Will data centres avoid the curse of the mega-project?

Big building projects follow an iron law, argues Danish geographer Bent Flyvbjerg: they come in “over budget, over time, over and over”. Will that apply to the AI infrastructure boom too? Global data centre investment will total $31.6tn by 2050, consultancy PwC estimates, and to say builders are in a hurry is an understatement. There is much room for error, human and otherwise.

Data centre builder SB Energy is bullish, at least about its own chances. An offshoot of Japan’s SoftBank, it has filed for a New York initial public offering, hoping to lure investors with a pipeline of projects including Ohio’s enormous PORTS-Pike. It supplies buildings and systems, but not chips and servers, so isn’t exposed to the price wobbles of those essential AI commodities. OpenAI and SoftBank have signed decade-plus contracts; all SB has to do is build — ideally on budget, on time, over and over.

This is bold, because in data centre building, many things can go awry, from materials shortages, planning delays, labour scarcity and access to power. Getting permission to plug into the grid, known as “interconnection”, can take two, five, seven or even 10 years, depending on the state. SB’s data centres will mostly be co-located with new power plants, which bring construction and planning challenges of their own.

Then there’s a new challenge for American builders: we, the people. President Donald Trump has blustered about Americans who would rather be “backwards and poor” than have a data centre in their back yard, but he is at odds with the public mood. A recent Gallup poll shows 71 per cent “somewhat” or “strongly” oppose local construction of these giant AI sheds.

SB thinks it can overcome those obstacles. It has never built a data centre itself, although a recently acquired consultancy arm has completed 15 in the past. It has filed for more interconnections than it needs, and being early is surely an asset, but it doesn’t guarantee timely approvals. Filings concede that “growing public resistance” to AI is an ongoing risk.

Time literally is money. Some data centre builders find a tenant only when the project is substantially under way. In SB’s case, the tenants are lined up already, but that creates deadline risk: if the project is late, SB could face enforced discounts and cash flow shortages. In some cases, its filings warn, the customer can buy out a delayed project outright, potentially at a low price.

Such risks can be priced, with a cool head. But in this market? At the $50bn it reportedly seeks, SB’s enterprise value would be in the region of 400 times its full-year ebitda based on what it made in the first six months of 2026. OpenAI has taken warrants in SB that pay out if the company’s equity valuation hits $80bn, suggesting Sam Altman’s tech group is bullish.

At least SB has deep-pocketed backers, in case the iron law of infrastructure gaffes holds true. And if data centres really are essential for global prosperity, these risks will have to be taken by someone. The shale boom showed America can pull off impressive mega-feats of construction, and quickly too. The bust that followed showed they don’t always work out as intended.

[email protected]

Lex sends a newsletter every Wednesday with additional insights, the pick of the week’s columns and links to what we’re reading. Premium subscribers to the FT can sign up