The explosive growth in AI-driven data center development is being accompanied by more rigorous underwriting standards, growing concern over technology obsolescence, and expectations that some projects will fail even as demand continues to outpace supply.

That was the conclusion several senior partners at Sidley Austin shared at the firm’s infrastructure media roundtable on 24 September, where members of the infrastructure and finance teams delved into how investors remain eager to deploy capital into digital infrastructure and data centers but are increasingly taking a more disciplined underwriting approach.

“A lot of projects have died over the last couple of years, and that’s a good sign because that means that people are not just saying, ‘Oh look, I can just throw this money into any old project and it’ll work out’,” said Tim Chandler, a Houston-based private equity partner at Sidley.

Seeing these failures, investors are placing a greater emphasis on sponsor quality, execution risk, and a developer’s ability to navigate complex and changing regulatory and interconnection challenges. Access to land and equipment, and plans to build or procure power are no longer guarantees to a project’s success.

There was a period where capital was chasing the theme, explained Chandler. Perspectives have changed, he added.

Tara Higgins, co-head of Sidley’s Energy, Transportation and Infrastructure group, said higher collateral requirements are helping separate serious developers from speculative entrants.

“People are looking for more and more money up front, and that is sorting through the players surprisingly quickly,” Higgins said.

Without access to capital, smaller players are finding it harder to pursue large-scale projects.

“You kind of have to be a big girl to do a data center,” she added.

Brien Wassner, partner and global co-leader, private equity, took Chandler and Higgins’ points a step further.

“We do expect a number of data center projects to fail,” he said. “There’s going to be winners and losers in this situation. There’s going to be bankruptcies around.”

Failing projects

The panel discussed the various potential causes of failed projects, from regulatory pauses on interconnection queues to community opposition.

“One hundred percent; community pushback will cause [some] data center developments to fail. Full stop,” said Chandler.

While opposition does not necessarily reduce overall demand for data center capacity, panelists said it could delay projects, increase costs, and redirect capital to areas that are more supportive of growth.

Rather than signaling weakness in the sector or declining demand, however, Wassner added, the anticipated failures reflect the unprecedented volume of projects competing for limited pools of grid capacity and capital.

Wassner said the underlying need for data center capacity remains strong enough that the market still faces a shortage of supply.

“If things keep going the way they are, we can’t build enough data centers,” he said.

Panelists suggested that project failures are a natural evolution of a rapidly growing market. Failed projects may still possess valuable land positions, transmission access, permits or equipment contracts that can be acquired by better-capitalized developers.

The discussion also highlighted a potential opportunity for distressed investors and private credit providers.

“There’s going to be buyers of those bankrupt assets that will then deliver them a second life,” said Wassner.

Higgins flagged debt funds as a potential safety net for some of the riskier infrastructure assets, which could provide capital to projects facing financial difficulties.

“They [debt funds] could be a safety net if things don’t go well for some of these data centers,” Higgins explained, and the panel concurred that lenders willing to acquire stalled developments could play an important role in bringing projects across the finish line.

“We’re going to continue to see creative structures for capital deployment,” said Wassner.

Lenders that are willing to fund restructurings or acquire stalled developments could play an increasingly important role in bringing projects across the finish line, panelists said.

“The money’s still there,” said Daniel Allison, a Houston-based infrastructure finance partner at Sidley. “The model is a little more sophisticated on, ‘What’s the risk here? How would it react if there’s a delay of six months?’”

A VCR problem?

Panelists said technological obsolescence is affecting both data center and GPU financings, but lenders are still underwriting deals, relying on the credit strength and judgment of hyperscalers.

Unlike traditional infrastructure assets such as pipelines, transmission lines or power plants, data centers are tied to a rapidly evolving technology where advances in chips, computing efficiency and AI models can alter demand assumptions in a matter of years.

Higgins said financiers are increasingly asking what she described as the “VCR problem” when evaluating projects.

This question reflects growing uncertainty about whether facilities being financed today will remain competitive over the life of their debt, or will be replaced by newer technology just as VCRs were made obsolete by DVDs and streaming video. Higgins said lenders and investors are debating whether future generations of hardware could require less space, less power or different facility designs than current systems.

“Are these data centers the same size in five years? Is there a rate of innovation that makes the rack smaller, the equipment smaller?”

Chandler said similar concerns are emerging around the pace of improvements in AI models and compute efficiency. He noted that newer AI models are already delivering stronger performance at lower computing costs, raising questions about how elastic demand will ultimately prove to be.

“If, for example, suddenly they can train and use these models at a tenth of the cost, what does that mean for these companies?” Chandler asked.

While speakers stressed that these concerns have not slowed dealmaking, they are increasingly influencing financing assumptions, particularly around equipment depreciation and long-term asset values. Rather than relying solely on projected AI demand growth, investors are beginning to model how advances in chips and software could affect the economics of facilities years after construction.

The result is a shift in underwriting focus: alongside questions about power availability and permitting, investors are now evaluating whether today’s state-of-the-art data centers will still represent premium infrastructure assets a decade from now.

Despite these concerns, panelists said lenders and advisors alike are taking comfort from the involvement of large hyperscalers such as Meta, Amazon, Google and Microsoft, whose balance sheets and technical expertise allow them to continually evaluate and adopt the most competitive computing technologies.