AI data center boom isn't going bust, but the 'pause' is trending

CNBC

27-04-2025

Sprawling, power-thirsty data centers have become as much a staple of suburbia in some places as shopping malls and soccer fields. However, when Microsoft pulled the plug on planned data centers in Ohio last month, it added to questions about whether the nascent data center boom had already gone bust. A Wells Fargo report last Monday saying some data centers planned by Amazon Web Services were being reconsidered added to market anxiety.
But the bust may have been over before it ever began. And if anything, a "pause" on some data center projects comes within a spending environment that remains strong.
"We continue to see accelerated scaling of AI deployments across the data center market, with strong demand signals reinforcing both our near- and long-term growth," said Giordano Albertazzi, CEO of Ohio-based data center supplier Vertiv on an earnings call last week. Its shared ended the week up 22%.
Amazon and Nvidia both reaffirmed last week that the data center market remains strong.
"There's been really no significant change," Kevin Miller, Amazon's vice president of global data centers, said at a conference organized by the Hamm Institute for American Energy. "We continue to see very strong demand, and we're looking both in the next couple years as well as long term and seeing the numbers only going up."
That doesn't mean the strategic thinking about how, where and when exactly to spend isn't changing as the AI market evolves and breakthroughs need to be digested. In the span of six weeks this year, China's DeepSeek burst onto the scene, President Trump's $500 billion AI-powered Stargate initiative was announced, and concerns over tariffs and trade wars roiled markets.
"All of that has created a scenario where the data center industry is taking a bit of a pause, broadly," said Pat Lynch, executive managing director for commercial real estate company CBRE's Data Center Solutions. "I think it is a temporary pause," Lynch added, noting that the project pipeline and its funnel remain significant and CBRE continues to execute deals. "I remain cautiously optimistic about future demand, particularly when you think of large AI training models," Lynch said.
Microsoft had pledged a $1 billion investment in Ohio-based data centers in the same area where Intel has planned chip factories, but the timeline has slowed.
"After careful consideration, we will not be moving forward with our plans to build data centers at the Licking County sites at this time. We will continue to evaluate these sites in line with our investment strategy," a Microsoft spokesperson said in a statement to CNBC.
A UBS report from last week concluded that among all the possible explanations for data center cancellations, it was most likely that Microsoft had overcommitted amid the AI rush, and was now zeroing in on the projects that currently make the most sense. It noted that Microsoft's leased capex was up 6.7x in the span of two years, with lease obligations of roughly $175 billion. "Microsoft bought up as much available leased data center capacity as it could in 2022-2024 and now has the visibility to eliminate some of these 'early-stage projects,'" UBS wrote. "We find the least support for the 'demand lull' explanation," its report added.
Anat Ashkenazi, Alphabet CFO, described the cloud supply-demand environment as "tight" after its latest earnings on Thursday. "We could see variability in cloud revenue growth rates depending on capacity deployment each quarter," she said. "We expect relatively higher capacity deployment towards the end of 2025."
"We're not seeing a retreat from demand but a strategic reallocation," said John Carrafiell, co-CEO of BGO, a global real estate investment manager with $83 billion in assets under management, including a significant data center portfolio. The most significant players, he says, are not pulling back, with Microsoft, Google, Meta, and Amazon planning to spend over $300 billion in capex this year-largely tied to AI infrastructure. And, he says, that doesn't include other major players, such as OpenAI and Oracle, both involved in the Stargate project.
"Rather than a bust, this is a reshuffling of the deck in an environment where power in particular, along with fiber, water, and land — are scarce and strategic," Carrafiell said. Long-term enterprise adoption will drive AI demand and data center demand for the next decade. "We aren't even in the first inning yet," he said.
Power is the lifeblood of data centers, but data centers aren't plug-and-play operations, requiring copious amounts of electricity for computing power and fans to keep the infrastructure cool. As generative AI adoption moves from early experimentation to enterprise-scale application, the need for low-latency, high-efficiency data centers near end-users will intensify, but it will take time for the right set of conditions to line up with the expected data center square footage.
"New data centers are increasing in size so dramatically that the grid cannot keep up," said Allan Schurr, chief commercial officer of microgrid developer Enchanted Rock. Three years ago, a large data center was 60 megawatts — enough power to supply 20,000 homes, but now he says new data centers to support all the uses of artificial intelligence are requesting 500 megawatts or more.
This rapid growth in electricity use is on top of new demand from manufacturing and the electrification of transportation, which together are weighing on supply and infrastructure. Data centers pose a unique challenge to utilities, which must ensure they can supply power to all customers, even in times of peak demand. "This is why some utilities are quoting long interconnection wait times for data centers," Schurr said. "Utilities need to invest in new substations and may also need to expand transmission and generation, all of which takes time," he added.
CBRE has seen data centers go from comprising 2% of its portfolio in 2022 to 10% in 2024, and Lynch expects that to keep growing, and power proximity is driving the current marketplace, as data center builders seek areas with access to plentiful power. Georgia, Texas, and Ohio all check a lot of the boxes builders are looking for, and if an area doesn't have the grid or infrastructure capacity, it needs to be able to scale up fast.
"Having large power availability inside of 36 months is attractive to clients," Lynch said.
Three percent of the world's power is now tied up in data centers, according to Datacenters.com.
Schurr said Enchanted Rock's data indicates there is plenty of power available to meet demand — most of the time. Of the 8,760 hours of the year, the grid is only under stress for a fraction of them. "If we can alleviate demand on the grid for those 100 to 500 hours, the long interconnection delays can be shortened," he said.
There is an important distinction to be made between the idea of a broader slowdown and some of the recent pauses enacted by major technology companies, according to McKinsey & Company senior partner Pankaj Sachdeva, who researches data center development and expects an ebb and flow.
"It will not be linear," he said.
Based on recent McKinsey modeling, which does not include tariffs impact, the data center market is expected to grow in the 20%–25% range over the next five to seven years, but year to year there will be variations in the growth rate.
Tariff changes will introduce new cost pressures across AI and data center supply chains, particularly with critical mineral tariffs on the horizon.
"These disruptions will elevate hardware costs, impact sourcing strategies, and require businesses to rethink their long-term procurement models," said John Archer, senior delivery principal and supply chain transformation Leader at Slalom Consulting. In the short term, AI and cloud providers will need to implement cost-mitigation strategies such as renegotiating supplier contracts and optimizing inventory.
"Longer-term, a push towards geographic diversification, co-manufacturing in tariff-friendly regions, and deeper integration of AI-driven supply chain analytics can be expected to adapt to evolving trade policies," Archer said.
One factor that hasn't changed is that compute power is currently expensive, and much more of it is needed for AI software and hardware, according to Suresh Venkatesan, CEO of POET Technologies, a publicly-traded company that develops power solutions for data centers. "The explosion in AI challenges data centers to find more efficient solutions because AI requires compute power in such volume that it's unlike anything we have ever witnessed," he said. "While one data center project may hit a wall, others are likely to spring up, because there is no indication of a slowdown in demand for connectivity," he added.