1. Introduction

Every answer a chatbot gives arrives from a building. That building holds racks of computers, the power feed to run them, and the cooling to carry away the heat they produce. The industry calls it a data center, and the three sections that follow describe how large these buildings have become, where operators put them, and what work the machines inside actually do. Capacity gets measured in power rather than in floor space or machine count, because power is the binding constraint: a site can only hold as much computing as its electric supply and its cooling can sustain. The unit is the watt. A megawatt (MW) is one million watts. A gigawatt (GW) is one thousand megawatts, which is roughly the output of one large nuclear reactor. A 100 MW campus draws about what 80,000 American homes draw on average. The three sections draw on DataCenters_8.3_Capacity_Evolution.pdf, DataCenters_8.4_Geographic_Distribution.pdf, and DataCenters_8.5_Workload_Distribution.pdf. Treat the percentages as estimates. Sources disagree, and AI construction moves the figures faster than any published table can track.

2. Capacity Evolution

Sixty years separate a corporate machine room from an AI campus, and the power draw over that span rose by a factor of ten thousand. Each era in the table below added a workload the previous generation of buildings could not hold.

Era	Dominant driver	Typical facility	Scale
1960s-1980s	Mainframe computing	Corporate or government machine room	10-100 kW; one to a few machines
1990s	Client/server; early web hosting	Purpose-built server room	100 kW-1 MW
2000s	Dot-com boom; search engines; early SaaS	Colocation and early hyperscale	1-10 MW per facility
2010s	Cloud computing; mobile; video streaming	AWS, Azure, GCP mega-campuses	10-200 MW per campus
2020-2024	Pandemic-driven cloud migration; early AI	Hyperscale campuses and GPU clusters	100-500 MW per campus
2025+	Large-scale AI training and inference	AI-optimised gigawatt campuses	500 MW-1+ GW per campus

Global installed capacity reached roughly 60 to 70 GW of IT load by 2024, having doubled about every four years through the cloud era. The AI build-out that began in 2022 and 2023 is compressing that interval. Analysts project more than 200 GW worldwide by 2030 if the announced construction proceeds on schedule, which would triple the 2024 figure in six years. Water follows power. A 100 MW facility can consume 1 to 3 million gallons per day for evaporative cooling, which is roughly what a town of 10,000 to 30,000 people uses. In water-scarce regions that draw forces the issue, and operators turn to dry cooling or to liquid immersion, which submerges the hardware in a fluid that carries heat away without evaporating.

3. Geographic Distribution

Operators site these buildings where several conditions align: reliable and affordable power, fibre connections with low delay, a technical workforce, political stability, tax incentives, and a cool climate that cuts the cooling bill. The shares below approximate global installed IT capacity as of 2025.

Region	Share	Key concentrations
United States	~35%	Northern Virginia (Ashburn) is the largest data center market on earth. Other clusters: Dallas-Fort Worth, Silicon Valley, Chicago, Phoenix, Atlanta, Seattle
Canada and Mexico	~4%	Toronto, Montreal (cheap hydro), Vancouver; Queretaro, Mexico City, Monterrey
Central and South America	~3%	Brazil (São Paulo) dominates, with growth in Chile, Colombia, and Argentina
European Union and UK	~22%	Amsterdam, Frankfurt, Dublin, London, and Paris form the FLAP-D cluster; the Nordics draw operators with cold climate and renewable power
Russia	~2%	Moscow dominates; law requires Russian user data to stay on Russian soil
China	~15%	Beijing, Shanghai, Shenzhen, Guangzhou; the Eastern Data, Western Computing plan moves new capacity to Guizhou and Inner Mongolia
Southeast Asia	~6%	Singapore is the hub; land and power limits push growth to Johor Bahru, Jakarta, Bangkok, and Ho Chi Minh City
India and Australia	~7%	Mumbai, Chennai, Hyderabad, Pune form the fastest-growing major market; Sydney and Melbourne serve the Pacific
Rest of world	~6%	UAE and Saudi Arabia are investing in AI-scale campuses; also South Africa, Japan, South Korea

Two forces pull capacity away from the dominant clusters. The first is latency, meaning the delay between a request and its reply. Light in fibre costs about 1 millisecond of round trip per 100 kilometres, so a user in Mumbai served from Virginia waits 150 to 200 milliseconds before anything happens, which a person notices in an interactive application. The second is data sovereignty: rules in the European Union, India, Russia, China, and elsewhere require certain data to stay inside national borders, so operators build locally even where the economics favour one large site elsewhere.

4. Workload Distribution

AI is not what most of this machinery does. The table below gives approximate shares of global data center computing by category for 2024 and 2025, drawn from IEA, EPRI, and Lawrence Berkeley National Laboratory reports.

Category	Share	What runs there
Enterprise and business IT	~25%	Company systems for accounting, sales, email, and internal applications; the largest category, and shrinking in relative terms as companies move to the cloud
AI and machine learning	~18%	Training large models and answering queries; the fastest-growing category, up from about 10% in 2022
Video streaming	~15%	Netflix, YouTube, TikTok, Disney+; mostly storage and format conversion at scale
Social media	~8%	Feed ranking, content storage, ad targeting, and notifications
E-commerce	~6%	Transactions, inventory, logistics, recommendations, and fraud detection
Search engines	~5%	Crawling the web, updating the index, and answering queries quickly
Financial services	~5%	Trading, fraud detection, payment processing, and regulatory reporting
Communications	~4%	Email, video conferencing, voice calls, and messaging
Gaming	~4%	Game servers, cloud gaming, and esports infrastructure
Scientific computing	~3%	Climate simulation, genomics, drug discovery, and particle physics
Government and defense	~3%	Intelligence analysis, logistics, satellite data, and classified facilities
Healthcare IT	~2%	Health records, medical imaging archives, genomic databases, and clinical AI
Cybersecurity	~1%	Threat monitoring, malware analysis, and security operations
Other	~1%	Connected-device backends, naming and certificate infrastructure, and content delivery

The AI share is the one number in the table that moves fast. It stood near 10% in 2022 and near 18% in 2024 and 2025. Asked to search the web for a current figure, Claude returned an ABI Research estimate of 38% for 2026, per DataCenters_8.5_Workload_Distribution.pdf. Whether that number proves right matters less than its direction: the category that barely registered five years ago is on course to become the largest consumer of the world’s computing capacity.