In December 2024 Amazon released Power Usage Effectiveness (PUE) data for many of their AWS cloud regions, and along with existing data from Microsoft Azure and Google Cloud, there is finally enough data to make some comparisons for regions spread around the world and see what changed from 2022 to 2023.
PUE is a measure of how efficiently a data center uses energy. In addition to the energy that the computers themselves use, there is energy needed to cool the data center, and losses in the transmission and conversion of electricity on its way to the computers. Energy is measured (and paid for) at the meter as it enters the building, and a PUE of 1.15 means that an extra 15% of the total energy used by the computers is used for cooling and overheads. PUE varies between about 1.04 and 2.0 in practice.
All the cloud providers run very efficient data center hardware configurations, and they have, in general, become more efficient over time. However, it is harder to cool data centers in warm and humid environments, so PUE for data centers in the tropics will tend to be at the high end of the range, and PUE for data centers in cold and dry parts of the world are at the low end of the range.
One way to use less energy for cooling is to use more water, so there is also a natural tension between PUE and Water Usage Effectiveness (WUE in liters/kWh). The cloud providers have all recently invested heavily in optimizing for WUE as well, so the latest data centers tend to have good PUE and good WUE, but this requires the best and latest technology. Lower-cost and older data centers typical of the kind that enterprises own tend to be closer to 2.0 PUE and have a high WUE as well. WUE was explained and compared across cloud providers in the story I wrote for The New Stack in July 2024.
Some of the regions deployed by cloud providers around the world are hosted by local service providers rather than in dedicated data centers built by the cloud providers. In this case the PUE is not included in the public numbers for two reasons, one is that it’s hard to attribute and allocate part of a shared resource when there is no information about what is in the rest of the datacenter, and the other is that the over-all PUE is often proprietary information that is owned by the service provider and they don’t allow it to be shared openly. Some additional PUE estimates may be available privately under a non-disclosure agreement so it’s worth asking your provider if you are operating in a region that doesn’t have a public PUE number.
Amazon/AWS PUE Data
The new PUE information from Amazon is described on their sustainability page. They also talk in some detail about new data center technology that will result in a PUE of 1.08 for what they are currently building. There is a short PUE Methodology pdf that basically says that they are following the relevant international (ISO) and European (CEN) standards. The global average 2023 PUE for AWS is 1.15, with AMER at 1.14, EMEA at 1.12 and APAC at 1.28. They say their best individual data center facility in Europe has a PUE of 1.04. Still, their best region is Melbourne, Australia, with a PUE of 1.08, and their worst is Hyderabad, India, with a PUE of 1.50. From year to year, AWS shows improvements in many regions.
AWS PUE data as stated in January 2025.
Microsoft Azure PUE Data
Microsoft published a lot of detailed information for 2022 as data center fact sheets including PUE estimates for all of their regions, but when they were updated in December 2024 the numeric data was omitted. Instead, a webpage provides a summary that provides information for only 11 out of 27 regions, that is offset from the calendar year and doesn’t cover all of 2023. We contacted the currently responsible teams within Microsoft and found that whoever disclosed the original 2022 data had moved on, and the 2022 data has since been removed from their website. We’ve archived it as part of the GSF cloud region metadata table.
The Microsoft web page PUE data as stated in January 2025.
Microsoft discloses their best PUE of 1.11 in Wyoming, USA, and their worst at 1.35 in Illinois. Their sustainability targets don’t mention PUE.
Google Cloud Platform GCP PUE Data
Google has been disclosing PUE data on a quarterly basis for years and includes it in their 2024 annual sustainability report with data for five years, 2019–2023. For the same reasons as AWS and Azure discussed above, they don’t report public PUE data on every region they operate in.
GCP PUE data as stated in January 2025.
Google’s best PUE result is 1.07 in Oregon, and its worst is 1.19 in both Singapore and Nevada. The AWS PUE discussion mentions that as new empty data center buildings are added to a region, it can cause PUE to be worse for a while until they start to fill up with equipment. This can be seen in the GCP regional data. Some locations support Google products that are not part of GCP regions, and there is a mapping of GCP cloud regions to the Google PUE data that we performed with help from Google engineers as part of producing the GSF cloud region metadata table.
Comparisons Between AWS, GCP and Azure
While all the major cloud providers have industry-leading PUE numbers that are likely to be much better than local data center alternatives, the numbers above show that GCP has the best transparency, with more data over a longer time period, and the best overall PUE for the regions being disclosed. Microsoft went from providing the most data for 2022 to the least data for 2023 and also has the highest PUE values overall. AWS sits in-between, with two years of data, and PUE values that are worse than GCP but mostly better than Azure overall. AWS includes data for regions like Hyderabad with higher PUE that aren’t disclosed at all by other cloud providers, and they have an excellent stated goal of 1.08 for their new data center builds.
Regional Comparisons
If you can choose which cloud provider to use in a particular region, and you’d like to make the most efficient use of the energy needed, then the above PUE data provides data for some scenarios.
Virginia is the biggest cloud region in the world, AWS PUE is 1.15, Azure is 1.14 and GCP is 1.08. For the nearby regions in Ohio AWS PUE is 1.12, GCP is 1.10.
Singapore is a major region in Asia for all the cloud providers, with a challenging tropical climate. AWS PUE is 1.30, Azure is 1.34, and GCP is 1.13 or 1.19 for each of their two facilities.
Ireland is one of the largest regions in Europe. AWS PUE is 1.10, Azure is 1.19, and GCP is 1.08.
What Is a Good PUE Assumption for the GPUs Powering the AI Boom?
There is a lot of concern about how much power the massive build-out of GPU capacity is going to need. Whatever power the GPUs use directly needs to be multiplied by the PUE of that datacenter or cloud region before accounting for its demand on the grid or its carbon footprint. Looking at the above data, I think there are two situations to think about. GPUs that are put into older enterprise data centers tend to overwhelm the existing infrastructure that wasn’t designed for very high power density, so the PUE is likely to be bad, I’d guess 1.5 or higher. However, the massive new data centers being purpose-built for GPU deployments are, in many cases, limited by their available power sources. The latest, most efficient designs will allow more GPUs to be powered and cooled in a given location, and I’d assume a PUE of 1.08, regardless of who is building it.
Green Software Foundation Real-Time Cloud Project
I’ve been leading the GSF real-time-cloud project for the last year or so. We’ve spent a lot of time discovering many sources, interfaces and products that collect and report energy and carbon data and documented their relationships in a large Miro flow chart. We’ve also published regional metadata collected from GCP, AWS and Azure and summarized it into a single table covering the 2022 data sets. We’re in the process of updating it to include the latest data releases to cover 2023 and are planning to produce estimates for 2024 and 2025 before the cloud providers disclose their data so that workloads that are running today can have some data to use that we think is the best guess available.
The post Cloud PUE: Comparing AWS, Azure and GCP Global Regions appeared first on The New Stack.