Hyperscale Data Center Statistics 2026: Count, Capacity and Spend

Meta anticipates 2026 capital expenditures, including principal payments on finance leases, to be in the range of $130 billion to $145 billion, a gauge of the spending behind hyperscale data centers. Across US data centers, Lawrence Berkeley National Laboratory (LBNL) estimates 74% of servers were present in hyperscale and colocation space types in 2023, a proportion that will rise to 85% by 2028.

No public statistical agency publishes a count of hyperscale buildings, so the footprint comes from the operators themselves. AWS reports 124 Availability Zones within 39 Geographic Regions, Google Cloud lists 43 global regions and 130 zones, and Microsoft describes a global Microsoft Cloud of over 400 data centers in 70 regions. At Alphabet, purchases of property and equipment of $44,924 million in Q2 2026 exceeded net cash provided by operating activities of $39,069 million.

Key Takeaways

• Microsoft’s additions to property and equipment reached $115,948 million in the twelve months ended June 30, 2026, against $64,551 million a year earlier, an increase of 79.6%.
• Alphabet’s free cash flow was negative $5,855 million in Q2 2026 as its spending on property and equipment climbed.
• Amazon’s trailing twelve-month free cash flow was negative $7,604 million in Q2 2026.
• LBNL’s 2024 report estimates hyperscale facilities alone held 8% of servers in 2014 and 25% in 2020, a slower shift than its earlier report projected.
• JLL says hyperscalers account for 59% of 2026 data center tenant demand, with 28 GW of owner-occupied capacity currently under construction.
• Google reports a trailing twelve-month (TTM) PUE of 1.09 across all its large-scale data centers (once they reach stable operations), while the Uptime Institute’s 2025 Global Data Center Survey put the global average PUE of respondents’ data centers at 1.54.
• Epoch AI finds aggregate cash capex across Microsoft, Amazon, Alphabet, Meta and Oracle is on track to overtake operating cash flow around Q3 2026.

Editor’s Choice

• AWS in North America: 31 Availability Zones within 9 Geographic Regions.
• Oracle remaining performance obligations: up $209 billion year-over-year to $664 billion.
• AWS annualized revenue run rate: $169 billion, per Amazon.
• Meta capital expenditures, Q2 2026: $31.08 billion, including principal payments on finance leases.
• Hyperscale, AI-focused data center capacity: 100 MW or more, consuming as much electricity annually as 100,000 households, per the International Energy Agency (IEA).
• North America data center absorption, first half of 2026: 25 GW, double the level from a year ago, per JLL.

Hyperscale Data Center Spending by Company

• Alphabet’s purchases of property and equipment totaled $132,402 million over the trailing twelve months to Q2 2026.
• Amazon’s purchases of property and equipment were $173,028 million over the twelve months ended June 30, 2026, with proceeds from property and equipment sales and incentives of $4,021 million, a net $169,007 million, up 64% on Amazon’s trailing measure.
• Meta’s outlook was narrowed from a prior outlook of $125 billion to $145 billion, raising the bottom of the range.
• Microsoft’s additions to property and equipment were $35,802 million in the three months ended June 30, 2026, compared with $17,079 million a year earlier, about 2.1 times the prior-year quarter.
• Over the twelve months ended June 30, 2026, Microsoft’s additions to property and equipment were $115,948 million, against $64,551 million a year earlier, an increase of 79.6%.
• Epoch AI names Microsoft, Amazon, Alphabet, Meta, and Oracle as the world’s largest hyperscalers.
• Each company defines its spending line differently, so the figures are never added together.

Source: Alphabet, Amazon and Microsoft earnings releases (July 2026); Meta Q2 2026 results (July 2026)

About This Data

• 15 sources: 5 company earnings releases filed with the SEC or posted to investor relations sites, 5 company infrastructure pages and announcements, 2 reports from public research bodies (LBNL and the IEA), 2 Epoch AI data pages and 1 commercial real estate research release from JLL.
• Publication window: LBNL’s December 2024 report to October 2026.
• Selection: company filings and official pages, public research bodies and named research publishers only; paid market-research report sellers excluded.
• LBNL and IEA figures are over 12 months old and carry their data year.
• Figures are reviewed on a rolling basis and updated when sources publish new editions.

How much do hyperscalers spend on data centers?

Meta expects 2026 capital expenditures, including principal payments on finance leases, in the range of $130 billion to $145 billion. Microsoft’s additions to property and equipment were $115,948 million for the twelve months ended June 30, 2026. The quarterly trend sits in the wider data center capex and pipeline figures. That spending now runs into a limit set by the same companies’ cash.

Hyperscaler Capex vs Operating Cash Flow

• Alphabet’s purchases of property and equipment rose from $23,953 million in Q3 2025 to $44,924 million in Q2 2026, about 1.9 times the Q3 2025 level.
• In the quarter ended June 30, 2026, Alphabet’s net cash provided by operating activities was $39,069 million, up from $27,747 million, while purchases of property and equipment rose to $44,924 million from $22,446 million a year earlier.
• Alphabet’s free cash flow, which it defines as net cash provided by operating activities less capital expenditures, was negative $5,855 million in Q2 2026.
• Alphabet issued a combination of Class A stock and Class C stock and mandatory convertible preferred stock for aggregate net proceeds of $49.6 billion, including for capital expenditures to scale AI infrastructure and global compute, in June 2026.
• Alphabet also entered into an equity distribution agreement to sell up to $40.0 billion of its Class A stock and Class C stock.
• Oracle reported a record Q1 operating cash flow of $23 billion, up 184%, while free cash flow was negative $5 billion for the quarter.

Amazon shows the same crossing on a trailing basis. Its operating cash flow for the twelve months ended June 30, 2026 reached $161,403 million, up 33%, below net purchases of property and equipment of $169,007 million over the same period, which leaves trailing twelve-month free cash flow at negative $7,604 million.

Key finding: Epoch AI, summing SEC filings for Microsoft, Amazon, Alphabet, Meta and Oracle, finds aggregate cash capex across hyperscalers is on track to overtake operating cash flow around Q3 2026, and most hyperscalers have already turned to external financing to fund their growing investments in AI infrastructure. Borrowed and raised money now funds part of the build-out.

Spending is disclosed in detail each quarter; the number of buildings is not.

Recent Developments

• October 12, 2026: Epoch AI said the 94 AI data centers it covers have a combined 14.1 GW of IT power capacity, and it estimates its coverage of global deployed AI computing capacity at 44%.
• September 10, 2026: Oracle reported the delivery of 850 MW of additional data center capacity and said it booked more than $30 billion of additional AI cloud contracts in Q1.
• August 11, 2026: JLL reported that North America data center demand reached an all-time high in the first half of 2026, five times that of two years prior, with neoclouds at 11% of demand and pure-play AI companies at 7%.
• August 6, 2026: Microsoft announced the general availability of its India South Central cloud region in Hyderabad, which comprises three Availability Zones.
• July 30, 2026: Amazon reported that AWS segment sales increased 37% year-over-year to $42.2 billion.
• July 29, 2026: Microsoft said Azure revenue surpassed $100 billion for the first time in its fiscal year.

How Many Hyperscale Data Centers Are There

• AWS has announced plans for 7 more Availability Zones and 2 more AWS Regions in the Kingdom of Saudi Arabia and Chile.
• AWS defines a Region as a physical location where it clusters data centers, and calls each group of logical data centers an Availability Zone.
• Each AWS Region consists of a minimum of three isolated, physically separate AZs, according to AWS.
• Epoch AI’s database covers 94 AI data centers with 18.3 GW of total facility capacity, against 14.1 GW of IT power before cooling and other infrastructure overhead.
• Zones, regions, and data centers are different units, so the rows cannot be added up.

Source: AWS Global Infrastructure (September 2026); Google Cloud locations (October 2026); Microsoft (September 2025); Epoch AI (October 2026)

How many hyperscale data centers are there?

No public statistical agency publishes a count of hyperscale data centers. Operators report their own units instead. AWS clusters data centers into Regions and calls each group of logical data centers an Availability Zone, Google Cloud counts global regions and zones, and Microsoft counts the data centers and regions of the Microsoft Cloud, so the totals in the table above measure different things. The cloud revenue those footprints earn is growing at very different speeds.

Cloud Revenue Growth Behind the Build-Out

• Oracle reported Q1 Cloud Infra (IaaS) Revenue up 121% in USD to $7.4 billion for its quarter ended August 31, 2026.
• Oracle also reported remaining performance obligations up $209 billion year-over-year to $664 billion for the same quarter.
• Alphabet said Google Cloud revenues reached $24.8 billion, led by an increase in Google Cloud Platform (GCP) across enterprise AI Solutions and enterprise AI Infrastructure, in Q2 2026.
• Microsoft reported Microsoft Cloud revenue of $59.3 billion, up 27% year-over-year, and revenue in Intelligent Cloud of $39.3 billion, an increase of 32%.
• Amazon said AWS net sales growth was its fastest in 18 quarters.
• Amazon put AWS at a $169 billion annualized revenue run rate in its Q2 2026 results.
• These growth rates sit inside the wider cloud computing market figures.

Counting hyperscale buildings starts with agreeing what qualifies as one.

What Is a Hyperscale Data Center

• The IEA says a conventional data center may be around 10-25 megawatts (MW) in size.
• The IEA says a hyperscale, AI-focused data center can have a capacity of 100 MW or more, consuming as much electricity annually as 100,000 households.
• The average power consumption of the AI data centers Epoch AI tracks is typically 60-80% of capacity, due to idle time and maintenance.
• The IEA adds that AI-focused data centers are increasing in size to accommodate larger and larger models and growing demand for AI services.
• LBNL sizes hyperscale facilities at an average of 30,000 square feet per module, not per entire facility or campus.
• The IEA also finds that nearly half of data center capacity in the United States is in five regional clusters. The Northern Virginia data center market is covered separately.

Source: IEA, Energy and AI (2025); LBNL, 2024 United States Data Center Energy Usage Report

What is a hyperscale data center?

A hyperscale data center is a very large facility that a cloud or internet company builds and runs for its own services. The IEA puts a hyperscale, AI-focused data center at a capacity of 100 MW or more, against around 10-25 MW for a conventional data center. LBNL averages 30,000 square feet per hyperscale module.

How much power does a hyperscale data center use?

The IEA says a hyperscale, AI-focused data center can have a capacity of 100 MW or more, consuming as much electricity annually as 100,000 households, while Epoch AI says the AI data centers it tracks typically draw 60-80% of capacity, due to idle time and maintenance. Those large facilities now hold most of the country’s servers.

Hyperscale vs Colocation vs On-Premises Data Centers

• LBNL reports that in the early 2010s, internal data centers dominated the industry, with hyperscale and colocation providers housing less than 15% of the servers.
• Hyperscale and colocation together reached 74% of servers in 2023 and will rise to 85% by 2028, according to LBNL’s estimate.
• LBNL’s 2016 report estimated that in 2014, 24% of servers were present in hyperscale data centers, and that by 2020 this number would rise to 40%; current data shows a slower shift.
• LBNL’s 2024 report puts hyperscale facilities alone at 8% of servers in 2014 and 25% in 2020.
• LBNL’s “internal” space type is the on-premises data center.

By the numbers: LBNL estimates the hyperscale and colocation share of US servers moved from 20% in 2014 to 57% in 2020 and 74% in 2023. From 2014 to 2020 the hyperscale-only share roughly tripled, and by 2023 most of the server base had left enterprise rooms for facilities run by cloud companies and landlords.

What is the difference between hyperscale and colocation data centers?

A hyperscale data center is run by one cloud or internet company for its own workloads, while a colocation operator leases space and power to many tenants. The line blurs, as JLL says hyperscalers meet capacity needs through both leasing and self-building. LBNL sizes colocation space types at 11,000 square feet per module. The largest data center operators are covered separately.

Hyperscale Share of Data Center Server Energy

• LBNL estimates that by 2023, hyperscale and colocation data centers accounted for almost 80% of US server energy.
• LBNL expects hyperscale and colocation will account for over 90% of server energy consumption, primarily driven by AI workloads, by 2028.
• These shares are model estimates, not metered readings.

Source: LBNL, 2024 United States Data Center Energy Usage Report

The overall power picture sits in the data center electricity use figures. The tenants leasing new capacity show who is driving that growth.

Hyperscalers’ Share of Data Center Demand

• JLL says hyperscalers remain the dominant force, accounting for 59% of 2026 data center tenant demand.
• Hyperscalers are meeting capacity needs through both leasing and self-building, with 28 GW of owner-occupied capacity currently under construction, per JLL.
• JLL reported North America data center demand reached 25 GW of absorption in the first half of 2026, double the level from a year ago.
• JLL said the market was driven by hyperscalers, neoclouds and pure-play AI companies competing aggressively for scarce capacity.
• JLL’s shares cover North America and do not sum to 100.

Much of that owner-occupied pipeline appears in the data center construction figures. Size also changes how efficiently those facilities run.

Hyperscale Data Center Efficiency (PUE)

• Google reports a trailing twelve-month (TTM) PUE of 1.09 across all its large-scale data centers (once they reach stable operations).
• Google calculates its overhead energy use at 83% less than the industry average, using its 0.09 overhead against 0.54.
• LBNL finds that by 2023, the average PUE across hyperscale and colocation data centers is under 1.4, with around 75% of servers installed in those data centers.
• LBNL projects that by 2028, the average PUE will fall to between 1.15 and 1.35.
• Google notes that a PUE closer to 1.0 means nearly all the energy is used for computing.

Worth noting: LBNL attributes the fall in US average PUE from 1.6 in 2014 to 1.4 in 2023 primarily to the shift towards larger data centers (hyperscale, colocation) that have a lower PUE. PUE is self-reported, and Google’s figure covers only its stable large-scale sites, so it is a vendor claim and not an independent benchmark.

Efficiency also depends on how busy each server is.

Server Utilization by Data Center Type

• LBNL assumes hyperscale data centers average 45% in 2014, rising to 50% in 2027.
• In LBNL’s model, utilization for conventional servers in all space types increases slightly through 2028.
• Under those assumptions, a hyperscale server’s 2027 utilization is 2.5 times the 20% assumed for internal and small data centers.
• These are LBNL model assumptions for conventional servers, not measured averages.

Higher utilization does not cap total demand, which is still climbing.

Hyperscale Data Center Electricity Demand

• The IEA Base Case projects global electricity consumption by data centers to reach around 945 TWh by 2030, representing just under 3% of total global electricity consumption.
• The IEA says data center electricity consumption grows more than four times faster than the growth of total electricity consumption from all other sectors from 2024 to 2030.
• LBNL puts US data center energy use at 176 TWh by 2023, representing 4.4% of total electricity consumption.
• IEA figures are global; LBNL figures are US-only.

Source: IEA, Energy and AI (2025); LBNL, 2024 United States Data Center Energy Usage Report

Water use shifted the same way.

Hyperscale Data Center Water Use

• LBNL estimates that in 2014, data centers consumed 21.2 billion liters of water, with 64% in internal data centers.
• By 2023, hyperscale and colocation account for 84% of the 66-billion-liter total, while internal data centers fell to just 12%, according to LBNL.
• LBNL expects hyperscale data centers in 2028 to consume between 60 and 124 billion liters.
• LBNL projects internal data centers falling to just 2% of the total by 2028.
• The data center water use figures cover the wider picture.

The biggest single sites show where hyperscale is heading.

Largest Hyperscale AI Data Centers

• Epoch AI reports that the largest known AI data center by IT power is Colossus 2, owned by SpaceXAI (xAI) in Memphis, TN, drawing 946 MW.
• The largest AI data center by compute is Colossus 2, with 1,112,000 H100-equivalents, per Epoch AI.
• Epoch AI estimates Google Pryor (North) at 637,000 H100-equivalents.
• Epoch AI says total facility power is 20-50% higher due to cooling and other infrastructure overhead.
• Microsoft says it has multiple identical Fairwater data centers under construction in other locations across the US.

Three of those five sites belong to Amazon, Microsoft, and Meta, part of the AI data center build-out. The largest US data center sites add campuses outside AI.

Conclusion

Meta’s outlook of capital expenditures, including principal payments on finance leases, in the range of $130 billion to $145 billion and LBNL’s estimate that 74% of US servers were present in hyperscale and colocation space types in 2023 describe the same shift from two sides. The money is filed each quarter, while the buildings appear only in each operator’s own units.

The IEA Base Case projects global electricity consumption by data centers to reach around 945 TWh by 2030, and Epoch AI expects aggregate cash capex across hyperscalers to overtake operating cash flow around Q3 2026. The next round of quarterly filings will show whether that crossover holds for the group as a whole.