Data Center
Last reviewed
Jun 7, 2026
Sources
43 citations
Review status
Source-backed
Revision
v2 ยท 7,218 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
43 citations
Review status
Source-backed
Revision
v2 ยท 7,218 words
Add missing citations, update stale details, or suggest a clearer explanation.
A data center is a dedicated physical facility that houses computer systems, networking equipment, storage arrays, and the supporting infrastructure (power distribution, cooling, fire suppression, and security) required to run them at scale. Data centers form the operational backbone of the modern internet, cloud computing, and large-scale machine learning workloads.
In the 2020s, the data center has become the central piece of AI infrastructure. Training and serving large language models, image and video generators, and recommendation systems requires tens of thousands of accelerators wired together with high-bandwidth networks, fed by megawatts of electricity, and cooled with liquid because the underlying chips run too hot for traditional air systems. Building, financing, and powering these facilities is now one of the largest categories of capital investment in the global technology sector, with hyperscale operators committing hundreds of billions of dollars per year to new construction.
The term "data center" covers a wide range of facility types, from a small server closet inside an office to a multi-building hyperscale campus consuming over a gigawatt of power. What unifies them is the function: aggregating compute, storage, and network resources in an environment engineered to keep that hardware running continuously and reliably.
The lineage of the modern data center begins with the mainframe computer rooms of the 1940s and 1950s. The ENIAC, completed at the University of Pennsylvania in 1945, occupied an entire room and required forced-air cooling to keep its vacuum tubes operational. Through the 1950s, similar rooms were built at U.S. military and government installations, including West Point, the Pentagon, and CIA headquarters, to house early electronic computing systems for ballistics calculations and cryptanalysis.
The transition from vacuum tubes to transistors in the mid-1950s, and then to integrated circuits in the 1960s, made computing equipment smaller and cooler but did not eliminate the need for purpose-built rooms. Bell Labs developed the first transistorized computer, TRADIC, in 1954, and IBM introduced its first fully transistorized commercial machine in 1955. By the 1960s organizations were constructing dedicated "computer rooms" inside office buildings to house IBM System/360 and similar mainframes. Many of the conventions that still define data centers, such as raised floors for cabling and air distribution, equipment racks, and overhead cable trays, originated in this era. Operators learned, often the hard way, that uncontrolled humidity, electrostatic discharge, and dust were as dangerous to the equipment as outright power loss.
The enterprise data center took its modern form in the 1980s and 1990s. Falling prices for minicomputers and then commodity x86 servers led companies to deploy hundreds or thousands of machines in central facilities, often built inside corporate headquarters or at suburban campus sites. The growth of client-server architecture and the emergence of the World Wide Web after 1993 dramatically expanded demand for server capacity, and the first commercial colocation facilities, where multiple tenants leased space in a shared building, appeared in the late 1990s. Equinix, founded in 1998, became one of the most prominent operators in this segment, focusing on carrier-neutral interconnection points where networks could exchange traffic. The dot-com investment boom of 1999 to 2001 funded a wave of new colocation construction, much of it overbuilt; the survivors became the foundation of the 2000s internet.
The early 2000s brought the rise of internet hyperscalers. Google, Amazon, Microsoft, Yahoo, and Facebook all began constructing very large facilities engineered around their own software stacks. Google in particular pioneered the use of commodity hardware in highly redundant, software-managed clusters and published influential research on warehouse-scale computing, including Luiz Andre Barroso and Urs Holzle's widely cited 2009 monograph on the topic. Amazon launched Amazon Web Services in 2006, opening hyperscale data center capacity to outside customers and effectively creating the public cloud market. Facebook open-sourced its data center designs through the Open Compute Project in 2011, accelerating standardization of high-efficiency server, rack, and power shelf designs across the industry.
From roughly 2010 onward, cloud computing reshaped the industry. Workloads migrated off enterprise floors and onto facilities operated by AWS, Microsoft Azure, Google Cloud, and a handful of other providers. Hyperscale operators standardized on highly efficient designs with custom-built power distribution, evaporative or outside-air cooling, and software-defined networking. Average industry PUE fell from above 2.0 in the late 2000s toward roughly 1.55 by the late 2010s as best practices spread, although improvement at the industry-wide level has since plateaued. [2]
The most recent inflection point is the AI buildout that began accelerating in 2023. The release of ChatGPT in late 2022 triggered an industry-wide race to acquire GPUs and the buildings to put them in. By 2024 and 2025, hyperscalers and a wave of new "neocloud" specialists were committing to single-site campuses measured in gigawatts of power capacity, signing multi-decade nuclear power purchase agreements, and rebuilding the cooling and power distribution stack around densities an order of magnitude higher than traditional enterprise infrastructure. Capital expenditure plans from the four largest U.S. hyperscalers reached the high hundreds of billions of dollars per year, with much of the increase directed at AI-specific facilities. [21]
A modern data center is a stack of interdependent systems. The IT equipment, the racks that hold it, and the power and cooling that keep it running are all designed together.
| Component | Function | Typical examples |
|---|---|---|
| Server racks | Standard 19-inch or 21-inch enclosures that hold IT equipment | Open Compute Project Open Rack, traditional 42U/48U cabinets |
| Compute servers | CPUs and accelerators that run application workloads | Dual-socket x86 servers, GPU servers, ARM-based instances |
| Storage | Persistent data, both block and object | All-flash arrays, distributed object stores, tape libraries |
| Networking | Switches, routers, optical interconnects, structured cabling | Top-of-rack switches, spine-leaf fabrics, optical line systems |
| Power distribution | Brings utility power to racks, conditions and backs it up | Substations, switchgear, transformers, PDUs, busways |
| Backup power | Carries load through utility outages | Uninterruptible power supplies (UPS), battery banks, diesel or natural gas generators |
| Cooling | Removes heat from IT equipment | Computer room air handlers (CRAH), chilled water plants, cooling towers, liquid cooling distribution units |
| Fire suppression | Protects equipment without water damage | Inert gas systems (FM-200, Novec), pre-action sprinklers |
| Physical security | Controls access to the facility | Mantrap entries, biometric controls, perimeter fencing, on-site guards |
| Building management | Monitors and controls all systems centrally | DCIM software, BMS controllers, environmental sensors |
A single rack can hold anywhere from a handful of dense GPU servers to dozens of one-rack-unit compute or storage nodes. Racks are organized into rows separated by hot and cold aisles, with cold supply air drawn into the front of the equipment and hot exhaust air captured at the back for return to the cooling plant. In modern facilities, hot or cold aisles are often physically contained with curtains or roof panels to prevent the two air streams from mixing, which significantly improves cooling efficiency.
The IT load itself is typically organized in tiers. Compute servers handle application logic; storage nodes hold persistent data, increasingly on flash media; and a layer of network equipment, beginning with top-of-rack switches and aggregating into spine and core layers, ties everything together. Storage capacity per facility is often measured in petabytes for cloud providers and exabytes for the largest hyperscale operators. Networking inside a hyperscale facility frequently uses Clos or fat-tree topologies that provide non-blocking connectivity between any two servers, while large AI clusters often layer a separate, dedicated GPU-to-GPU fabric on top of the general-purpose network.
The most widely cited framework for data center reliability is the Uptime Institute's Tier Classification System, which rates facilities from Tier I (basic) to Tier IV (fault tolerant) based on the redundancy of their power and cooling infrastructure and their ability to be maintained without taking IT load offline. [1]
| Tier | Description | Annual uptime | Maximum annual downtime | Redundancy |
|---|---|---|---|---|
| Tier I | Basic capacity, single non-redundant distribution path | 99.671% | ~28.8 hours | None (N) |
| Tier II | Redundant capacity components for power and cooling | 99.741% | ~22 hours | N+1 capacity components |
| Tier III | Concurrently maintainable, multiple distribution paths with one active | 99.982% | ~1.6 hours | N+1, dual-powered equipment |
| Tier IV | Fault tolerant, multiple active distribution paths | 99.995% | ~26.3 minutes | 2N or 2N+1, fully compartmentalized |
Tier ratings apply to the design of the facility, not to its operations; the Uptime Institute also offers separate certifications for constructed facilities and for ongoing operational sustainability. Many hyperscale operators do not pursue formal Tier certification because their distributed software architecture allows individual sites to fail without affecting application availability, but the framework remains the dominant industry vocabulary for describing physical reliability.
Power and cooling consume the majority of the engineering effort that goes into a data center. The two are tightly linked: every watt delivered to an IT load must be removed as heat, and the efficiency of that removal directly affects the facility's total electricity consumption.
The industry's standard efficiency metric is Power Usage Effectiveness (PUE), defined as the total facility energy divided by the IT equipment energy. A PUE of 1.0 would mean every watt entering the building is used by computing equipment, with zero overhead for cooling, lighting, or power conversion. Real facilities cannot reach this ideal.
| PUE level | Typical operator | Notes |
|---|---|---|
| ~1.55-1.58 | Industry average | Uptime Institute Global Data Center Survey 2023-2024; the figure has been roughly flat for five years |
| 1.20 or below | Best-in-class new builds | Common target for new hyperscale construction |
| ~1.16 | Microsoft global fleet (2024) | Reported in Microsoft sustainability disclosures |
| ~1.09-1.10 | Google and Meta global fleets | Among the lowest reported by any major operator |
The Uptime Institute's 2025 Global Data Center Survey reported an industry-average PUE of 1.54, essentially unchanged for a sixth consecutive year, even as hyperscale operators (Google, Meta, Microsoft, and Amazon) continued to report fleet figures in the 1.10 to 1.15 range and colocation and enterprise facilities averaged higher, between roughly 1.58 and 1.80. [22] The same survey found that adoption of direct liquid cooling, while rising sharply at AI sites, remained a minority practice across the broader industry: about 22 percent of operators reported using direct liquid cooling against 75 percent still relying on perimeter air cooling, with high rack densities cited as the main driver of liquid-cooling deployment. [23] [18]
For most of the industry's history, data centers have relied on air cooling: chilled supply air pushed through perforated floor tiles or overhead ducts, drawn through servers by their internal fans, and returned to a chilled water or direct expansion plant. This works well up to roughly 20 to 30 kilowatts per rack, beyond which the volume of air required becomes impractical.
Direct-to-chip liquid cooling circulates a coolant through cold plates attached directly to CPUs and accelerators, removing heat far more efficiently than air. Immersion cooling submerges entire servers in a dielectric fluid; a single-phase variant uses a non-boiling oil, while two-phase systems use fluids that vaporize on the chip surface and condense overhead.
The rise of high-density AI accelerators has forced a shift away from air. The thermal design power of an NVIDIA H100 GPU is 700 W, and the NVIDIA Blackwell generation pushes individual packages above 1,000 W. Heat flux at the chip surface for GB200 systems reaches roughly 500 to 600 watts per square centimeter, a level that cannot be removed by air at any practical flow rate.
A traditional enterprise server rack draws roughly 5 to 15 kW. AI training racks operate at far higher densities. The NVIDIA GB200 NVL72, which connects 36 Grace CPUs and 72 Blackwell GPUs in a single liquid-cooled rack, draws approximately 120 kW continuously, supplied through four 30 kW power shelves at 480 V three-phase input. [14] [15] Industry roadmaps describe future racks reaching 250 kW and beyond as new generations of accelerators arrive. This shift is rewriting the assumptions behind data center electrical design, busway sizing, and floor loading.
NVIDIA's next-generation Vera Rubin platform, announced for volume shipment in the second half of 2026, extends this trajectory. The Vera Rubin NVL72 rack unifies 72 Rubin GPUs and 36 Vera CPUs with ConnectX-9 SuperNICs and BlueField-4 DPUs, requires 100 percent liquid cooling with no air-cooled configuration offered, and is positioned by NVIDIA to deliver roughly ten times more tokens per megawatt than the GB200 NVL72 for reasoning workloads. [24] To support the higher per-rack power these systems require, NVIDIA, Siemens, and Fluence published a joint reference electrical and power architecture in 2026 that pairs on-site battery storage with the grid feed to smooth the large, rapid load swings characteristic of synchronized training jobs. [25]
The transition has practical consequences across the facility. Floor loading, historically specified at perhaps 150 to 250 pounds per square foot, must increase substantially to support liquid-cooled racks that may weigh several thousand pounds when fully loaded with coolant. Power distribution that once relied on overhead bus bars rated at 400 amps per rack must now handle several times that. Cooling distribution units (CDUs) and secondary cooling loops, previously confined to specialized HPC sites, are becoming standard elements in mainstream commercial colocation facilities. Some operators have begun engineering entire new building shells around 100 to 250 kW rack densities, with structural and mechanical systems sized accordingly from day one.
"AI data center" has become a distinct category, characterized by extreme rack power density, liquid cooling, ultra-high-bandwidth east-west networking between accelerators, and physical layouts optimized for very large training jobs that span an entire building or campus.
A handful of single-site projects illustrate the new scale of AI infrastructure. xAI's Colossus supercomputer in Memphis, Tennessee, came online in July 2024 with approximately 100,000 NVIDIA H100 GPUs, built out in 122 days. [9] xAI announced an expansion to roughly 200,000 GPUs within months and has continued adding capacity, including H200 and GB200 systems, while pursuing a second site (Colossus 2) in the Memphis area.
Meta announced the Hyperion data center in Richland Parish, Louisiana, in December 2024, describing it as the company's largest facility to date. [10] [11] The roughly 4 million square foot project is planned to scale to approximately 5 GW of capacity over time, with construction continuing through 2030. Meta has agreed with Entergy Louisiana to fund new generation and transmission infrastructure to support the load.
The Stargate initiative, a partnership between OpenAI, Oracle, and SoftBank, has its first site in Abilene, Texas. The Abilene campus is designed to house roughly 450,000 NVIDIA GB200 accelerators across eight buildings, with the first two buildings completed in 2025. [12] [13] Power capacity at Abilene is being capped at approximately 1.2 GW after grid interconnection delays prompted the partners to scale back further on-site expansion.
Microsoft operates a large network of AI-focused campuses for OpenAI workloads in locations including Mt. Pleasant, Wisconsin and several sites in Arizona and Texas. Google's largest AI sites host pods of TPU accelerators wired together with custom optical circuit switches.
AI training facilities are designed around the dominant accelerator generation of the moment. The Hopper-generation NVIDIA H100 and H200 GPUs powered most large training runs from 2023 through 2024. Blackwell-generation B200 and the rack-scale GB200 NVL72 began deploying in volume in 2025. Google's TPU line, including TPU v5p (8,960 chips per pod, ~460 petaFLOPS) [17] and the sixth-generation Trillium, [16] occupies the bulk of Google's internal AI capacity. Amazon Web Services has deployed Trainium and Trainium2 silicon at scale in its own facilities. Specialized vendors, including Cerebras with its wafer-scale engine and Groq with its inference-optimized LPUs, occupy smaller but growing footprints.
Large AI clusters depend on extremely high east-west bandwidth between accelerators. Within a single rack, NVIDIA NVLink and NVSwitch provide all-to-all GPU connectivity at up to 1.8 TB/s of bidirectional bandwidth per GPU in current generations. Between racks, two main fabrics dominate: InfiniBand, with current-generation NDR running at 400 Gb/s per port and the upcoming XDR generation at 800 Gb/s; and high-performance Ethernet, led by NVIDIA Spectrum-X, which adapts congestion control and load balancing techniques from InfiniBand. Both rely on Remote Direct Memory Access (RDMA) to move data between GPUs without involving the host CPU.
A further development in 2025 was the extension of these fabrics across multiple buildings and even multiple sites. Microsoft's "AI superfactory," which links its Fairwater campuses in Mount Pleasant, Wisconsin and Atlanta, Georgia, uses a dedicated AI wide-area network (AI WAN) built on roughly 120,000 miles of new fiber to let geographically separated buildings operate as a single training system, with the two-story Atlanta site using closed-loop liquid cooling that consumes almost no water in operation. [26] [27]
A frontier model training run today consumes power on the order of hundreds of megawatts continuously over weeks. Colossus's initial 100,000-H100 cluster required roughly 150 MW of facility power. The Stargate Abilene campus is being built out toward 1.2 GW. Hyperion is targeted at 5 GW. By 2026, multiple hyperscalers were openly discussing single sites in the 5 to 10 GW range, scales without precedent in commercial computing infrastructure.
Because of these scales, power availability has displaced land and fiber connectivity as the binding constraint on data center site selection. Operators now choose locations primarily based on the time required to interconnect to the grid and the availability of firm electricity supply. This has driven new construction toward regions with surplus generation, including parts of Texas, Wyoming, North Dakota, and rural Louisiana. Historically dominant hubs such as Northern Virginia, Dublin, Singapore, and Amsterdam have seen new restrictions or moratoriums imposed by local utilities and governments unable to extend transmission and generation fast enough to meet demand. The grid-interconnection queue (the time required between requesting a substation tie-in and receiving energization) is now commonly five to seven years in major U.S. markets, a timeline that distorts site choice for projects that need power within two to three years to remain competitive. [28]
A small number of companies own most of the world's hyperscale data center footprint. The cloud hyperscalers dominate by owned capacity, while a tier of specialized colocation providers leases shell-and-power to large tenants. A new wave of "neoclouds" has emerged since 2023, focused exclusively on renting out GPU capacity for AI workloads.
| Operator | Type | Focus |
|---|---|---|
| Amazon Web Services | Hyperscale cloud | Largest global cloud footprint, broad workload mix plus AI |
| Microsoft Azure | Hyperscale cloud | Tightly integrated with OpenAI workloads |
| Google Cloud | Hyperscale cloud | Internal Google services plus external cloud, heavy TPU presence |
| Meta | Hyperscale (private) | Recommendation systems, ranking, generative AI |
| Oracle | Hyperscale cloud | Major Stargate partner, large GPU buildouts for OpenAI |
| Equinix | Colocation and interconnection | Carrier-neutral interconnection across major metros |
| Digital Realty | Colocation and wholesale | Large global footprint, hyperscale and enterprise customers |
| CoreWeave | Neocloud | GPU-as-a-service, multi-hundred-megawatt footprint |
| Lambda | Neocloud | GPU-focused cloud and on-prem clusters |
| Crusoe | Neocloud | AI-optimized greenfield campuses, partner on Stargate Abilene |
| Nebius | Neocloud | European-headquartered, expanding GPU capacity |
CoreWeave reported approximately 590 MW of connected power as of the third quarter of 2025 and has stated targets in the 800 MW to 1 GW range by the end of 2026. Lambda Labs has signed leases for over 320 MW of capacity. Crusoe is the build partner for the Stargate Abilene campus and has announced a multi-gigawatt pipeline. Oracle Cloud Infrastructure is the contracted compute provider for Stargate Abilene and has committed to 4.5 GW of capacity for OpenAI across multiple sites.
By the end of 2025, CoreWeave had grown well beyond those earlier figures, reporting more than 850 MW of active power across 43 data centers, about 3.1 GW of total contracted power, and a revenue backlog of $66.8 billion, while guiding to more than 1.7 GW of active power by the end of 2026 and full-year 2025 revenue above $5.1 billion. [29] The company set 2026 capital expenditure guidance of $30 billion to $35 billion, more than double its 2025 outlay, and described a target of more than 8 GW of data center capacity by 2030. [29]
Global data center capacity is highly concentrated in a small number of metropolitan regions, often clustering near fiber crossroads, low-cost power, or favorable tax regimes.
Northern Virginia, anchored by Loudoun County and the town of Ashburn, is the world's largest single concentration. The region has roughly 5,000 MW of operating capacity (more than twice that of the next-largest hub) and accounts for a significant share of all internet traffic exchanged in North America. [19] The cluster grew from the original MAE-East internet exchange and has expanded continuously since the late 1990s.
Other major U.S. clusters include the Phoenix metro (which trades cheap land and reliable grid power against extreme summer heat that complicates evaporative cooling), the Dallas-Fort Worth area, Silicon Valley, the Pacific Northwest along the Columbia River (favored for cheap hydroelectric power), and Atlanta.
In Europe, the FLAP markets (Frankfurt, London, Amsterdam, Paris) historically anchored the industry, with Dublin emerging as a major Irish hub thanks to its tax regime and the early presence of U.S. hyperscalers. Iceland and the Nordic countries have grown as destinations for cooling-intensive workloads given their cold climates and abundant geothermal and hydroelectric power.
In Asia, Singapore is the dominant Southeast Asian hub, though it has imposed moratoriums on new construction at various points to manage power and land. Tokyo, Hong Kong, and Sydney serve regional markets. China hosts large clusters in Beijing, Shanghai, and the Guizhou province, where the central government has steered new construction inland for power and strategic reasons.
The AI buildout has pushed new capacity into rural locations primarily for power reasons. Wyoming, the Texas Permian Basin, North Dakota, and rural Louisiana have all seen new gigawatt-scale projects announced because they offer surplus generation, available land, and faster grid interconnection than constrained metropolitan markets.
The two years from 2025 to 2026 saw the largest sustained surge of data center construction and capital commitment in the industry's history, driven almost entirely by demand for AI training and inference capacity.
Combined capital expenditure by the four largest U.S. hyperscalers (Amazon, Microsoft, Alphabet, and Meta) reached roughly $388 billion to $410 billion in 2025 and was guided sharply higher for 2026, with most estimates placing the four companies' combined 2026 plans in the $600 billion to $700 billion range, on the order of a 60 percent to 77 percent year-over-year increase. [21] [30] Industry analysts estimated that around three-quarters of this spending was directed at AI-related infrastructure. [21] Microsoft alone reported $37.5 billion of capital expenditure and finance leases in its fiscal second quarter ending December 2025, up about 66 percent year over year, and chief executive Satya Nadella said the company had added nearly one gigawatt of capacity in that quarter while remaining capacity-constrained against demand. [31]
The buildout was concentrated in a small number of very large campuses, several of which crossed the gigawatt threshold for the first time:
| Campus | Operator | Location | Reported scale and status (2025-2026) |
|---|---|---|---|
| Colossus / Colossus 2 | xAI | Memphis, Tennessee | Expanded toward about 2 GW of site capacity with roughly 555,000 NVIDIA GPUs after xAI acquired a third building in late December 2025; powered partly by on-site gas turbines and Tesla Megapack storage [32] [33] |
| Prometheus | Meta | New Albany, Ohio | Described by Meta as among the first roughly 1 GW AI data centers, targeted to come online in 2026 [34] |
| Hyperion | Meta | Richland Parish, Louisiana | Planned to scale toward about 5 GW of IT capacity, with operations expected around 2030 [11] [34] |
| Stargate Abilene | OpenAI / Oracle / Crusoe | Abilene, Texas | First two of eight buildings operational from September 2025; designed for about 450,000 GB200 GPUs at roughly 1.2 GW [13] [35] |
| Project Rainier | AWS / Anthropic | New Carlisle, Indiana | $11 billion campus brought online in October 2025 with about 500,000 Trainium2 chips; planned to scale to roughly 2.2 GW [36] [37] |
| Fairwater | Microsoft | Mount Pleasant, Wisconsin and Atlanta, Georgia | Two-story Blackwell sites (GB200/GB300) linked into a multi-state "AI superfactory" over a dedicated AI WAN [26] [27] |
xAI's Colossus expansion was among the most aggressive: after building its first 100,000-GPU cluster in 2024, the company acquired a third Memphis building on December 30, 2025 and pushed toward roughly 2 GW of total site capacity and about 555,000 GPUs, backed by an approximately $18 billion silicon investment, with much of the power supplied by on-site natural-gas turbines (including a fleet operated with Solaris Energy Infrastructure) and Tesla battery storage. [32] [33]
The Stargate initiative expanded well beyond its Abilene flagship. In September 2025, OpenAI, Oracle, and SoftBank announced five additional U.S. sites (in Shackelford County and Milam County, Texas; Dona Ana County, New Mexico; Lordstown, Ohio; and an unnamed Midwest location), which the partners said brought Stargate's planned capacity to nearly 7 GW against an overall target of about 10 GW and roughly $500 billion of investment. [12] Additional Stargate sites were announced with Vantage Data Centers in Wisconsin and with Related Digital in Saline Township, Michigan, each described as providing more than a gigawatt of capacity. [12] In March 2026, however, the partners scrapped a planned 600 MW expansion at Abilene that would have taken the flagship site to roughly 2 GW; reporting attributed the decision to financing disputes, power-infrastructure timing of about a year, and a preference to deploy NVIDIA's newer Vera Rubin generation at new sites rather than add Blackwell capacity that would be a generation behind by the time power was available, while the broader 4.5 GW Oracle-OpenAI agreement remained on track. [35]
Anthropic and Amazon deepened their own infrastructure tie-up: the $11 billion Project Rainier campus in New Carlisle, Indiana came online in October 2025 with roughly 500,000 AWS Trainium Trainium2 chips (described by AWS as the world's largest cluster of non-NVIDIA AI chips), and in April 2026 the two companies announced an expansion to as much as 5 GW of compute for training and serving Claude, with Anthropic committing more than $100 billion to AWS over ten years and Amazon investing a further $5 billion (with an option for $20 billion more) in Anthropic. [36] [37] [38]
By 2026 the binding constraint on the buildout was not capital but physical power and equipment supply. Of roughly 12 GW of new U.S. data center capacity expected to come online in 2026, only about a third was under active construction, and analysts estimated that between 30 percent and 50 percent of large facilities scheduled to open during the year would be delayed or canceled, with a substantial block of announced capacity showing no signs of construction. [39] Long-lead electrical equipment was a central bottleneck: U.S. lead times for large power transformers stretched to roughly four years, and switchgear was reported sold out into 2028. [39] [40] On the generation side, GE Vernova's gas-turbine backlog reached about 100 GW in the first quarter of 2026 (up from roughly 83 GW at the end of 2025 and headed toward a projected 110 GW by year-end), with delivery slots for new heavy-duty turbines from the major suppliers running about five years out. [41] Morgan Stanley analysts projected a U.S. power shortfall on the order of 49 GW by 2028 as individual data center sites scaled to between 1 GW and 4 GW each. [39]
Data centers consumed approximately 415 TWh of electricity in 2024, or about 1.5 percent of global electricity consumption, according to the International Energy Agency. [3] The IEA's Electricity 2024 report projected that data center electricity use could roughly double by 2026, reaching between 650 and 1,050 TWh annually. [3] The agency's later analysis projected continued growth at roughly 15 percent per year through the end of the decade, with global data center demand approaching 945 TWh by 2030 in its base case (just under 3 percent of total global electricity). [4]
Subsequent IEA analysis sharpened these figures, projecting global data center electricity consumption rising from about 485 TWh in 2025 to roughly 950 TWh in 2030 and around 1,200 TWh by 2035 in the base case, with electricity use by AI-focused accelerated servers growing about 30 percent per year, far faster than the roughly 9 percent annual growth for conventional servers. [42] In the United States specifically, studies from Lawrence Berkeley National Laboratory and the Electric Power Research Institute projected that data centers could consume on the order of 8 percent to 12 percent of national electricity by 2028 to 2030, up from roughly 4 percent in 2023, although the range across published forecasts is wide because of uncertainty about AI adoption, hardware efficiency, and how many announced projects are actually built. [43]
The United States accounted for roughly 45 percent of global data center electricity use in 2024, followed by China at 25 percent and Europe at 15 percent. [4] AI workloads are the fastest-growing component within this total, and they concentrate demand in a smaller number of physical locations than general cloud workloads, creating localized grid stress.
Water consumption is a related concern. Many large facilities use evaporative cooling towers that consume water directly through evaporation. The Green Grid's Water Usage Effectiveness (WUE) metric measures liters of water consumed per kilowatt-hour of IT load. [20] Industry averages are around 1.8 to 1.9 L/kWh, while the most efficient operators (Amazon at roughly 0.19 L/kWh and Microsoft at roughly 0.30 L/kWh) report dramatically lower figures by using designs that emphasize air-side economization or closed-loop cooling. [20] The newest AI campuses have pushed this further: Microsoft's Fairwater Atlanta site uses a closed-loop liquid cooling system whose initial fill is comparable to the annual water use of about 20 homes and which then consumes almost no water in operation. [27]
The magnitude of new AI loads has driven hyperscalers toward direct procurement of firm, low-carbon electricity, including a wave of nuclear power purchase agreements:
| Buyer | Seller | Asset | Capacity | Announced | Target online |
|---|---|---|---|---|---|
| Microsoft | Constellation Energy | Three Mile Island Unit 1 (renamed Crane Clean Energy Center) | ~835 MW | September 2024 | 2028 |
| Amazon Web Services | Talen Energy | Cumulus campus, served by Susquehanna nuclear plant | Initial 960 MW campus, expanded to 1,920 MW supply agreement | March 2024 (acquisition); June 2025 (expanded supply) | Phased ramp from 2025 |
| Kairos Power | Multiple small modular reactors (with TVA) | Up to 500 MW from up to seven SMRs | October 2024 | First unit ~2030, full deployment by 2035 |
The Microsoft agreement with Constellation Energy restarts Three Mile Island Unit 1 as the Crane Clean Energy Center, supplying about 835 MW of carbon-free power under a 20-year agreement, with restart targeted for 2028. [5] Amazon's relationship with Talen Energy began with the March 2024 purchase of the Cumulus data center campus adjacent to the Susquehanna nuclear plant and was expanded in June 2025 into a supply agreement scaling toward roughly 1,920 MW. [6] [7] Google's October 2024 agreement with Kairos Power covers up to 500 MW from a fleet of small modular reactors, with the first unit expected around 2030. [8]
In parallel, hyperscalers continue to sign large renewable power purchase agreements (solar, wind, and grid-scale storage) and to fund grid upgrades. The combination of nuclear baseload, renewables plus storage, and (in some markets) new natural gas generation is reshaping electricity planning across the United States. Meta's Hyperion campus, for example, is being supported by an agreement with Entergy Louisiana to fund seven new natural-gas generating units totaling more than 5.2 GW alongside roughly 240 miles of new 500 kV transmission lines and battery storage installations.
Carbon accounting at this scale is contentious. Large operators historically claimed "100 percent renewable" status by purchasing renewable energy credits annually matched to their consumption, even when the underlying grid hours were served by fossil generation. Google and Microsoft have both committed to a stricter "24/7 carbon-free energy" framework that requires hourly matching of clean generation to consumption at each grid location. Achieving this for a load that runs around the clock at hundreds of megawatts is significantly harder than annual matching, and is the proximate driver of hyperscaler interest in firm low-carbon resources such as nuclear, geothermal, and long-duration storage.
An edge data center is a small facility located close to end users or to data sources, designed to reduce the round-trip latency that would be incurred by sending traffic back to a centralized cloud region. Edge sites typically range from single shipping-container deployments at the base of cell towers to small multi-megawatt facilities at metropolitan exchange points.
Edge sites are functionally distinct from hyperscale facilities. They prioritize proximity over raw capacity. A regional edge site might host on the order of 100 to 200 racks; a metro-edge or micro-edge site might host only a handful. Their workloads include content delivery network (CDN) caches, streaming media transcoding, real-time gaming infrastructure, network functions for 5G mobile networks, and AI inference for latency-sensitive applications such as autonomous vehicles, industrial automation, and augmented reality.
Edge sites have become more important as AI inference workloads grow. While model training generally remains centralized in hyperscale campuses, inference for many consumer-facing applications benefits from being run physically closer to users. Edge GPU racks deployed for inference often run at higher densities than traditional edge workloads, on the order of 10 to 15 kW per rack.
Power and cooling at edge sites are simpler than at hyperscale facilities but face their own constraints. A telecom-tower-mounted micro data center may have only a few kilowatts of available utility power and must reject heat into outdoor air without the benefit of a chilled water plant. Modular prefabricated designs (often shipped as fully integrated containers) have become standard for rapid edge deployment. The category has been growing alongside 5G mobile network rollouts, which require small distributed compute footprints to deliver the network slicing and ultra-low-latency services the standard envisions.
A data center's value depends on its ability to keep IT load running through equipment failures, utility outages, and physical threats. Industry practice combines redundant infrastructure with layered physical security.
Redundancy is described in shorthand as N, N+1, or 2N. "N" denotes the minimum capacity required to carry the IT load. "N+1" adds one additional unit of capacity per system, so a single failure can be absorbed without losing load. "2N" duplicates the entire system, including independent distribution paths, so that an entire failure domain can go offline without affecting operations. Tier III data centers are typically built to N+1; Tier IV facilities are typically 2N or 2N+1, with full physical separation of redundant paths.
Backup power is provided in two stages. An uninterruptible power supply (UPS), normally based on lithium-ion or lead-acid battery banks, carries the load during the seconds-to-minutes window required for backup generators to start. Diesel or natural gas generators then provide long-duration standby power for as long as fuel can be supplied. Large facilities typically maintain on-site fuel for 24 to 72 hours of full-load operation, with contracted resupply for extended outages. Newer designs increasingly substitute battery energy storage systems for some generator capacity. Some of the newest AI campuses go further: Microsoft's two-story Fairwater Atlanta site was reported to omit traditional uninterruptible power supplies and standby generators entirely, relying instead on grid power supplemented by battery storage and the operator's distributed software resilience. [26]
Physical security typically follows a layered "onion" model: perimeter fencing and vehicle barriers; a controlled site entry with credential checks; a single-tenant access vestibule, often a mantrap with biometric verification; cabinet-level locks for individual customer cages; and continuous video surveillance and on-site security personnel. SOC 2, ISO 27001, and (for federal customers) FedRAMP audits are common compliance benchmarks for commercial operators. Higher-security facilities serving government workloads or regulated financial customers add further controls such as TEMPEST-shielded rooms, explosive ordnance-trained dogs, and secure compartmentalized information facility (SCIF) construction.
Reliability practices extend beyond hardware. Concurrent maintainability, the property that any single component can be taken offline for service without affecting IT load, is the central operational requirement of Tier III and Tier IV designs and shapes how facilities are run day to day. Maintenance windows for chillers, generators, and electrical gear are scheduled months in advance and executed against detailed methods of procedure (MOPs), with full operational rehearsals for high-risk activities. The Uptime Institute's Management and Operations (M&O) Stamp of Approval certifies operational discipline as a separate qualification from physical design tier, recognizing that even a well-designed facility can fail if procedures and staffing are inadequate.
The table below summarizes a selection of the largest and most-publicized AI-focused facilities and campuses as of early 2026. Reported figures for AI sites change frequently as build-out continues; targets shown are public commitments by the operator.
| Site | Operator | Location | Type | Approximate scale |
|---|---|---|---|---|
| Colossus | xAI | Memphis, Tennessee | Single-coherent training cluster | ~100,000 H100 GPUs at launch (July 2024); expanded to 200,000+ GPUs including H200 and GB200; ~150 MW initial facility load |
| Colossus 2 | xAI | Memphis, Tennessee | Hyperscale AI campus | Pushing toward ~2 GW site capacity and ~555,000 GPUs after a third building was acquired in December 2025 [32] [33] |
| Hyperion | Meta | Richland Parish, Louisiana | Hyperscale AI campus | ~4 million sq ft planned, scaling toward ~5 GW by ~2030 |
| Prometheus | Meta | New Albany, Ohio | Hyperscale AI campus | Among the first ~1 GW AI data centers, targeted online in 2026 [34] |
| Stargate Abilene | OpenAI / Oracle / Crusoe | Abilene, Texas | Hyperscale AI campus | ~450,000 GB200 GPUs across 8 buildings; capped at ~1.2 GW |
| Project Rainier | AWS / Anthropic | New Carlisle, Indiana | Hyperscale AI campus | ~500,000 Trainium2 chips online October 2025; scaling toward ~2.2 GW [36] |
| Fairwater | Microsoft | Wisconsin and Atlanta | Multi-site AI superfactory | Two-story GB200/GB300 sites linked over a dedicated AI WAN [26] |
| Susquehanna / Cumulus | AWS / Talen Energy | Salem Township, Pennsylvania | Nuclear-adjacent hyperscale campus | 960 MW initial campus, expanded supply agreement to ~1,920 MW |
| Crane Clean Energy Center (PPA) | Microsoft / Constellation | Pennsylvania | Nuclear-restart PPA for Microsoft data centers | ~835 MW (target restart 2028) |
| Mt. Pleasant | Microsoft | Wisconsin | Hyperscale AI campus | Multi-billion-dollar buildout supporting OpenAI workloads |