Investor appetite for data centre IPOs is reaching its limit

When it comes to building data centres, finance comes in many flavours. But judging by the stumbles of some would-be listing candidates, publicly traded equity currently seems the least palatable.

AirTrunk, a Blackstone-backed operator of server farms, has postponed its plans for a $1.5bn initial public offering in Singapore. It is the third such deal to trip up in less than a month. Firmus last week pulled a $5bn float in Sydney, while SoftBank-backed SB Energy has pushed back its own IPO in New York.

The need for data centre capital remains huge: the combined $12bn sought by the three is a drop in the $5.2tn ocean of investment McKinsey predicts for the sector by 2030. But investors’ demand for unseasoned new AI infrastructure stocks looks to be wavering.

Each of the trio came with its own quirks. AirTrunk was structured as a real estate investment trust and prices of these, like bonds, weaken as interest rates rise. Nvidia-backed Firmus had tripled the valuation it sought since raising funds privately three months ago. Both Firmus and SB Energy have negligible revenue and are still lossmaking. That makes traditional valuation techniques for IPOs all but useless.

It’s not hard to see why they want equity nonetheless. Cash raised from an IPO can be spread among projects at the company’s discretion, which isn’t always the case with debt. On top of that, a public-markets listing builds a platform that can be tapped again.

But data centres’ blend of property, infrastructure and rapidly changing technology looks better suited to other kinds of funding. For example, project finance, where debt is secured against an specific assets’ cash flows, enabling investors to pick the projects they like and not the ones they don’t. Companies with little revenue and big ambition, meanwhile, are generally more at home in the portfolios of venture capitalists.

Lenders also have less exposure to the risk that servers and chips will become obsolete surprisingly quickly. If chips burn out faster than expected, it’s equity investors who bear the first losses. How long such equipment will actually remain usable is still up for debate. Rating agency S&P, though, already assumes that data centres will last at most 25 years, half the maximum lifespan it is prepared to consider for assets in general.

Estimates vary wildly of just how enormous the investment needed to support AI compute will turn out to be. McKinsey’s total includes non-AI data related development as well; rating agency Moody’s estimates $3tn by 2031. Either way, that’s a lot of capital. It’s just that public-market stock investors might not be the logical people to provide it.

jennifer.hughes@ft.com