Hyperscale self-build owners/operators
BOTTLENECKA handful of hyperscalers design and fully occupy their own data centres, so the best AI capacity never reaches the open colocation market.
Cloud and internet companies that design, build and fully occupy their own data centres. They avoid third-party margins and optimise directly for AI workloads, capturing value through cost efficiency and scale.
Cloud and internet giants that design, build and wholly occupy their own facilities, managing the full capital programme and fit-out.
Why the concentration exists
Hyperscale self-build owners/operators are cloud and internet companies that design, construct, and fully occupy their own data center facilities rather than leasing from third-party providers. These companies own core campuses in strategic regions and carry most capital costs on their own books, often acquiring land without leveraging debt and self-funding construction. The Big Five operators—Amazon, Microsoft, Google, Meta, and Oracle—drive the bulk of new development, with combined capital expenditure for 2026 projected between $600 billion and $700 billion, representing a 36% year-over-year increase. Approximately 75% of this spending, exceeding $450 billion, is tied directly to AI infrastructure.[11][17]
Self-build enables hyperscalers to capture value through cost efficiency and scale by avoiding third-party margins, though it requires managing long equipment lead times and construction complexity. Utility-grade transformers currently have lead times of 60 to 90 weeks, medium-voltage switchgear takes 52 to 78 weeks, and generator sets require 40 to 60 weeks. Hyperscale campuses today routinely exceed 100 MW per phase, and a handful of announced AI campuses are planned in the 1 to 5 GW range, requiring massive coordination among utilities, construction partners, and workforce providers.[5][23][3]
What the evidence shows
Hyperscale self-build segment holds 54.1% market share in 2024.
gminsights.comHyperscale owners projected to account for 67% of all capacity by 2031.
datacenterdynamics.comHyperscale companies operate 41% of worldwide data center capacity; over half is self-built.
srgresearch.comWho supplies it
The four largest US hyperscalers—Amazon, Microsoft, Alphabet, and Meta—together operate roughly 19.4 GW of self-built IT capacity as of Q1 2026. Amazon Web Services operates approximately 6.2 GW of self-built IT capacity as of that date. In 2023, global hyperscale self-build capacity reached 15.5 GW operational, with 73% coming from AWS, Microsoft Azure, Google Cloud, and Meta. Amazon, Microsoft, and Google control roughly 59% of global hyperscale capacity, with the Big Five operators driving the majority of new development.[16][13][11]
Apple owns eight data centers in the US, Europe, and China with plans to build more, while also maintaining multi-year lease arrangements for cloud-computing services from providers like AWS and GCP. Alibaba Cloud operates massive data center campuses including the Zhangbei Data Center in Hebei and the Ulanqab Data Center in Inner Mongolia, with international facilities in Singapore, Indonesia, Dubai, Germany, Japan, the UK, and the US. Hyperscalers partner with Quanta Computer Incorporated as a primary original design manufacturer, with manufacturing across eight regions and customer support via service centers in 11 countries.[22][18][12]
Who controls it
What it depends on, and what depends on it
Hyperscale self-build facilities draw on extensive land acquisitions, power infrastructure, and specialized equipment to create purpose-built AI and cloud computing campuses. Apple's $2 billion self-built Mesa data center in Arizona spans 1.3 million square feet, while Google's self-built Council Bluffs facility covers more than two million square feet and has a power capacity above 100 MW. Microsoft operates Azure across 70 regions with over 400 data centers globally, including its NDv5 cluster Eagle featuring 14,400 H100 GPUs installed in 2023 and expected to run for years.[7][8]
These self-built facilities power the cloud services, AI training, and digital platforms that businesses and consumers rely on daily. Microsoft's data center campus near San Antonio, Texas, Google's facilities in The Dalles, Oregon and Northern Virginia, and Meta's Kuna data centers in Idaho represent examples of hyperscaler-owned campuses that support their global service delivery. Most hyperscalers own and operate a dozen or more of their own data center facilities in locations around North America.[17]
Where it sits in the stack
Takes in: Capital, land, construction management, internal procurement
Sends on: Owned, occupied compute capacity
What would break it
Leasing from third-party operators offers hyperscalers an alternative to self-build, providing capacity on shorter timeframes but requiring them to give up control or compromise on design, operations, and ownership. While self-builds will continue in the long run, the majority of hyperscaler capacity is expected to be leased from third-party operators over the next 10 to 15 years. Hyperscalers' growth is outpacing in-house delivery capabilities, driving a rise in build-to-suit deals with data center providers, with 120 to 130 hyperscale data centers expected to come online each year over the next decade.[2][1][19]
Project delays and cancellations demonstrate the execution risks inherent in self-build strategies. Microsoft is freezing 1.5 GW of near-term self-build data center projects that were previously scheduled to go online in 2025 and 2026. The company is also canceling hundreds of megawatts of data center leases across the US, shifting more investment into self-built AI campuses, while approximately 18% of its cloud infrastructure relies on third-party colocation.[6][10]
New entrants from the cryptomining world have attempted to enter the market by leveraging existing interconnection agreements to plan large data centers away from population centers but with ample power. Core Scientific's 570 MW deal with CoreWeave exemplifies this approach, with Core Scientific guiding 250 MW to be delivered to CoreWeave by the end of 2025. However, due to a lack of credentials and the self-build alternative, AI-first developers largely struggled to secure hyperscaler customers.[4][9]
What to watch
Meta is building a 5-gigawatt AI campus in Louisiana called Hyperion, scheduled for completion by 2028, which will require three dedicated natural gas power plants with a combined capacity of around $3 billion each. Microsoft announced its $80 billion data center investment plan in January 2026, with more than half allocated to facilities inside the United States. US data center construction starts reached $77.7 billion in 2025, up 190% from the prior year, with more than 60 major projects expected to break ground in the first half of 2026.[15][14]
Hyperscaler self-build capacity across Europe is expected to reach 4.2 GW in 2026, representing 24% year-on-year growth compared to 2025. As of Q4 2025, approximately 60% of Europe's operational hyperscaler self-build capacity is located in Ireland, the Netherlands, Sweden, and Belgium. The EMEA development pipeline is approaching 15 GW according to Cushman & Wakefield.[20][23]
Google partnered with the electrical training ALLIANCE for 2025 to 2026 to upskill 100,000 electrical workers and train 30,000 new apprentices, increasing the workforce pipeline by 70% over five years. By 2030, developers are projected to build two to four times more capacity in the next five years than they did in the last 25, requiring an estimated 140,000 additional skilled tradespeople. In the United States, federal legislation was introduced in late July requiring new facilities to achieve LEED certification to be eligible for existing federal tax incentives.[21][24]
Related nodes
Sources
- lek.com · 2025-06-03T20:22:22
- streamdatacenters.com · 2026-02-27T19:21:50
- dcbyte.com · 2026-01-12T07:28:32
- newsletter.semianalysis.com · 2025-10-09T18:12:49
- irecruit.co · 2026-05-22T09:08:42
- newsletter.semianalysis.com · Apr 28, 2025
- datacenterdynamics.com
- newsletter.semianalysis.com
- newsletter.semianalysis.com · 3Q2024
- globaldatacenterhub.com · Mar 02, 2025
- irecruit.co
- hyperscalers.com
- structureresearch.net · 2023
- aitooldiscovery.com · 2026
- aitooldiscovery.com · January 2026
- lambdafin.com · Q1 2026
- gravel2gavel.com
- blackridgeresearch.com · 2026-05-13
- glp.com
- cbre.co.uk · February 19, 2026
- irecruit.co
- ibm.com
- irecruit.co
- canada.constructconnect.com · late July
Full scorecard, owner shares, supply edges and the full tracked roster are in the desk letter.
GET THE BRIEFING