Some private-investment firms are hesitating to invest in dedicated electricity supplies to data centers running AI systems, warning that technical hurdles and strict reliability demands threaten so-called behind-the-meter power developers with the possibility of hefty financial penalties.

Large tech companies developing artificial-intelligence tools -- such as Amazon.com, Facebook's owner Meta Platforms, Google and Microsoft -- are increasingly looking for additional sources of power for their data centers as interconnection backlogs mean it can take years to hook up new projects to the power grid.

That creates openings for private-equity firms to bet on developers of dedicated power systems. But the obstacles inherent in those projects are giving some firms pause, starting with the requirement that generators must provide electricity continuously -- with no room for failure.

"Every single [data center] has a power project to go along with it. And although there are opportunities -- and we've looked at quite a few of them -- we're often struggling with some of the contractual complexities," Nicholas Hertlein, a managing director at infrastructure-focused Stonepeak, said during an industry conference in Houston last week organized by S&P Global. New York-based Stonepeak manages about $93 billion.

Data centers often include in their power-supplier agreements requirements that their electricity sources must supply "four nines availability," or operate 99.99% of the time, which allows them to shut down for less than an hour a year, private-equity managers say. That creates a potentially insurmountable technical challenge for the power-project developers and raises the risk of losses for their financial sponsors, the managers add.

"I think there's a big disconnect between developers agreeing to these stringent provisions in the contracts and when the sh -- hits the fan, they can't deliver it," said Himanshu Saxena, chairman and CEO of Greenwich, Conn.-based Lotus Infrastructure Partners. "If they have to pay LDs [liquidated damages], they'll get wiped out."

How behind-the-meter systems behave when required to supply power continuously to a data center -- as opposed to only serving as back-up during grid outages -- isn't fully understood, industry analysts say. The possible technical tangles are many: excessive equipment heating, asset degradation and intermittent fuel supply for natural gas-fired generators, as well as the difficulty of handling the often erratic electricity consumption of AI chips, according to the analysts.

"For those of us who build power plants, there is no way you can guarantee a three-nines, four-nines, five-nines availability," Saxena said during the conference, using gradations of reliability. He also contrasted data centers' requirements to the much lower dependability -- as much as 95% -- demanded by utilities from independent power producers.

Another disadvantage of behind-the-meter projects is they are typically limited to only one customer -- an undesirable situation even if the customer is a large tech company with a huge balance sheet, fund managers say.

"It may sound outstanding to have a $4 trillion counterparty sitting on the other end of a behind-the-meter power plant, but it also means you have no other alternative," said James Berner, a partner at asset manager BlackRock's roughly $200 billion infrastructure-investment arm, Global Infrastructure Partners.

"Having all your eggs in one basket, even if it's a hyperscaler basket, is in itself a risk," Berner added, referring to large tech companies using AI data centers.

The breakneck speed of AI expansion, and tech companies' willingness to pay higher prices for the electricity they need to power their data centers, will likely continue to lift demand for behind-the-meter generators. To profit from this trend, private-equity firms must be able to back solid project developers while mitigating investment risks, fund managers say.

"You'd better be very clear on your contract terms," Hertlein said.