AI Data Center Projects Raise Demand Questions For Investors

Even as artificial intelligence drives demand for data center capacity to unprecedented heights, investors say AI-only data center sites and leases to AI-specific tenants pose the greatest risk in the market.

Leaders across the sector are generally dismissive of concerns about a potential slowdown in demand and fears that data centers are being overbuilt. But prominent data center investors at Bisnow’s DICE Capital Markets event this month said there are segments of the development landscape for which the demand outlook is far less certain and in which there is a genuine risk of overbuilding data centers.

Demand-side risk is heavily concentrated in segments of the market exclusively focused on AI, the investors said. This means sites far from traditional data center hubs that have little utility outside of training AI models. It also means projects developed for AI startups — companies that require huge blocks of data center capacity but present credit risks compared to the tech giants driving the bulk of the industry’s growth. 

While these AI projects are still being built, panelists like Hydra Host Special Projects Manager Kai Golden said they require more caution and creativity to finance.

“The supply and demand dynamics are obviously very clear right now — you need more capacity — but five or 10 years from now, the demand might still be there, but it might be for different use cases or in different regions, so it's natural to expect some sort of rationalization based on those workloads changing,” Golden said.

“As an investor, as long as you're being compensated for the risk around that development and people are going into it with their eyes open, I think that's fine.” 

Earlier this year, a narrative gained traction on Wall Street around a potential overbuilding of data centers and digital infrastructure to support AI. Fears of an imminent dip in demand for data center capacity grew following the release of an innovative AI model by Chinese firm DeepSeek that was touted as yielding similar results as other large language models while using far less computing power. 

Prominent voices like Alibaba Group Chairman Joe Tsai said the U.S. was building more data centers than it needed. Around the same time, Microsoft and Amazon Web Services canceled or paused a handful of planned data center projects, adding fuel to the growing alarm about a potential AI infrastructure bubble. 

Data center and tech leaders refuted concerns that AI efficiency would reduce demand, arguing that it would instead increase it. Their bullishness has since been supported by the actions of the world's leading tech companies, which have ramped up their planned data center investments. Consequently, much of the post-DeepSeek apprehension regarding demand has faded.

But while this optimistic macro demand narrative is now widely accepted in the data center industry, leaders at the DICE Capital Markets event said some of the industry’s fastest-growing segments do face a risk of overbuilding and questions about the sustainability of demand.

These overlapping segments all involve AI: remote projects suitable for generative AI training and projects serving pure AI firms. 

Data centers are pushing into geographies that wouldn't have been considered for the sector's development just a few years ago, with major tech firms and third-party developers pursuing projects in places like rural Louisiana and North Dakota. This expansion is due in part to constrained power grids in the industry’s traditional markets, which have made the massive blocks of power needed for hyperscale campuses increasingly difficult to acquire. 

But the shift toward these tertiary markets has also been catalyzed by the rise of AI, particularly the need for large blocks of power to train generative AI large language models. Unlike traditional cloud workloads that have determined data center site selection, AI training is typically not latency-sensitive and therefore doesn't require close proximity to end users or major data center hubs. This has allowed these workloads to be placed in remote locations with cheap power. 

But there is skepticism that this demand will continue, panelists said, with a real risk that tenants may not want to renew existing leases when they expire in the coming years. More efficient AI training — like that pioneered by DeepSeek — may end up boosting overall demand for AI computing, but that demand could shift away from centralized training clusters and toward inference deployments closer to end users. 

Goldman Sachs Managing Director Sapna Sirohi said AI firms may move training architecture closer to users, and some AI models could prevail while others exit the artificial general intelligence race. In almost any scenario where there is even partial pullback in demand from any segment of the AI training landscape, Sirohi and other panelists said it will be these remote AI-only markets that will feel the pinch. 

“If you're in a world 15 years from now where there is oversupply on the AI training side, those tertiary markets where you don't have the demand and you don't have the right infrastructure, those are the first at risk of rationalization,” Sirohi said. “We talk about these tertiary markets and people going where there is power. The reality for infrastructure investors like us is it's about what is the longevity of that demand.”

While data centers in the AI-only markets are considered a risky bet, so are projects anchored by AI-only tenants. 

While cloud and social media giants Amazon, Microsoft, Google and Meta remain by far the largest players driving the data center building boom, a growing number of AI-specific startups are also looking for data center capacity at massive scale. The largest of these firms are “neocloud” providers — companies like CoreWeave and Lambda that sell graphics processing unit computing specifically for AI as a service. 

Such tenants present a far higher risk profile for investors and developers than the credit-grade tech giants that dominate the leasing landscape. Beyond just their lack of a long-term track record, industry leaders have expressed doubt about the fundamentals of their business model. 

One particular source of worry is that neocloud providers need to sign long-term data center leases of 10 to 15 years, but their contracts with customers can be for as little as a few hours or a week. Longtime data center attorney Jeffrey Moerdler compared this mismatch to the leases that sunk WeWork. 

“I have a serious concern about those GPU as a service companies,” he said.  “I expect we're going to see a bunch of those companies, as well as a bunch of AI companies that don't generate revenue, going bust in the next few years.”

Still, projects at this riskier end of the market are moving forward. For investors and lenders at DICE: Capital Markets, their willingness to engage in these projects depends on their ability to command higher returns that are commensurate with the higher risk and put mechanisms in place to protect themselves if the worst-case scenario comes to pass. 

Hydra's Golden said investors are willing to underwrite data center projects in uncertain markets or with higher-risk tenants if yields are high enough. These deals offer a potential entry into the data center sector for investors seeking higher returns than typically found in hyperscale investments.

“If you are building a data center in the middle of nowhere, which might have a very real renewal risk in 15 years’ time, as long as your development yields compensate you for that, that's fine,” Golden said. “Or if you're signing shorter contracts with the neocloud providers, maybe you want a higher risk premium for that and higher rates, which is what we're seeing in the market.”

Investors are also putting safety mechanisms in place to manage their downside risk, particularly when dealing with newer AI and neocloud firms. In addition to collateralizing deals with the GPUs these tenants own, developers, lenders and investors are demanding significant security deposits or other guarantees. Often, this means upfront payments that give the owner flexibility to pivot the project if the tenant goes under. 

“I've done a ton of those deals where we're getting several years' rent in some form of security or a substantial portion of the capex to build out the building,” Moerdler said. “If that customer fails after two or three years, you now have a much lower basis in the building, having applied the security to the construction cost, and now you're able to market it on a very competitive basis.”