Data Center Inflation Is Here
In the first weeks of the year, it has become clear that building data centers is getting more expensive. On a relative basis, the shift could benefit the cash-rich tech giants, particularly Google, while potentially hurting cash-burning startups.
This doesn’t mean giant data centers won’t get built. Investors are still enthusiastic about AI, and the tech giants that are funding the buildout are still rolling in cash. It could mean these companies see lower returns and some projects don’t get built or are scaled back.
The biggest and hardest-to-quantify cost is power, which has long been a major constraint on data center construction. Local governments have been complaining for more than a year about rising electricity costs, and the White House, which has woken up to concerns about affordability, weighed in.
Last week, the Department of Energy and several governors in the middle Atlantic region effectively demanded that companies building data centers agree to 15-year contracts to buy power from new power plants.
Auctions for power run by grid operator PJM have hit records, but the prices haven’t been high enough to get power companies to build new power generation. That means data center developers will have to spend more. The notably bipartisan effort is expected to generate $15 billion in new power plants.
Both Google and Microsoft appear committed to delivering their own power. As my colleague Ann Davis Vaughan wrote in our new AI Infrastructure newsletter, Google’s $4.8 billion purchase of power producer Intersect will speed its data center development. We also reported that Google is asking federal regulators to speed approval of data centers that bring their own power. Microsoft said last week it would pay more for electricity and subsidize the expansion of the electric grid.
Electric power means computing power and that could become the differentiator among AI companies. Those that can afford to build it will move ahead and those that can’t fall will fall behind. Computing power translates into revenue, a connection made clear by OpenAI’s chief financial officer, Sarah Friar, over the weekend. The company’s computing capacity roughly tripled in each of the last three years and so did its revenue.
If power was the only higher cost, it could be manageable. But there are others. Financing could get more expensive for two reasons. First, long-term interest rates hit their highest levels in more than four months this week, driven in part by President Donald Trump’s tariff threats, tied to his obsession with taking over Greenland.
Second, there will be an estimated $300 billion of AI debt issuance this year. While a lot of it will come from the tech giants, which have pristine credit, the sheer amount of borrowing could outweigh demand. That means yields could rise, making borrowing more expensive. The weight of that more costly debt will fall on the companies with less pristine credit, such as Oracle or CoreWeave. (There’s another more wonky way to think about it: Investors buy up the super-safe debt issued by the tech giants and sell off generally lower yielding U.S. government debt, which drives up interest rates for the rest of the economy.)
Then there are the costs of the gear needed to build data centers. You can start with commodities such as copper, which rose 45% last year. Demand for the metal is expected to keep growing over the next decade. AI data centers use far more copper than typical data centers do, for everything from electrical wires and components to cooling systems and servers.
Then there’s the manufactured goods. The wait time for new natural gas turbines is three to four years, making it a terrific seller’s market. That’s one reason companies such as GE Vernova are raising prices. The company has boosted margins and expects profits to grow as price increases flow through its order book. GE Vernova lifted its dividend, boosted share buybacks and is generating piles of cash. Its stock was among the market’s best performers last year.
The wait time for gas turbines is just one of many factors causing delays for data centers, as my colleague Anissa Gardizy, who also writes the AI Infrastructure newsletter, reported. There have been delays installing water-cooling systems, which Nvidia’s superfast chips require, and delays getting those chips up and running. There have also been delays connecting to the power grid and getting local approval.
The longer it takes to open a data center, the more it will cost and the harder it will be to earn a return on that investment.
I’m focusing on rising costs right now because I think their impact will start showing up this year. In extreme cases, developers will be squeezed for cash and might need to raise money to finish their projects. Higher costs might also make some cancel planned projects before they get off the ground.
As we saw with OpenAI, more computing capacity means more revenue. Those who can’t pay will fall behind.
New From Our Reporters
