How much a 1 GW AI data center costs: about $30 billion, according to Bloomberg

Bloomberg has published an analysis of the cost of building a single AI data center. According to the piece, Meta, Amazon, Microsoft and Alphabet will spend some $750 billion on capital expenditure this year, mostly on data centers. The article compares that figure to the GDP of Ireland or Belgium and says it exceeds half of what S&P 500 companies spent in total last year. It is also equivalent to a $5,500 check for every US household.

The piece draws on an exclusive breakdown from Cushman & Wakefield, the real estate consultancy that advises on data center construction. That breakdown covers the cost of the structure (the “shell”) of a one-gigawatt AI facility in Abilene, Texas, where some of the world's largest computing centers are going up. One gigawatt is equivalent, according to the piece, to the consumption of about 750,000 homes at any given moment. Bloomberg Intelligence separately provided estimates of the cost of the hardware inside such a facility, and the article uses the high end of that analysis.

The conclusion is that building and equipping a theoretical one-gigawatt data center in the Abilene area could cost a developer about $30 billion. The line-item breakdown does not appear in the downloaded text, so I cannot confirm the detailed split.

The article attributes the rising costs to several factors:

• Construction materials such as steel and cement have soared because of trade wars, fuel costs and geopolitical conflicts.

• The AI boom itself has driven up the price of everything from chips to copper. According to Bloomberg Intelligence data, the short-term memory used in data centers costs more than five times what it did a year ago.

• Demand for equipment that generates and transmits electricity is so high that suppliers have order backlogs stretching beyond 2030.

• Local opposition and regulatory setbacks are forcing plans to be redrawn and more expensive alternatives to be used for resources such as power. The piece cites the case of Oracle, which issued a force majeure notice last month on a large data center in New Mexico because a gas pipeline key to the project suffered regulatory delays.

For John McWilliams, of Cushman & Wakefield's research group, spending like this has never been seen before. He also says that, given these costs, the figure no longer seems so outlandish, because these facilities are extremely expensive to develop.

The article further projects that more than 60 gigawatts of this capacity will be added in the US over the next four years, tripling current levels. Adding construction in other countries, from Brazil and China to the UK and Europe, estimates that the AI boom will cost trillions of dollars appear consistent. The spending is being felt in the markets for gas turbines, transformers, construction, real estate, hardware and networking, and has reached private credit, bonds and equities.

Bloomberg underscores the risks of this bet. These facilities are being built on the assumption that AI demand will keep growing and that the chips they house will keep delivering returns for years. However, some of the leading model developers are not yet profitable, so that demand is uncertain. The piece notes bubble warnings from Michael Burry and Ray Dalio. Burry points, among other things, to “circular deals,” such as those in which hardware makers like Nvidia partner with and invest in key customers like OpenAI. According to Burry, all the capital being spent and lent by the richest companies on the planet will not be enough to keep the bubble from eventually bursting.

Note to the reader: the text received does not include the line-item breakdown, which is the article's most numerically detailed section. This summary is limited to what appears in the text.