Microsoft Slowed AI Spending. Now It’s Playing Catchup

The Information The Information —

Microsoft is playing catch-up in the AI power race by grabbing big sites in Texas and West Virginia for data centers fired by natural gas.

It wasn’t long ago that the company had a big lead.

In the early days of the AI boom, Microsoft’s in-house energy team amassed a 9-gigawatt war chest of grid-powered data center sites that was the envy of the industry, according to people with knowledge of its pipeline and Wall Street tallies of announced projects.

That war chest was the equivalent of nine nuclear reactors.

Microsoft also held an advantageous place in line for connecting to several regional grids, the people said.

Then in late 2024 and early 2025, as Microsoft’s infrastructure capital expenditures looked like they could exceed the $80 billion it had budgeted that fiscal year, Chief Financial Officer Amy Hood put a curb on spending.

Microsoft’s energy team had to walk away from or slow down several data center deals in the works in both the U.S. and Europe.

To an extent not previously reported, Google, Oracle and others swooped in, shortening their waits for power in the Midwest and mid-Atlantic.

In some cases AI competitors grabbed actual sites Microsoft had released, and in others they were able to secure grid power because it relinquished its spot in utility transmission queues, according to members of infrastructure teams that negotiated power in those locations.

Oracle claimed a big chunk of available utility capacity for its data center under construction for Open AI in Port Washington, Wis., said a developer involved in those negotiations.

Google moved forward more quickly with two Indiana campuses because Microsoft was in wait-and-see mode, another energy professional with knowledge of the hyperscalers’ moves said.

Microsoft would never get back its place in line, although it has resumed developing several of the projects on revised timelines.

Alistair Speirs, the company’s general manager for its Azure cloud infrastructure, said, “Microsoft’s global infrastructure approach is built on flexibility and optionality, based on the near-term and long-term demand signals we see from customers.”  

Microsoft acknowledged last October that its Azure cloud business was short of capacity. “I thought we were going to catch up.

We are not,” Hood said on an earnings call.

Microsoft will be “capacity constrained” through at least the end of its fiscal year this June, she said. 

Hood told investors on an earnings call this January, “We need to make sure we’ve got power and land and facilities available” so it can put chips in them when they’re done “as quickly as we can.”

In recent months, as Microsoft stepped up its model development efforts, teams under cloud computing chief Noelle Walsh have signed deals or reached preliminary agreements with third-party developers to fill the computing power gap.

These partners, including newer AI infrastructure players Nscale and Crusoe as well as energy giant Chevron and investor Engine No. 1, promised to bypass the grid by building private, off-grid gas-power generation.

The starts and stops were painful for the Microsoft energy team that reported to Walsh, because securing power on the strained U.S. power grid can take years.

Several members have left, including Microsoft’s top energy executive, Bobby Hollis, who announced his departure on March 31.

Hollis did not respond to a request for comment.

Microsoft’s decision to forfeit power on the grid may have hurt its aim to host AI workloads. “Microsoft is going to fall behind Google” on AI compute capacity, one person with knowledge of the projects told me. “Google has an amazing team.

They continue to push.” 

Even so, that person and two other infrastructure managers familiar with the projects  acknowledged that Hood’s restraint at a moment of peak AI mania may still prove to be the right call.

That’s because Microsoft is putting more of the risk of delivering these data centers on its partners and may be in a position to grab attractive sites in the future when less-well-capitalized land and power speculators fail.

This also reduces its exposure to rapid changes in AI hardware and software.

Microsoft also has big campuses coming online now, including in Wisconsin and Georgia.

What has surprised a number of infrastructure professionals is how many of Microsoft’s new sites intend to rely on natural gas and operate independently of the grid, given the company’s longstanding commitments to cleaner energy sources and its deep relationships with power utilities.

Microsoft contracted with Crusoe for a 900 MW gas-fired data center complex adjacent to the flagship complex for OpenAI and Oracle that Crusoe is building in Abilene, Texas.

Microsoft is in exclusive talks for a 2.5 GW site in the Permian oilfields marketed by Chevron and Engine No. 1.

With both Texas projects, the developers offered gas turbine equipment they’d ordered years ago.

In West Virginia, Microsoft has a https://www.theinformation.com/newsletters/the-information-finance/nscale-crashes-big-leagues-data-center-developers">preliminary agreement with Nscale, which recently hired Microsoft’s former head of AI infrastructure for Azure, Nidhi Chappell, to power 1.35 GW of top-of-the-line Vera Rubin chips from Nvidia with fleets of Caterpillar gas turbines.

The state passed policies designed to speed building. 

Self-generated power is typically more costly because it requires more backup power to match the reliability of the grid.

If final estimates are too high, some say, Hood could still balk at projects.

Today, Microsoft is again moving fast to expand in Wisconsin and Indiana but is dealing with a more delicate political environment in those states.

To address growing local opposition to data centers, Microsoft asked to rescind certain property tax abatements already granted by the town of La Porte, Ind.

A pleased City Council member, Tim Franke, said in March that the yearly property taxes from Microsoft now have the potential to “double or triple” the city budget, and LaPorte Schools Superintendent Sandra Wood called the deal “significantly more lucrative.”

The company is also rapidly building again in Europe, spending $30 billion in the U.K. alone and using Nscale to construct a supercomputer that Microsoft has https://url3396.theinformation.com/ls/click?upn=u001.71kYkaWDpGOJSzbGrs4y1cGuC-2BUGDYiHRKTwPfdtj7rZMQwycZe8u-2FUIkxrtXMotc4JAEQTztG6wC0lU7GmwVbBb-2FFgWoEx9Amru75mQTJqEE1jXJ-2Fj2Qj9t-2BPDExKuLkLGA6-2FajvEAwjPgyFlbKY3vw-2FIoj8FZhISZMtFu-2Fi1W71JvnktBr-2B5dxL7vwmIE-2BocQI7KtVWHeDrVc9gkdfpeQ7KbGLGNlrBZW-2Bajhc4J2cEGUS4g1bzaY993CjIg6aCe56_GwpShWH4g0-2BfyOwxIsRujnOdqmodjYF42Fq5fLL7njGMVjl7jek1zgzOKbb1pdRZVA0zKnw7OvPJA0XRtTfk5oK7CEkBtwVj3cHnuCxnP-2BX4cirkJGk-2FpoBjJJisQxuMP6TKb5zml84wo4bIFshbv9FTOLPc2rxJ9U5ANvJToiIQGEkZLVVqt7R1ixG7RSnSslFctT3ZcI5cuUkhdHH6L9COSyhq7-2B9OAlP60yP3AW0syDCeg7L8RVNEVUSo20-2BXq1jVC7fIlv36j8Rzhyta-2FNNjqrITGye8y7vP9vQgX2M38J1V7Stg0jnV8caWMQ1AuaB6PxDt6y0pKl2Lm35GYw-3D-3D">dubbed that country's largest as well as other facilities in Norway and Portugal.

In other news:

New From Our Reporters

Upcoming Events

https://www.theinformation.com/events/financing-the-ai-revolution-2026" style="text-decoration: none !important;">Monday, April 27 — Financing the AI Revolution

Join The Information at the New York Stock Exchange on Monday, April 27, to hear from top executives and investors on how the rapid buildout of AI is reshaping tech, finance, and capital markets

https://www.theinformation.com/events/financing-the-ai-revolution-2026">More details

https://www.theinformation.com/events/ai-agenda-live-2026" style="text-decoration: none !important;">Wednesday, September 23 — AI Agenda Live SF 2026

Save the date for The Information’s annual AI Agenda Live in San Francisco, where top AI researchers, founders, investors and executives come together for a day of conversations about the breakthroughs and applications shaping the future of AI.

https://www.theinformation.com/events/ai-agenda-live-2026">More details

Read full article at The Information →