The data-centre investment spree shows no signs of stopping
If investment in data centres is about to slow, nobody told Mark Zuckerberg. On January 29th, during an earnings call, Meta’s boss boasted that the social-media giant had plans to build an artificial-intelligence (AI) data centre “so big that it’ll cover a significant part of Manhattan if it were placed there”.
His timing was conspicuous. Only two days earlier the share prices of firms from Nvidia, a chipmaker, to Dell, a manufacturer of servers used in data centres, had nosedived in response to the release of a new AI model created by DeepSeek, a Chinese firm. Its training costs were a fraction of those for similarly powerful Western models, raising questions over how much computing power—and investment—is needed to develop ai.
Although many of those share prices have since recovered, the episode has brought increased scrutiny to the huge sums of money that are being spent on data centres. Meta and America’s three big cloud-service providers—Alphabet, Amazon and Microsoft—splashed a combined $180bn on data-centre infrastructure last year. Add in spending by smaller tech firms, telecoms providers, big enterprises and data-centre operators such as Digital Realty and Equinix, and the figure rises to around $465bn. Land, buildings and peripheral gear such as electrical equipment make up about 30% of that, with chips, server racks, networking kit and the like accounting for the rest. Cashed-up private-equity firms such as Blackstone have been lured in by the spending boom, undertaking a record $70bn-worth of data-centre deals last year.
From northern Virginia to Johor Bahru in Malaysia, there are now as many as 11,000 data centres around the world, by one estimate. They collectively guzzle some 55 gigawatts (GW) of power (about as much as the generating capacity of the Netherlands) and take up around 300m square feet of land (more than 4,500 football pitches). The Americas are home to a little over 50% of capacity, measured by power consumption, with Asia accounting for just under 30% and Europe, the Middle East and Africa making up the remaining 20% (see map).
The spending spree shows no signs of letting up. On February 4th Sundar Pichai, Alphabet’s boss, said during an earnings call that his firm’s capital expenditure would hit $75bn this year, mostly for data centres, up from $53bn last year and more than analysts had expected. At least another 63GW-worth of data centres are set to come online over the next few years, according to Omdia, a research firm.
Yet those worried that the world will soon be awash in excess data-centre capacity need not be, for two reasons. The first relates to supply, of which there is currently too little. According to CBRE, a property firm, just 2.8% of data-centre space in North America is unoccupied; capacity is often “pre-leased” even before it is built, typically for 10-15 years.
Alphabet, Amazon and Microsoft all blamed capacity constraints for slower growth than expected in their cloud divisions in the last three months of 2024. In an effort to keep up with demand, the cloud giants have been leasing facilities from data-centre operators. But they, too, have been struggling to expand supply quickly enough. Jon Lin, an executive at Equinix, points out that it can now take three years to build a data centre, up from 12-18 months not long ago.
Supply chains are part of the problem. CBRE notes that a growing number of projects are in the “under-construction phase for extended periods” because of power shortages. The wait for transformers, a critical electrical component for data centres, can be years.
Governments around the world have also gummed up development. In Malaysia, which has emerged as a regional data-centre hub, officials have rejected applications for new facilities to conserve water and power. Last year Dublin axed a data centre planned by Alphabet; other cities have put in place temporary moratoriums on new facilities. In Virginia, home to a quarter of America’s data-centre capacity, local authorities are tightening zoning rules to limit where data centres can be built.
The second reason not to expect a data-centre glut relates to demand, of which there will still be plenty in the years ahead. At the end of 2024 AI soaked up only about a tenth of data-centre capacity, according to Goldman Sachs. Before the DeepSeek ruckus, the investment bank predicted that AI’s data-centre usage would triple by the end of 2026. Even so, it would still have accounted for only about a quarter of the total (see chart).
Various other forces are propelling demand for data centres. Many companies are still in the process of digitising their operations and moving them to the cloud. Internet usage continues to rise as consumers in poor countries gain access and those in rich countries gorge on video content. Devices from fridges and cars to industrial machinery now generate oodles of data, too.
Recentred
Even if demand from AI decelerates because models require less computing power to train, data-centre operators say that their facilities could be repurposed for other applications. The biggest difference between AI and regular data centres is how much power they consume per square foot: running whizzy AI chips in particular requires lots more energy. That means cooling systems have to be upgraded, though not much else. Switching between uses is “pretty seamless”, says an executive at one data-centre operator.
What is more, plenty of techies are optimistic that cheaper AI will increase demand for the technology. Some also point out that new “reasoning” models use more computing power when responding to queries, in order to generate better answers. Whether those effects will be enough to offset a fall in the number-crunching capacity needed for model training is not yet clear. In the meantime, the data-centre bonanza goes on. ■
To stay on top of the biggest stories in business and technology, sign up to the Bottom Line, our weekly subscriber-only newsletter.