Welcome to AI Infrastructure, a new newsletter focused on how the data center boom is reshaping the energy sector and turning Silicon Valley into a rising industrial power.

Ann Davis Vaughan has covered the intersection of energy and technology as both an investment analyst and journalist for two decades, including 14 years at The Wall Street Journal. Based in Houston, she has researched electrification and previously evaluated dozens of companies that form the backbone of the economy for Select Equity Group, a multibillion-dollar investment firm.

She and the rest of our infrastructure team at The Information bring you exclusive news and analysis on the biggest infrastructure buildout in generations.

Two key aspects of the AI race are misunderstood: The gigawatts of electricity needed for the most ambitious AI data centers aren’t coming online nearly as fast or as easily as the announcements about them did. And the acute shortage of “AI compute”—power and servers—is actually producing some thrilling innovations in energy, climate tech, industrial automation and computing much faster than government-led policies and stimulus plans ever could.

Let’s start with the reality check and finish with the thrill.

Breathless chatter about the AI race makes it seem like the roughly 1 trillion dollars per year that McKinsey estimates will be spent on the infrastructure buildout between now and 2030 is already producing fully electrified AI cities on a hill, one after another.

I’ve been traveling the country to see the steel and concrete rising (and, when I can get inside, the racks rolling in), and it may surprise you how many of the multigigawatt sites you hear about won’t hit full power capacity for years. JPMorgan Chase estimates that U.S. firms this year struck deals to add 9 GW of AI data centers, up from 4 GW of deals last year. If and when that capacity does come online, it’ll be just 13% of the way to the roughly 100-plus GW of AI compute that banks like JPMorgan and Goldman Sachs conservatively project will be coming to the U.S. by 2030. (RAND and others project even higher figures of more than 300 GW.)

Despite public boasts by Elon Musk and other AI leaders about electrified clusters of hundreds of thousands of chips, and announcements of sites so big that people in the data center industry started to joke about “bragawatts,” we’re just barely starting to consolidate one or more gigawatts of data center capacity on a single campus to prove that denser clusters will produce better AI.

Unless dreams quickly come true for nuclear fusion or AI servers in space, it doesn’t look likely that OpenAI CEO Sam Altman can get his hoped-for 250 GW of capacity by 2033, or that xAI founder Elon Musk can soon get to his stated goal of a terawatt (1,000 GW)—equivalent to all of the U.S. electricity produced today—by some unknown date.

OpenAI has been furiously hunting down potential data center sites from Wisconsin to Michigan to move beyond the power limitations of its existing sites. For instance, Crusoe, a developer of OpenAI facilities in Abilene, Texas, just announced it had “topped off” the eighth building there, though that merely refers to the erection of the eighth structural frame. There’s much more to come before chips are installed and all systems are go to achieve 1.2 GW of capacity, expected by mid-2026. In January, Oracle CEO Larry Ellison said the Abilene site could grow to as many as 20 buildings, but the Texas grid operator can only supply power for eight buildings until new transmission lines can be built, sometime in the several years, people involved in the projects say. Even getting equipment like transformers, which adjust voltage from the grid or self-generated power before it flows inside the data center, has been so difficult that an OpenAI manager told me in September they’d almost considered assembling their own.

Hidden Setbacks

There are many other infrastructure setbacks and pivots for high-profile U.S. data centers you haven’t read about (a power transformer from China had to be rebuilt, a fiber installation flooded a nearby aquifer, a gas pipeline was maxed out), and some you likely have, such as CoreWeave’s troubled Texas project with Core Scientific. In Pennsylvania, where a bipartisan state government and energy CEOs are working to usher in a golden age of gas-fired AI, the regional grid regulator recently failed to approve even one of 12 proposals to address electricity demand growth. That has left the busiest computing region in the country, also home to Data Center Alley in Northern Virginia, in limbo. A federal proposal may supersede those proposals, but rising power bills are making the debates politically fraught.

As Microsoft, Amazon and Meta become the single largest electricity customers in states where they’re developing facilities, they’ve faced an increasing public backlash that could shape next year’s midterm elections. Just two projects in Wisconsin from Microsoft and OpenAI could require the doubling of available capacity at Wisconsin’s main utility, We Energies, to about 11 GW from 5.3 GW by decade’s end, if they grow as big as the companies envision, according to many people involved in the power planning.

Such friction is pushing these companies to develop or use energy sources outside the public electrical grid, which could speed up their projects. A leading architect of such behind-the-meter power, KC Mares, believes as much as 4 GW of private power installations got underway in 2025, and he knows of 10 GW of projects that will be started next year. But it’s a complicated dance: Mares says makers of small gas turbines, which the industry turned to because waits for big ones hit five to seven years, are also now virtually sold out until 2028. The massive off-grid renewables, batteries and gas sites—like the one Google is developing with Intersect, a clean energy and data center builder—will still draw some utility electricity and demand heroic logistics; Intersect put in equipment orders long ago.

The great news is that the severity of these challenges, and the risk that precious server chips will sit idle, unable to morph into AI products, has unleashed a torrent of innovation beyond AI models.

The Boom’s Innovation Dividends

AI doomers severely underappreciate the dividends of this infrastructure race. Power scarcity is catalyzing a badly needed reinvention of a broken, century-old grid. The $250 billion that the largest tech companies spend annually on research and development—over and above capital expenditures—is helping make fossil energy cleaner and nonfossil energy cheaper. We’re on the cusp of massive leaps in data center and power electronics efficiency. Tech and energy firms are rallying around common goals of producing more energy—ideally (although secondarily) with fewer emissions. We are also finally rethinking permitting and regulatory systems that didn’t serve us well.

Google and Amazon, for example, are pushing the adoption of grid-enhancing technologies like sensors and advanced conductor wire. We’ve known about these technologies for years, but utilities didn’t have an incentive to adopt them because they profit from erecting big plants.

Nvidia and Emerald AI, an energy software startup named as one of The Information’s 50 most promising startups of the year, are demonstrating that it’s possible to shift computing workloads from one data center node to another to avoid straining the grid and causing power bills to spike.

My next column will be about how the data center buildout has triggered capex supercycles in other industries. A powerful array of companies and institutions are angling to get what they want from this boom—and not just money—on different and longer timelines, and that’s why I believe it’s too simplistic to say the industry has created a single, speculative bubble.

In other news:

NextEra Energy said it is collaborating with Exxon on a new hyperscale AI campus equipped with carbon capture; Exxon identified two potential sites in Mississippi and Louisiana.
One of the companies thankful for AI’s insatiable power habit, advanced geothermal leader Fervo Energy, just raised a $462 million, oversubscribed Series E funding round. Fervo applies oil and gas fracking techniques to inject water and produce clean steam power in regions without natural geothermal resources. Google stepped in to help it sell power to utilities by backstopping the extra costs above local power rates.
BloombergNEFnow says U.S. data centers may need nearly 40% more power by 2035 than it predicted just a few months ago.
Shares of AI energy campus developer Fermi Energy fell sharply Friday after the company said its first prospective tenant terminated a $150 million deal to fund construction on Fermi’s audacious gas and nuclear megasite in West Texas.

Rowan Digital Infrastructure, a data center developer in the U.S. and elsewhere in the Americas, has explored a sale valuing it at $10 billion.

New From Our Reporters

At AI’s Biggest Event, Some Researchers Said the Field Needs an Overhaul

ChatGPT Nears 900 Million Weekly Active Users But Gemini is Catching Up

‘ChatGPT for Doctors’ Startup Doubles Valuation to $12 Billion as Revenue Surges

DeepSeek is Using Banned Nvidia Chips in Race to Build Next Model

UAE Fund MGX Quietly Becomes One of the Biggest Data Center Financiers

China Weighs Nvidia Chip Purchase in Emergency Meetings With Tech Companies