In a capital-intensive industry where money is the weapon that catapults startups to scale and overwhelms rivals, financiers can have a lot of sway.

This must be what private credit lenders are banking on as they write big checks to AI data center operators, who will need to spend as much as $30 billion just to bring a gigawatt of computing power online. Not content with earning double-digit yields, some are also dabbling with equity investment, with the hope of capturing more upside.

Public investors are now pushing back, asking whether AI data center developers can generate much shareholder value at all. Known as neoclouds, these companies buy processing chips and then lease them to customers that need computing power to build and run AI models.

Sydney-based Firmus Grid Ltd.’s failed $30 billion initial public offering last week was the first reality check. Blackstone Inc. is one of Firmus’ largest equity investors, owning 6.7% of the company. In August, it participated in a $2 billion funding round with a valuation of more than $10.5 billion, months after leading a $10 billion debt financing facility. The investment firm reportedly negotiated a wide range of clauses to protect its equity interests, in the event that an IPO fails to materialize or tanks upon listing.

And then came AirTrunk, acquired in 2024 in what was Blackstone’s biggest-ever investment in Asia. On Tuesday, news broke that the Australian data center operator is delaying a planned IPO of a real estate investment trust in Singapore. AirTrunk had considered offering a roughly 6% yield for its REIT, up from a previously envisaged 5.5%. That was not lucrative, considering that global investors can earn more than 5% from the US Treasuries without all that risk.

Granted, Firmus’ $30 billion ask was astronomical, not the least because it was a huge leap in valuation from just two months earlier. Of its 912 megawatt pipeline, just 46MW had been built. By comparison, CoreWeave Inc., a pioneer of the neocloud model, will have more than 1.85 gigawatts of capacity by year-end; it has a market value of $45 billion.

The problem with data centers is more deep-rooted than Firmus’ tone-deaf offering, however. Highlighting backlogs and pipelines is no longer enough. In the past, investors used valuation metrics more suited to fast-expanding firms, comparing a company’s enterprise value to its revenue, or earnings before interest, taxes, depreciation and amortization. But neocloud valuations have been steadily falling, a sign that stock markets are now less willing to reward rapid revenue growth on its own.

Rather, investors are asking when AI data centers can break even. Case in point: CoreWeave is expected to generate $7.6 billion EBITDA this year, but is still bleeding free cash flow. It plans to spend as much as $39 billion in capital expenditure, a lot more than what the company earns.

Investors are also starting to see a mismatch between technology and financing. While the debt is fixed and the developer is on the hook for repayments, the value of their equipment is not. Processing chips are on a three- to six-year depreciation schedule, and neoclouds might have to book impairments if the chips lose their technological relevance too quickly. After paying off their debt, how much residual value is there left for shareholders?

This is a question Blackstone must prepare for, since it still needs to monetize its equity stakes. AirTrunk is planning to go public in 2027; so is Lambda, another late-stage unicorn it has recently invested in. Given how fast sentiments are turning, the $1.3 trillion asset manager’s AI data center deals are looking shaky.