SUSTAINABILITY

AI for sustainability

At Google, we have a unique opportunity to help lead the transition to a more sustainable future by investing in AI solutions that help individuals, cities, and other partners in their efforts to collectively reduce one gigaton of their carbon equivalent emissions annually by 2030. We are also pursuing a bold goal to reach net-zero emissions across our operations and value chain by 2030, which includes operating our data centers and campuses on 24/7 carbon-free energy.


We’ve been bringing AI into our products and services for over a decade to improve the lives of as many people as possible. This includes everything from collaborating with the US Forest Service to update their fire spread prediction models for the first time in ~50 years; providing traffic light optimization recommendations to city and state governments; and working with cities to reduce extreme heat through Heat Resilience insights.

SUSTAINABILITY

AI for sustainability

At Google, we have a unique opportunity to help lead the transition to a more sustainable future by investing in AI solutions that help individuals, cities, and other partners in their efforts to collectively reduce one gigaton of their carbon equivalent emissions annually by 2030. We are also pursuing a bold goal to reach net-zero emissions across our operations and value chain by 2030, which includes operating our data centers and campuses on 24/7 carbon-free energy.


We’ve been bringing AI into our products and services for over a decade to improve the lives of as many people as possible. This includes everything from collaborating with the US Forest Service to update their fire spread prediction models for the first time in ~50 years; providing traffic light optimization recommendations to city and state governments; and working with cities to reduce extreme heat through Heat Resilience insights.

Project green light

Reducing stop-and-go traffic in cities

Green Light uses AI to optimize traffic lights to reduce vehicle emissions in cities, mitigating climate change and improving urban mobility

Project Green Light
Road traffic is a major source of greenhouse gas emissions, especially at city intersections where pollution can be 29x higher than on open roads. Half of intersection emissions come from stop-and-go-traffic, which could be prevented with optimized traffic light timing. However, current optimization methods are costly and provide limited information.

Green Light uses AI and Google Maps driving trends to model traffic patterns and build intelligent recommendations for city engineers to improve traffic flow. Early numbers indicate a potential for up to 30% reduction in stops and 10% reduction in greenhouse gas emissions.1 By optimizing each intersection, waves of green lights can be created, helping to reduce stop-and-go traffic in cities.
Project Green Light

Fuel-Efficient routing

Helping people save money and fuel

Fuel-efficient routing leverages AI in Google Maps, so people can get to their destinations as quickly as possible while minimizing fuel or battery consumption. Fuel-efficient routing is estimated to have helped prevent more than 2.9 million metric tons of carbon equivalent emissions since launching in 2021 – that’s like taking approximately 650,000 fuel-based cars off the road for a year.2
Learn more

PROJECT contrails

Using AI to help airlines tackle contrails

We’re using AI and satellite imagery to reduce climate-warming contrails

Project Contrails
Clouds created by contrails (short for condensation trails) account for roughly 35% of aviation’s global warming impact. Google Research teamed up with American Airlines and Breakthrough Energy to bring together huge amounts of data – like satellite imagery, weather, and flight path data – and used AI to develop contrail forecast maps to test if pilots can choose routes that avoid creating contrails. After these test flights, we analyzed satellite imagery and found that the pilots were able to reduce contrails by 54%.3
Project Contrails

solar savings estimator

Estimating potential solar savings for homeowners

The solar savings estimator in Google Search uses AI to generate a 3D rendering of any home’s roof, then assesses factors like roof space, weather, tree shade, estimated installation costs, local utility rates, and available incentives to help homeowners understand potential return on investment. If installing solar panels isn’t right for a home, users can also check the estimated savings from joining nearby community solar projects.

Solar Savings Estimator

flood forecasting

Making critical flood forecasting information universally accessible

Since 2018 we’ve made progress applying AI to forecast riverine floods. By building a breakthrough global hydrological AI model and combining it with publicly available data sources, we are able to predict floods up to seven days in advance. We are providing forecasts on our Flood Hub platform in more than 100 countries on five continents covering sites where more than 700 million people live.4
Flood Forecasting

firesat

Providing wildfire information to affected communities

We’re using AI to create breakthroughs in wildfire detection

FireSat

Google Research has been partnering with the US Forest Service to expand our existing fire simulation work and develop FireSat, a constellation of satellites dedicated entirely to detecting and tracking wildfires. After it launches, it will provide global high resolution imagery that is updated every 20 minutes, enabling the detection of wildfires that are roughly the size of a classroom.

FireSat

heat resilience

Tackling extreme heat in cities using AI

To lower city temperatures and keep communities healthy, Google Research is continuing its efforts to use AI to build tools that help address extreme heat. Our new Heat Resilience tool applies AI to satellite and aerial imagery, helping cities to quantify how to reduce surface temperatures with cooling interventions, like planting trees and installing highly reflective surfaces like cool roofs.

Environmental Insights Explorer

Responsibly managing the environmental impact of AI

Model optimization
Efficient infrastructure
Emissions reductions

Optimizing models to be faster and more efficient

Learn more

We’ve long been at the forefront of AI and machine learning, evolving years of deep learning research to develop techniques like quantization, which has boosted LLM training efficiency by 39% on Cloud TPU v5e5 – enabling models that are higher quality, faster, and less compute-intensive to serve. We also help developers reduce their digital footprint with tools like the Go Green Software guide.

Building the world’s most energy-efficient computing infrastructure

Learn more

Our data centers are among the most efficient in the world, and we regularly work to optimize their use of electricity, water, and materials. We’ve built world-leading infrastructure for the AI era, including Trillium, our sixth-generation TPU, which is over 67% more energy-efficient than TPU v5e.6 We’ve also identified combined best practices that can reduce the energy required for AI model training by up to 100x and associated emissions by up to 1,000x.7

Aiming to reach net-zero emissions by 2030

Learn more

We’ve set a goal to achieve net-zero emissions across our operations and value chain by 2030. As we continue to advance the future of AI, we’re working to reduce the embodied carbon impact of growing machine learning demand at our data centers. We’re actioning our net-zero mission by optimizing machine placement, promoting the reuse and upcycling of technical infrastructure hardware, and collaborating with partners.

Model optimization

Optimizing models to be faster and more efficient

We’ve long been at the forefront of AI and machine learning, evolving years of deep learning research to develop techniques like quantization, which has boosted LLM training efficiency by 39% on Cloud TPU v5e5 – enabling models that are higher quality, faster, and less compute-intensive to serve. We also help developers reduce their digital footprint with tools like the Go Green Software guide.

Learn more

Efficient infrastructure

Building the world’s most energy-efficient computing infrastructure

Our data centers are among the most efficient in the world, and we regularly work to optimize their use of electricity, water, and materials. We’ve built world-leading infrastructure for the AI era, including Trillium, our sixth-generation TPU, which is over 67% more energy-efficient than TPU v5e.6 We’ve also identified combined best practices that can reduce the energy required for AI model training by up to 100x and associated emissions by up to 1,000x.7

Learn more

Emissions reductions

Aiming to reach net-zero emissions by 2030

We’ve set a goal to achieve net-zero emissions across our operations and value chain by 2030. As we continue to advance the future of AI, we’re working to reduce the embodied carbon impact of growing machine learning demand at our data centers. We’re actioning our net-zero mission by optimizing machine placement, promoting the reuse and upcycling of technical infrastructure hardware, and collaborating with partners.

Learn more

1 Reductions in stops estimates are based on early data points from Google’s analysis of traffic patterns before and after recommended adjustments to traffic signals that were implemented during tests conducted in 2022 and 2023. Emissions reductions estimates are modeled using a Department of Energy emissions model. A single fuel-based vehicle type is used as an approximation for all traffic, and it is not yet adjusted for local fleet mix. These data points are averaged from coordinated intersections, and are subject to variation based on existing scenarios. We expect these estimates to evolve over time and look forward to sharing continued results as we perform additional analysis.

2 Google uses an AI prediction model to estimate the expected fuel or energy consumption for each route option when users request driving directions. We identify the route that we predict will consume the least amount of fuel or energy. If this route is not already the fastest one and it offers meaningful energy and fuel savings with only a small increase in driving time, we recommend it to the user. To calculate enabled emissions reductions, we tally the fuel usage from the chosen fuel-efficient routes and subtract it from the predicted fuel consumption that would have occurred on the fastest route without fuel-efficient routing and apply adjustments for factors such as: CO2e factors, fleet mix factors, well-to-wheels factors, and powertrain mismatch factors. We then input the estimated prevented emissions into the EPA’s Greenhouse Gas Equivalencies Calculator to calculate equivalent cars off the road for a year. The cumulative figure covers estimated emissions prevented after fuel-efficient routing was launched, from October 2021 through December 2023, while the annual figure covers estimated emissions prevented from January 2023 through December 2023. Enabled emissions reductions estimates include inherent uncertainty due to factors that include the lack of primary data and precise information about real-world actions and their effects. These factors contribute to a range of possible outcomes, within which we report a central value.

3 Using satellite imagery, large-scale weather data, and flight data, we trained a contrails prediction model. For this trial, we partnered with American Airlines to integrate contrail likely zone predictions into the tablets that their pilots used in flight so they could make real time adjustments in altitude to avoid creating contrails. We evaluated the model’s performance using satellite imagery, comparing the number of contrails produced in flights where pilots used predictions to avoid contrails, to the number of contrails created in flights where pilots didn’t use contrail predictions.

4 The estimated population covered is based on the forecasted flood risk area, using the WorldPop Global Project Population dataset.

5 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.

6 This calculation is based on internal data, as of May 2024.

7 “The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink,” Computer, vol. 55, July 2022.