurred in regions with medium or high water stress, exacerbating local vulnerability even as Google celebrates replenishment programmes.

Microsoft, for its part, reports a drop in water consumption from almost 8 billion liters in 2023 to almost 6 billion liters in 2024. This initially appears positive until one notes the switch in methodology: Microsoft now estimates water withdrawals based on water use efficiency metrics, rather than direct measurements. In short, the numbers changed because the method did.

Notably, Microsoft acknowledges data limitations and estimates that 46% of its withdrawals still occur in water-stressed areas. Whether the new methodology leads to underreporting remains unclear, but it certainly hinders year-over-year comparisons.

THE WAY FORWARD: ACCOUNTING, TRANSPARENCY, AND OVERSIGHT

What emerges from a closer look at these documents is not a textbook illustration of greenwashing, but a case of obfuscation-by-complexity: figures selectively communicated, accounting methods favouring best-case narratives, and a continuing disconnect between environmental marketing and environmental materiality. The reports do contain valuable information, much of which is independently verified, and these companies are clearly investing in decarbonisation and water stewardship initiatives. But sustainability accounting remains structured in ways that allow corporations to present a best-case image to the public while burying less flattering realities in appendices and footnotes.

Carbon accounting debates, particularly around market- vs. location-based emissions, need urgent regulatory clarification as market-based emissions can sometimes be misleading for an uninitiated reader. Market-based methods may incentivise investment in renewable energy, but they obscure actual physical emissions. Similarly, unreported enabled emissions remain a problematic blind spot. Without their inclusion, the full lifecycle environmental cost of AI cannot be understood.

Finally, selective transparency undermines democratic oversight. Only experts with time and training can parse these lengthy, complex documents. The result is an information asymmetry where companies control the narrative and citizens, journalists, and even policymakers struggle to respond effectively. Clarity and honesty in accounting methods, more transparency and increased public oversight over the environmental costs of data centres seem more necessary than ever to ensure public acceptance and a sustainable future for AI technologies.