A single AI data center now costs about $30 billion to build.
Meta Platforms, Amazon, Microsoft and Alphabet together expect to spend about $750 billion on capital expenditures this year, with data centers absorbing most of the money. That sum exceeds the annual economic output of Ireland and Belgium. It is more than half of what every company in the S&P 500 spent on capital projects last year. Spread across US households, it comes to about $5,500 each.
John McWilliams, a researcher at real estate services firm Cushman & Wakefield, said his team never expected to see four companies commit that kind of money to computing infrastructure in a single year. "We've never seen it," he said.
Cushman & Wakefield advises clients on data center construction and publishes market reports. The firm gave Bloomberg an exclusive cost breakdown for the shell of a single-gigawatt AI data center in Abilene, Texas, a fast-growing computing hub in West Texas. Bloomberg Intelligence built its own estimates for the hardware inside the building. At the high end of both analyses, a developer spends about $30 billion for a site that handles a gigawatt of computing power. A gigawatt roughly matches what 750,000 homes draw at the same moment.
The hardware takes the larger share of that bill. Servers, networking gear and short-term memory prices have run upward all year. Memory used in data centers now sells for more than five times its price of a year ago, according to Bloomberg Intelligence. Chipmakers and network equipment vendors report order books that stretch for years.
The building shell costs less but rises fast too. Prices for steel and cement have climbed as trade policy shifts, fuel costs and conflicts abroad squeeze supply chains. The AI boom has done the same for copper.
Electric equipment is the bottleneck. Utilities and developers queue for gas turbines, transformers and switchgear, and supplier backlogs now extend past 2030. A project that cannot secure a turbine waits years or pays a premium for a used unit.
Zoning fights and permitting delays add their own costs. Developers redraw site plans, pay more for power arrangements, or relocate projects. Oracle sent a force majeure notice last month on a large New Mexico data center after a gas pipeline serving the site ran into regulatory trouble.
The United States plans to add more than 60 gigawatts of data center capacity over the next four years, tripling a total that already leads the world. Brazil, China, the UK and Europe are adding their own. Trillions of dollars of investment and debt ride on the build-out, and the money flows through private credit, bond and equity markets to suppliers of turbines, transformers, construction services, land and servers.
The bet rests on demand. Every one of these buildings assumes companies will keep renting AI compute and that the chips inside will earn their keep for years. The firms building the most advanced models still lose money.
Michael Burry, who rose to fame betting against US housing before the 2008 crisis, has pointed to deals in which chipmakers such as Nvidia invest in customers like OpenAI and then sell them hardware. Ray Dalio, founder of Bridgewater Associates, has raised similar concerns about the cycle.
"All the capital being spent and lent by the richest companies on Earth," Burry said, will not stop the bubble from bursting if demand falls short.
Developers answer with a simpler calculation. The concrete, the power connections and the permits hold value long after the current cycle cools, and the chips inside get replaced while the site stays rented. That logic funds $30 billion holes in the Texas dirt, 60 gigawatts at a time.