Why AI Data Centers Cost So Much, and Why Power Is the Next Bill

The cost of an AI data center is no longer mainly the building. It is the chips inside, the power systems that feed them and, increasingly, the grid connection that determines whether the site can switch on at all.

Microsoft's latest annual report shows the scale. The company added $115.9 billion of property and equipment in the fiscal year ended June 30, up from $64.6 billion a year earlier, and it obtained $24.6 billion of finance-lease assets in addition. In the June quarter alone, Microsoft said capital spending was $41 billion, "including the impact from higher component pricing."

The mix explains why the bill is so large and why it turns over so fast. Amy Hood, Microsoft's chief financial officer, told investors in July that roughly two-thirds of the quarter's capital spending went to "short-lived assets, primarily CPUs and GPUs." Microsoft's annual report says servers and network equipment last two to six years. Buildings last much longer, five to 15 years under the policy in that report, which Microsoft is extending to 25 years from fiscal 2027.

Put simply, the expensive part of an AI data center has the shortest life. A company that builds one commits to replacing most of its contents within a few years.

Nvidia, the main supplier of those chips, has put a price on the build. On the company's Aug. 26 earnings call, Chief Executive Jensen Huang said each gigawatt of data center now represents about a $60 billion investment, up from about $30 billion five years ago. Nvidia puts its own revenue opportunity at about $40 billion per gigawatt with its coming Vera Rubin generation, up from about $25 billion with Grace Blackwell. That is about two-thirds of Huang's figure, by a Rackline calculation; the rest covers the building, land, power and cooling equipment and other suppliers' hardware.

Power equipment is rising as a share of that cost. As racks move toward a megawatt each, Nvidia has proposed shifting data-center power distribution to 800 volts of direct current, with new converters, switchgear and on-site energy storage to absorb the rapid swings in demand that AI training produces. Each of those is a new line in the budget.

Then there is the grid. In its October Short-Term Energy Outlook, the U.S. Energy Information Administration forecast electricity demand from the commercial sector, which includes data centers, to grow 2.8% in 2027. Electricity consumption across the country in the third quarter of 2026 was 4% higher than a year earlier, the agency said, helped by a hot summer.

Grid operators are struggling to keep up. The Federal Energy Regulatory Commission on Oct. 9 accepted PJM Interconnection's definition of large loads of 50 megawatts or more and a registry to track them, but rejected the rest of the grid operator's plan for handling them, including a service that would have let PJM curtail new large loads early in a supply emergency. Days earlier, FERC had accepted PJM's plan for a one-time purchase of up to 6,831 megawatts of new capacity but suspended it until Feb. 28, 2027. PJM says it faces a capacity shortfall starting in the 2027-28 delivery year, driven in large part by data-center load growth.

For developers, that turns power into a cost of time as well as money. A building and its chips can be ready long before a grid connection is.

The markers to watch are near. Microsoft reports fiscal first-quarter results on Oct. 28, after guiding to capital spending of more than $50 billion for the quarter. The next PJM steps after FERC's two rulings, and the EIA's November outlook, will show whether power supply is catching up with the money being spent on chips.