Read Google’s 10th annual Environmental Report

Jun 27, 2025

|

Our annual report details how we’re working to address the increased energy demands of AI to enable positive impact, and how AI can be used to build a more energy-efficient and resilient world.


Kate Brandt

Chief Sustainability Officer, Google


a picture of trees from the base of a forest

Today we’re releasing our 10th Environmental Report.

This year’s most important highlight demonstrates a pivotal milestone: We reduced our data center energy emissions by 12%. 1

Delivering record clean energy while investing in future energy breakthroughs

Despite a 27% increase in electricity demand to power our data centers, we successfully decoupled our operational energy growth from its associated carbon emissions. This was largely due to more than 25 clean energy projects we’d contracted over the past several years — some as far back as 2019 — coming online in 2024. Together, they added 2.5 gigawatts of new clean energy to the grids that served our operations last year. That’s roughly equivalent to more than 4 million solar panels. 2 This record investment pushed our carbon-free energy use from 64% to 66% on an hourly basis, 3 proving that with focused effort, it’s possible to power AI and all our products and operations with cleaner energy as we scale.

We continue to build on these gains. In 2024, we made our largest-ever procurement of clean energy, adding 8 gigawatts to our portfolio, more than we’ve ever done in a single year. 4 But you can’t run the future on yesterday's infrastructure and grid. Powering the next wave of innovation — especially AI — requires us to solve some of the most significant challenges of our time: not only generating vast amounts of clean, reliable energy, but also using it efficiently.

So we’re focused on building and operating the world’s most energy-efficient data center infrastructure, optimizing our models and hardware to use less electricity, and pushing the frontiers of advanced energy development.

In 2024, Google data centers used 84% less overhead energy than the industry average. 5 And we’re leading the industry in making AI models better, faster and more efficient, for example through techniques like quantization which have sped up large-language model training efficiency by 39%. 6 Ironwood is our seventh-generation and most powerful, capable, and energy efficient Tensor Processing Unit (TPU) — using nearly 30 times less energy than our first Cloud TPU from 2018. 7

We’re also actively investing in the creation of next-generation solutions like advanced nuclear and enhanced geothermal. These powerful energy sources ensure we can continue to promote economic growth and, most importantly, build and deliver the benefits of AI to people all over the world.

Advancing AI innovation through our products and research

Our teams are putting AI to work right now, and the impact is impressive. Take just five of our AI-powered products. In 2024, they helped others collectively reduce an estimated 26 million metric tons of CO2 equivalent from their own footprints. 8 To put that in perspective, that’s like taking the energy use of over 3.5 million U.S. homes offline for a year. 9 This is happening today with products like Nest thermostats automatically saving energy, our Solar API helping installers design better systems, and fuel-efficient routing in Google Maps.

And it goes beyond emissions. We’re deploying AI to help communities become more resilient. We’ve built tools like FireSat to help detect wildfires earlier and Flood Hub for flood prediction, giving people critical time to prepare and stay safe. Ultimately, we’re putting helpful technology in the hands of billions to solve real-world problems.

Progressing toward our climate moonshots required broader change

We’ve always set bold ambitions for ourselves — what we call “moonshots.” But as we head toward 2030, it's clear that the challenges are increasingly complex and it’s important to be up front about the hurdles we’re facing.

The reality is, while our data center energy emissions are down, total emissions — encompassing the activities of third parties in our supply chain that we rely on — have increased by 11% year-on-year. 10

This is largely driven by the emissions from our supply chain as we continue to grow, as well as the fact that we operate globally, and the world isn’t decarbonizing fast enough or at a uniform pace. In some key regions, like parts of Asia Pacific, the grid is still in the early days of its clean energy transition. For both Google and our suppliers, we can’t outpace the infrastructure that isn’t available on the ground, which is why we’re partnering with others to accelerate solutions.

When it comes to the next-generation clean technologies we need — the solutions that go beyond wind and solar like enhanced geothermal, advanced nuclear, nuclear fusion, and more — they’re still in their early stages. They’re not yet deployed at the scale required and can remain expensive to get online.

Looking ahead

We’ve learned a tremendous amount over the last several years. We now have a better understanding of what it takes to get further down the path toward our climate moonshots — including what’s actually possible in different parts of the world, and how we can apply our resources and unique capabilities to have the greatest possible impact.

We remain focused on our environmental ambitions as we advance the frontiers of AI and all the ways it can benefit humanity.

1

We use the term data center energy emissions to describe scope 1 and 2 (market-based) emissions that result from our Google data center operations.

2

According to the U.S. Department of Energy, 1 gigawatt is roughly equivalent to 1.887 million photovoltaic panels, meaning 2.5 gigawatts is roughly equivalent to more than 4 million photovoltaic panels.

3

For details about this calculation, refer to the Methodology section in the Appendix of our 2025 Environmental Report.

4

The total gigawatt figure represents primarily PPAs, and includes some generation from targeted clean energy investments where we also receive EACs. Actual generation may vary from the signed amounts based on changes during construction or project terminations.

5

According to the Uptime Institute's 2024 Global Data Center Survey, the global average PUE of respondents’ data centers was 1.56. The overhead energy use comparison was calculated as follows: (1 - (Google’s overhead energy use [0.09] divided by the industry average overhead energy use [0.56])) x 100 = 84%.

6

This estimate is based on our internal analysis comparing the BFLOAT16 / INT8 model step time ratio measured on the MLPerf 3.1 GPT-3 175B model. The results (11,798ms / 8,431ms = 139%) can be interpreted as a 39% speed improvement and, in turn, training efficiency.

7

These calculations are based on internal data, as of March 2025. Google’s TPU power efficiency relative to the earliest generation Cloud TPU v2 is measured by peak FP8 flops delivered per watt of thermal design power per chip package.

8

To estimate aggregate enabled emissions reductions, we first estimated annual reductions for five products individually (Google Earth Pro, Solar API, Nest thermostats, fuel-efficient routing, and Green Light) and then combined the totals. For details about the individual calculation methodologies, refer to their respective endnotes in our 2025 Environmental Report.

9

Greenhouse Gas Equivalencies Calculator,” U.S. Environmental Protection Agency, November 2024, accessed June 2025.

10

In 2024, we adjusted the boundary for our carbon reduction ambition—in line with SBTi’s scope 3 guidelines—to exclude certain scope 3 emissions activities. For more details, refer to the Ambition-based carbon footprint and Methodology sections in our 2025 Environmental Report.