Bubble Trouble?

Despite the unprecedented investment and hype, a majority of artificial intelligence initiatives fail to generate measurable financial returns.

A survey by IBM and Oxford Economics of 2,000 CEOs across 33 countries and 24 industries between February and April 2025 found that only one-quarter of their AI initiatives delivered the expected return on investment (ROI). A major slowdown or collapse in AI adoption could precipitate an oversupply of data center capacity.

Although most financial experts and industry analysts are bullish on the enormous build-out of the global data center ecosystem, many nonetheless believe a moderate and temporary correction is likely, with a few predicting a full-on crisis similar to the bursting of the dotcom bubble 25 years ago.

Market pessimists cite the dislocation between current investments in data centers and current revenue. Hyperscalers are spending hundreds of billions of dollars on new data centers and artificial intelligence infrastructure, but revenue directly attributable to AI services is a fraction of this investment, leading some to question the immediate ROI.

Still, the average forward price-to-earnings (P/E) ratio for top tech firms at present is about 26 to 38 times earnings, according to various sources—substantially lower than the over 66 times P/E ratio at the peak of the 2000 dotcom bubble. While not every hyperscaler betting big on AI and the data centers needed to realize its full potential will win the day, many will come out just fine.

Among those with an upbeat perspective for sustained AI demand driving data center growth is Peter Hanbury, senior partner and head of data center infrastructure at management consultancy Bain.

“We’re doing a lot of work with clients around enterprise AI, and I can tell you that it’s getting strong traction,” he says. “There’s less of a challenge now around monetizing AI, given already realized ROI, the rapid explosion in AI training, and growing confidence that it will drive inference to apply to specific use cases.” Inference is when an AI model is trained in processing new data to make a prediction or a decision.

Jonathan W. Welburn, professor of policy analysis at the RAND School for Public Policy, is also relatively sanguine about industry prospects. Looking back at the dotcom bubble, he points out that a few companies like Pets.com bit the dust but others survived, encouraging newer players to join them in what became today’s e-commerce ecosystem.

“I see something analogous happening in the AI ecosystem, where some tech companies build out ideas that other enablers make successful,” Welburn says. “The whole point of the markets is to allocate capital to areas of need and right now we’re moving historically fast in the allocation of capital to AI. When moving fast, some misallocations of capital are inevitable. The markets will eventually correct… where there has been an overallocation of capital, but a ‘bubble’ [like 25 years ago] is not how I would describe it.”

Insurance brokers close to the action also weighed in on the bubble phenomenon. “My feeling is that if you believe that AI is a useful tool that will continue to gain prominence in society at large, then the fundamental scaffolding provided by data centers will remain important. The infrastructure will continue to be required for the computing power as AI becomes more involved in our daily lives,” says Dave Nicholson, global client leader for reinsurance solutions at Aon.

Tom Quigley, Communications, Media, and Technology Practice leader at Marsh, leans more toward the capitalist reality that many data center builders will survive and others will either fail or struggle to compete. “Hyperscalers that choose the right partners and right contractual terms and build highly sustainable [data center] facilities have a better chance of surviving than someone…finding a suboptimal building that isn’t purpose-built for a data center and putting the equipment into the space,” he says.