Built at any cost: How AI infrastructure is rewriting the cloud sustainability rulebook

And why the hyperscalers are hoping you won’t notice

The rapid expansion of AI infrastructure by hyperscalers is dramatically increasing global datacentre electricity consumption, outpacing sustainability commitments and creating new challenges in transparency and environmental reporting.

Summary for IT leaders:

• Despite operating a fraction of the physical footprint of AWS or Microsoft, Google records higher market-based Scope 2 emissions than either
• Oracle's Scope 2 emissions look artificially low because a significant portion of its infrastructure is embedded in datacentres operated by its rivals.
• Oracle also overwrites its historical emissions data annually, making year-on-year comparison impossible.
• All three major hyperscalers claim 100% renewable energy matching but contracting renewable energy and delivering it onto the grid are different things. AWS brought 2.7 GW online in 2024. Google brought 2.5 GW. Microsoft won't say.
• Behind-the-meter gas generation is quietly filling the gap between clean energy ambition and operational reality and no ESG report addresses it.
• Amazon refuses to publish water consumption data. The most recent figures available are from a leaked 2021 internal memo.
• Google's water consumption, when converted from the gallons format it prefers, is disproportionately high relative to its footprint which is consistent with over-reliance on evaporative cooling.
• Google generated more waste than Microsoft in 2024 despite being a fraction of the size.
• Oracle reports one waste category in pounds and another in litres, with no explanation of what the liquid is. Oracle has not responded to Computing's request for clarification.
• The absence of data is not a neutral finding. It is a deliberate choice.

2024 was the inflection point

2024 was the year that the AI infrastructure build out began in earnest, and whilst the AI infrastructure that the hyperscalers contracted and started to build hasn’t come online yet, the available data suggests that 2024 was the inflection point.

According to the International Energy Authority (IEA), annual increases in electricity consumption from the datacentre sector were reasonably steady up to 2023. The sector consumed an additional 28 TWh in 2021, 27 TWh in 2022, and 36 TWh in 2023.

In 2024, that increment shot up to 55 TWh, placing the total amount of electricity consumed by datacentres globally at 415 TWh, roughly equivalent to the entire annual electricity consumption of a medium-sized European country.

Based on the facilities currently under construction, that annual growth is expected to almost double to 104TWh in 2027 and hit 110 TWh in 2028. This does of course assume that all the facilities currently under construction, such as Stargate Abilene, will be completed.

The hyperscalers we’re looking at here aren’t accountable for all that electricity consumption. However, hyperscale operators control approximately 44% of global datacentre capacity. - an increase from 20% in 2017. That share is expected to reach 61% 2030.

Scope 2 comparison

To understand the impact of this increased electricity consumption, we need to look at Scope 2 emissions. These are emissions generated indirectly through the purchase of energy. Each company reports two figures: market-based emissions, which reflect the specific emissions associated with the electricity a company has contractually purchased, and location-based emissions, which reflect the average carbon intensity of the local grid where electricity is consumed. Location-based figures are always significantly higher.

Amazon is always going to have greater energy emissions in absolute terms than its peers because the rest of its business is based on massive fulfilment warehouses and transport infrastructure.

What makes this dataset striking is that despite operating a fraction of the physical footprint of either AWS or Microsoft, Google records higher market-based emissions than either.

AWS runs approximately 900 datacentres globally across more than 30 regions, an estimate based on credible reporting because no hyperscaler publishes its datacentre footprint. Microsoft operates at least 400 datacentres. Google, by contrast runs an estimated 28-30 campuses globally, albeit larger, more consolidated sites.

The scale difference makes Google's emissions profile hard to explain.

The same pattern is discernible in the location-based data. Google emitted approximately 1.3 million mtCO2e (metric tons of carbon dioxide equivalent) more than Microsoft despite a proportionately much smaller footprint.

Google’s ESG reporting claims that its datacentres “are some of the most efficient in the world.”

The data paints a more nuanced picture.

Oracle Scope 2 emissions of 1.2 million mtCO2e appear low for a hyperscaler of its scale. The explanation lies in multicloud architecture. A significant portion of OCI infrastructure is embedded inside datacentres operated by the other three hyperscalers. This arrangement reduces the Scope 2 emissions Oracle is legally required to report. By rights, the same emissions should sit within the ‘Purchased Goods and Services’ category within Scope 3 – which Oracle is not compelled to report.

There is a further transparency problem with Oracle's emissions reporting. Year-on-year comparison is impossible because historical reports are overwritten annually. Old reports exist but default back to the current year's data. Oracle has not responded to Computing's requests for access to historical reports.

Renewables – contract versus delivery

The amount of renewable energy generating power for the hyperscalers in 2024 has a significant effect on total emissions. This is where the gap between narrative and reality becomes most apparent.

Amazon, Google and Microsoft are all forthcoming about how much renewable energy they contracted in 2024. Microsoft signed agreements for an additional 19 GW of new renewable energy. Google contracted 8 GW and Amazon a further 6 GW. Oracle is less forthcoming.

But contracting renewable energy and delivering it onto the grid are two different things, and the gap between them can span years. AWS brought an additional 2.7 GW of renewable energy online in 2024. Google brought 2.5 GW online. Microsoft does not publish this figure.

Microsoft has contracted 40 GW of new renewable energy since 2020 and earlier this year announced it had achieved 100% matching of its global annual electricity consumption with renewable energy. That is a genuine milestone. But without disclosure of how much of that contracted capacity is generating power - estimates suggest 19 GW is currently live, though this is a cumulative figure - it is impossible to assess how much of the matching claim reflects operational clean energy and how much reflects carbon accounting.

This matters because all three companies use the same methodology: annual, global matching, which allows surplus renewable procurement in regions with abundant clean energy to offset fossil fuel consumption elsewhere.

Amazon and Google have both maintained this 100% match for two consecutive years. Google has gone further, setting a 2030 target to match consumption with carbon-free energy on an hourly, location-specific basis, which is a far more demanding standard. In 2024, it achieved 66% against that measure, despite a 27% rise in electricity consumption.

Renewable energy certificates, which underpin much of the matching claim, are better understood as a carbon accounting tool than a guarantee of additional clean energy generation. All three companies have achieved real milestones in renewable energy procurement. But their infrastructure build-outs have grown faster than the renewable energy that their reporting implies is powering them.

What is powering them?

A bitter irony of the growing public opposition to datacentres is that datacentre operators are increasingly opting for “behind-the-meter" power. This involves datacentre operators building their own fossil-fuelled generators and power plants to avoid placing a massive strain on local electricity grids.

Researchers such as Cleanview track gas permits for datacentres across the US. Whilst permits and built capacity are two very different things, gas turbines emit millions of tons of CO2e per annum. According to a recent report in Wired, Microsoft is reportedly looking into purchasing power from a Chevron-backed natural gas project in West Texas. This project alone, according to the state permit, could emit more than 11.5 million tons of greenhouse gases each year.

It simply isn’t feasible to believe that hyperscalers can build out this capacity and stay within the carbon targets they set themselves. To ask us to believe both things is to take us for fools.

No ESG report dwells on this gas-powered datacentre revolution.

Scope 3

We’ve spent a lot of time analysing Scope 2 findings, but Scope 3 emissions (indirect emissions in a company’s value chain) constitute between 95% and 98% of overall emissions, so they matter. They are also very difficult to quantify because most occur upstream, so you need supplier data to estimate them.

All hyperscalers use carbon credits to offset a big chunk of Scope 3 emissions, but are not obligated to show the scale of emission pre-offset or explain their carbon accounting methodology. Carbon credits very in quality and the opacity of accounting makes meaningful comparison difficult.

The only hyperscaler that makes the full extent of its Scope 3 emissions available is Amazon (although you have to look hard to find it.) Google, Microsoft and Oracle all report market adjusted totals with minimal explanation of the carbon accounting involved.

Microsoft does report on the total carbon removal credits contracted. The table below indicates just how much that spiked in 2024.

As you’d expect from the increases in CapEx in 2024, all hyperscalers saw chunky increases in the emissions associated with capital goods. This category includes emissions embodied in buildings like datacentres.

Microsoft, Amazon and Google all recorded big rises in upstream travel and distribution. All those buildings need inputs, and these emit carbon before they get to site. Fuel and Energy related activities (which counts emissions upstream from producing and delivering energy) also rose for Oracle and Microsoft, but declined for Amazon and Google.

The big three also saw rises in emissions associated with business travel and employee commuting.

Maybe time to rethink those back to the office mandates?

Water consumption

Amazon still does not publish total water consumption, preferring instead to emphasise water use efficiency (WUE). When offered the opportunity to share data for this article, Amazon once again declined.

Amazon’s report claims that the company is 54% of the way towards its “water positive by 2030” target, but provides no data to back that claim up.

In 2025, a leaked memo indicated that the company has chosen being criticised for opacity, rather than what it perceives would be the greater cost of transparency.

That memo did provide some figures based on 2021 data so in the absence of anything more recent we use them here, noting that water consumption will have grown alongside datacentre expansion. Primary consumption is given at 7.7 billion gallons, converted into megalitres for comparison with Google.

Neither company fares well. Google presents its figures in millions of gallons rather than megalitres (probably because the absolute number looks smaller) yet its consumption is striking. Microsoft and Amazon are making greater use of direct-to-chip and closed-loop cooling. Google, with an estimated WUE above 1 litre per kWh, appears not to be.

However, there are other factors to consider in addition to absolute volumes.

Amazon recently settled a $20.5 million lawsuit bought by residents of Morrow Country in Eastern Oregon who have suffered harm from drinking water polluted with nitrites. The problem originates from agricultural runoff, but as that water is taken from local water sources for cooling AWS datacentres, evaporation means that the water that exits these datacentres is even more polluted. It’s a second order problem, but a problem, nonetheless.

To be clear, Amazon is not the source of nitrites and in settling has not admitted to wrongdoing, but it has agreed to help the regions residents gain access to cleaner water.

Microsoft also had to spend big in 2024, but on water replenishment projects to keep it on track for its ‘water positive by 2030’ target. Total contracted water volume replacement more than doubled to 50 million cubic metres.

Waste

Of the four hyperscalers under the microscope, only two fully report waste volumes generated. Again, Google is out ahead based on its footprint in comparison to the others. Google created 6,300 more tons of waste than it did in FY23, taking it to 58,500 metric tons overall. It doesn't sound like a lot until you consider that Microsoft generated 41,290 metric tons.

Of this, all hyperscalers manage to recycle or refurbish roughly similar proportions of this waste.

Microsoft diverted 88.1% away from landfill and Google 84%. Amazon chooses not to publish waste volumes but will share a diversion rate of 85%.

Oracle publishes very limited data. Electronic waste is the only waste category reported - in pounds at 10,853,361. For the sake of comparison, this translates into 4,923.08 tons. However, Oracle claims that only 0.2% of that goes to landfill.

There is another mysterious category on non-electronic waste reported in litres – 8,430,157 to be precise. We can’t convert that to tons because we don’t know the density of the mystery liquid.

Oracle has not responded to Computing’s request for clarification on this matter.

Transparency

None of the hyperscalers examined here provides a complete picture. Amazon declines to publish total water consumption and won't explain how it calculates progress toward its water positive target. Microsoft won't say how much renewable energy it actually brought online in 2024. Google presents water data in units that minimise the apparent scale of consumption. Oracle wipes its historical emissions data annually, making year-on-year comparison impossible.

None has responded meaningfully to our requests for clarification and further data.

It’s a pattern of selective disclosure, and each omission avoids bringing an inconvenient truth to light. The renewable energy matching claims all four companies make are genuine achievements, but it’s a cherry-picked metric, reported at the expense of more granular data that tells a more complicated story.

The AI infrastructure buildout is making this worse, not better. As CapEx accelerates and emissions rise, the gap between what these companies are committing to and what they are delivering is getting wider, and the reporting frameworks are struggling to keep pace – especially in Oracle’s case.

Cloud procurement decisions aren’t directed solely by such concerns, but they are informed by them. If a vendor won't tell you how much water their datacentres consume, or how much of their renewable energy is generating clean power (and where it is generating it) rather than sitting in a pipeline, the absence of that data is a material fact about how seriously that vendor takes accountability.