AI Capacity Race: Billions Pour into Data Center Boom

The global pursuit of artificial intelligence, particularly the quest for superintelligence, is fueling an unprecedented boom in data center construction. These sprawling facilities, often nondescript from the outside, are becoming the new symbols of computing’s frontier, housing the advanced infrastructure vital for training large language models developed by industry leaders like OpenAI, Google, and DeepSeek. This intense demand is driving a spending surge by major tech companies and even nation-states, pouring billions into ever-larger data center clusters.

At the core of these facilities are racks of powerful processors known as Graphics Processing Units (GPUs). Unlike traditional Central Processing Units (CPUs) that handle tasks sequentially, GPUs excel at parallel processing, making them ideal for the massive, concurrent calculations required to train complex AI models. Many AI data centers house tens of thousands of these GPUs, with a single rack of the latest AI chips demanding the same power as 10 to 15 racks in a conventional data center. As Chase Lochmiller, CEO of Crusoe, a start-up building a major data center for OpenAI, puts it, “The data center is the computer.”

Despite a brief period of investor concern in early 2025 following the emergence of a highly efficient AI model from DeepSeek, the infrastructure-building frenzy has only intensified. Microsoft, Alphabet, Amazon, and Meta collectively plan to increase their capital expenditures to over $300 billion in 2025. IT consultancy Gartner projects total data center spending to reach $475 billion this year, a 42 percent jump from 2024. Some forecasts, like McKinsey’s, suggest an even greater need, predicting $5.2 trillion in data center investment by 2030 to meet global AI demand. Meta founder Mark Zuckerberg recently announced his company’s investment of “hundreds of billions of dollars into compute to build superintelligence,” including plans for data center clusters large enough to cover most of Manhattan. Jensen Huang, Nvidia’s CEO, emphasized the essential nature of this infrastructure, stating, “I don’t know any company, industry [or] country who thinks that intelligence is optional.”

However, building these next-generation AI facilities is far more complex, costly, and energy-intensive than previous computational systems. Andy Lawrence, executive director of research at Uptime Institute, notes the significant gamble involved: “To suddenly start building data centers which are so much denser in terms of power use, for which the chips cost 10 times as much, for which there is unproven demand and which eat all up available grid power and suitable real estate — all that is an extraordinary challenge and a gamble.”

The dramatic increase in power demand from new AI chips has revolutionized data center design. Nvidia’s latest processors generate so much heat that traditional air conditioning is insufficient. Steven Carlini, Vice-President of Innovation and Data Center at Schneider Electric, explains that “Everything has been turned upside down,” with cooling and power equipment now occupying 70 percent of the facility’s footprint, compared to servers. Whereas a “big” data center 20 years ago might have required 20 megawatts of electricity, today’s AI facilities are designed for a gigawatt or more. This rapid evolution is so profound that Meta reportedly demolished a data center under development in Texas in 2023 to redesign it for higher-powered chips before restarting construction.

The relentless demand for computing power is also spurring a real estate boom as hyperscalers – large cloud providers like Amazon, Microsoft, and Google – develop vast clusters of data centers. These “AI factories” are often built for a single company or even a nation-state, a departure from the traditional model of shared servers. Key factors for location selection include cheap land, tax incentives, access to subsea cables, and, critically, abundant and affordable energy. Areas like northern Virginia, Atlanta, Columbus, Dallas, and Phoenix have become major hubs due to these advantages. For instance, Crusoe is building eight data center buildings totaling 1.2 gigawatts in Abilene, Texas, for OpenAI, as part of its ambitious $100 billion Stargate project, which will include approximately 400,000 Nvidia GPUs provided by Oracle. Meta is constructing a 2GW facility in Richland, Louisiana, while Elon Musk’s xAI targets 1.2GW across multiple sites in Memphis, Tennessee. Amazon is developing a 2.2GW site for Anthropic in New Carlisle, Indiana.

The global race for AI capacity extends beyond the US. An Oxford University study found that nearly 95 percent of commercially available AI computing power is operated by US and Chinese tech groups. China’s drive has led to data center construction in remote regions like Xinjiang and Inner Mongolia. Following a détente, Nvidia is set to resume some AI chip shipments to China, though US export controls on the most powerful semiconductors remain. Malaysia’s Johor Bahru is emerging as an AI hub for Chinese developers. Gulf states are also investing heavily, with the UAE announcing a massive data center cluster for OpenAI and other US companies as part of the Stargate project, aiming for up to 5GW of power. Saudi Arabia’s new state-owned AI company, Humain, plans to build “AI factories” with hundreds of thousands of Nvidia chips. The EU, meanwhile, intends to mobilize €200 billion to become an “AI continent,” planning five “AI gigafactories.”

The escalating energy consumption of these facilities is a major concern. The International Energy Agency forecasts data center energy usage to climb from 415 terawatt-hours in 2024 to over 945 TWh by 2030, roughly equivalent to Japan’s current electricity consumption. This surge is pushing operators to utilize any available energy source; xAI, for example, used gas turbines in Memphis while awaiting grid connection. All four major hyperscalers have recently secured deals for nuclear power. The immense and constant power draw, coupled with demand spikes during AI model training, poses significant challenges for utility providers, risking grid instability and outages.

Water consumption is another critical issue. Hyperscale and colocation sites in the US consumed 55 billion liters of water directly in 2023, with indirect consumption (tied to energy use) estimated at 800 billion liters annually, comparable to the annual water usage of nearly 2 million US homes. Tech giants acknowledge this, with Microsoft reporting 42 percent of its water from “water stress” areas in 2023, and Google nearly 30 percent from watersheds with depletion risk. Data centers in drought-prone states like Arizona and Texas, and developments in Georgia, have sparked local concerns over water shortages and increased municipal water costs.

To combat the extreme heat generated by AI chips, advanced cooling methods are essential. About two-fifths of an AI data center’s energy goes towards cooling. While early data centers relied on industrial air conditioning, the increased chip density now necessitates more sophisticated solutions. Operators are installing pipes with cold water to transfer heat, often using large cooling towers that consume vast amounts of water through evaporation (about 19,000 liters per minute per tower). More efficient, closed-loop chiller systems are also being adopted. The latest innovation is “direct-to-chip” cooling, where coolant flows directly over heat-generating components. Portugal’s Start Campus, for example, plans to use seawater as a heat sink for its upcoming 1.2GW AI data center hub, circulating over 1.4 million cubic meters daily through heat exchangers before returning it to the ocean.

Despite the hundreds of billions of dollars already invested, Silicon Valley leaders show no signs of slowing down. Nvidia projects that its next-generation “Rubin Ultra” systems, due in two years, will cram over 500 GPUs into a single rack consuming 600 kilowatts, presenting fresh energy and cooling challenges. OpenAI’s Sam Altman envisions facilities “way beyond” 10GW, requiring “new technologies and new construction.” This ambition is underpinned by the “scaling law” of AI – the belief that more data and computing power will endlessly yield greater intelligence. This drives data center designers to constantly innovate, and the relentless construction shows no signs of abating. As Mohamed Awad, who leads the infrastructure business at chip designer Arm, concluded, “At some point, will it slow down? It has to. But we don’t see it happening any time soon.”