Google data centers
Hyperscale campuses running Search, YouTube, Gemini and Google Cloud AI workloads
Google data centers are the company-owned campuses and leased Google Cloud regions that power Search, YouTube, Gemini and Google Cloud Platform. They grew from clusters of cheap personal computers built for web search in the early 2000s into hyperscale sites running custom TPUs, Arm CPUs and optical switching, with a fleet-wide power usage effectiveness of 1.09 in 2025.
- Gartner estimated 2.5 million servers across Google sites in July 2016
- Fleet-wide power usage effectiveness of 1.09 recorded for 2025
- Data center electricity reached 30.8 million MWh in 2024, 95.8% of company use
- Jupiter optical switching cut network power draw by 41% and capital cost by 30%
- The Dalles campus in Oregon was built in 2006 on cheap hydro power
What it is
Google data centers are the physical plants behind Google Search, YouTube, Gemini and Google Cloud Platform. They mix large storage drives, rows of rack-mounted servers, internal and external networking, cooling and humidity control, and software that spreads load and survives hardware failures. The estate includes Google-owned hyperscale campuses and GCP regions that also draw on leased colocation space. Gartner put the server count at 2.5 million in a July 2016 report, though Google publishes no official figure and the fleet keeps growing and being refreshed. Modern sites run custom Tensor Processing Unit accelerators, Arm-based CPUs and optical circuit switching, and the whole fleet averaged a power usage effectiveness of 1.09 in 2025.
How it came about
The first machines date to around 1998, when the search engine ran at Stanford on a Sun Microsystems Ultra II, two Intel-donated Pentium II servers, an IBM RS/6000 and assorted disk boxes. By 2003 the team described a cluster of roughly 15,000 cheap PCs handling about 1,000 queries a second at peak. Software written in-house compensated for flaky hardware. From 2015 the company added its own application-specific chips alongside x86 processors, and in 2024 it brought out Axion, its first custom Arm server CPU. A heavily customised Debian Linux replaced a Red Hat-based system through 2013.
The compute
Early racks held 40 or 20 servers per side, drawing 10 kW each at 400 W per square foot. Servers of 2009 and 2010 were open-topped custom boxes with two multi-core processors, eight DIMM slots and SATA drives, each carrying a 12-volt battery to cut power costs. Tensor Processing Units arrived from 2015 for machine learning training and inference, followed by liquid-cooled v3 pods in 2018, optically linked v4 in 2021, Trillium in 2024, Ironwood in 2025 and the 8t and 8i chips. The Jupiter fabric uses modified Clos topologies and, since 2022, MEMS optical circuit switches that steer light between aggregation blocks without conversion.
Where the GPUs are
At least 12 major installations sit in the United States, with the biggest known ones at The Dalles in Oregon, Atlanta, Reston, Lenoir and Moncks Corner. Europe holds large centres at Eemshaven and Groningen in the Netherlands and Mons in Belgium, and the Oceania site is in Sydney. The Dalles complex, built in 2006 beside the Columbia River, covers about two football fields and uses cheap hydroelectric power plus surplus fibre. Google bought the Summa paper mill in Hamina, Finland from Stora Enso for 40 million euros in 2009 and spent 200 million euros there, adding 150 million more in 2012.
Customers and deals
Search, YouTube and Gemini run on the company-owned campuses, while Google Cloud Platform customers use regions that mix Google sites with leased colocation. The public network reached 67 exchange points and 69 locations, carrying 882 Gbit/s of open connectivity as of May 2012, and Google peers openly with many ISPs to keep distribution cheap. DNS round robin sends users to different clusters, then internal balancing picks the least busy web server.
Money and ownership
Google spent 40 million euros buying the Finnish paper mill site in 2009, then invested 200 million euros in the data center and promised a further 150 million euros in 2012. Electricity for the data centers came to 30.8 million megawatt-hours in 2024, which was 95.8% of the company's total power use and more than double the 2020 figure. Consumption grew another 37% during 2025 as cloud and AI demand expanded. Optical switching in the Jupiter fabric cut network capital spending by 30% and network power draw by 41%.
Its place in AI
The 2003 cluster design, written up by Luiz André Barroso, Jeff Dean and Urs Hölzle, showed that software fault tolerance could replace expensive supercomputers. That approach shaped how large-scale web services are built. Google's private network ranks third among ISPs in the Atlas Top 10 listing, behind Level 3. The company also experimented with floating barges off California and Maine in 2013, described at the time as learning centres, before stopping the work and selling the vessels in 2014.
Questions and criticism
Electricity use is the clearest pressure point: data center consumption more than doubled between 2020 and 2024 and then climbed again by 37% in 2025, making up almost all of Google's power draw. Reliability is handled in software rather than hardware, so many cheap machines can fail without disruption. Details of the private wide-area network stay undisclosed, and the floating barge project was kept secret until press reports in 2013.




Questions people ask about Google data centers
What runs inside a Google data center?
Search, YouTube, Gemini and Google Cloud Platform all sit on this infrastructure. Sites combine storage drives, rack servers, networking, cooling and humidity control, plus in-house software for load balancing and fault tolerance. Custom Tensor Processing Units and Arm-based Axion CPUs handle AI and general workloads.
How large is the fleet?
Google gives no official server count. Gartner estimated 2.5 million machines in a July 2016 report, a figure that shifts as capacity is added and hardware refreshed. At least 12 major installations exist in the United States, with further large sites in the Netherlands, Belgium, Finland and Australia.
What is the Jupiter fabric?
Jupiter is Google's internal software-defined network built on modified Clos topologies. It once used a hierarchical electrical spine giving more than 1 Pbit/s of bisection bandwidth across tens of thousands of machines. Since 2022, MEMS optical circuit switches steer light paths between aggregation blocks, cutting network power use by 41%.
Why build beside a paper mill or a river?
Sites are picked for cheap, plentiful power and fibre. The Dalles complex in Oregon sits on the Columbia River to use hydroelectricity and leftover fibre from the dot-com era. In Hamina, Finland, Google bought a Stora Enso paper mill for 40 million euros in 2009 because renewable energy was close by.
What software keeps it running?
Most of the stack is written in house. Google Web Server fronts the internet, Google File System and Colossus hold data, Bigtable and Spanner store structured records, Chubby handles locking, and MapReduce with Sawzall spreads work across the cluster. C++, Java, Python and Go are the favoured languages.
Good reading and sources
- Official websiteOfficial site
- Hiding in Plain Sight, Google Seeks More Power (Published 2006)The New York Times, Article
- Google Data Center FAQData Center Knowledge, Article
- Google's ship based datacenter patent application surfacesDCD, Article
- Google data centers on WikipediaEncyclopedia
- More photos on Wikimedia CommonsPhotos
- Wikidata recordOpen data
The facts on this page come from the Wikipedia article "Google data centers", written by its contributors and used under the CC BY-SA 4.0 licence, and from its Wikidata record (CC0). EveryNeocloud rewrote the text in its own words, with AI assistance, and compares every page with the article to make sure no passage is copied. It may be reused under the same CC BY-SA 4.0 licence with a link back.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.