A MESSAGE FROM OUR SPONSOR

McKinsey’s latest State of AI in 2025 report makes one thing clear: the next wave of AI advantage will be won by leaders, not algorithms.

High-performing organizations are three times more likely to have leaders who own and model AI usage, weaving it into strategy, culture, and execution.

The question is no longer if AI delivers value—it’s how leaders will capture what’s next.

Read the report at McKinsey.com/StateofAI

“There’s been a lot of talk about an A.I. bubble. From our vantage point, we see something very different. … Our customers’ financing is up to them.”
—Nvidia C.E.O. Jensen Huang on an earnings call after yet another blowout quarter. Four unnamed customers were responsible for over 60 percent of Nvidia’s total revenue. The company’s stock was down 3 percent at today's closing bell.

And now for the main event…

Last week, Bloomberg reported that two massive new data centers are sitting empty and unused in California—some nearly 100 megawatts of capacity going to waste. The local utility simply doesn’t have sufficient power to plug them in. Meanwhile, a report published earlier this month from AEP Ohio, a large electric utility, explained that it had received requests to study 36 new data center sites—and that the necessary regional upgrades to bring them online wouldn’t be ready until the last quarter of 2031 at the earliest. Similar reports have emerged from California, Arizona, Illinois, and elsewhere.

Over the past year, reports like these have become increasingly common as A.I. hyperscalers and startups around the country scramble to build data center capacity. For overburdened utility companies, the energy demands have become increasingly untenable. Already, the lack of adequate supply has fueled the industry’s push into nuclear energy and even outer space, and is quietly emerging as one of the field’s most pressing existential challenges. A Deloitte report from earlier this year predicted that energy demands from data centers would rise fivefold by 2035, to around 176 gigawatts. In 2023, total U.S. electrical capacity was around 1,200 gigawatts.

Naturally, strains across the energy grid are not uniform. Some states, like Ohio—by virtue of local laws and geographic amenities—have become hotbeds for data center development. According to one estimate, Ohio is home to more than 200 data centers, as co-locators such as Cologix and Vantage and hyperscalers including Microsoft and Google—but mainly Amazon Web Services—have nestled into the terrain. A.W.S. invested $6 billion in data center infrastructure in the state in 2022, $7.8 billion in 2023, and $10 billion last year to build out “advanced cloud infrastructure and compute power.”

A MESSAGE FROM OUR SPONSOR

McKinsey’s latest State of AI in 2025 report makes one thing clear: the next wave of AI advantage will be won by leaders, not algorithms.

High-performing organizations are three times more likely to have leaders who own and model AI usage, weaving it into strategy, culture, and execution.

The question is no longer if AI delivers value—it’s how leaders will capture what’s next.

Read the report at McKinsey.com/StateofAI

When I asked an Amazon spokesperson whether the Ohio data centers are maxing out on available power, I was told that all of the company’s Ohio data centers, including those currently operational and under construction, have “secured” power. This person added that most of this power is being drawn from the grid, but that the power demands at one site, in Hilliard, will be augmented by on-site fuel cells while transmission upgrades are completed. As to whether there will be sufficient energy for Amazon’s planned data centers, the spokesperson said that Amazon continues to invest in renewable energy projects and work with local utilities to plan for growth.

While some industry leaders have obfuscated the scope of the energy consumption problem, others have already addressed it publicly. On a podcast recently, Microsoft C.E.O. Satya Nadella said that the “biggest issue we are now having is not a compute glut, but the power, and the ability to get the builds done fast enough close to power.” He continued: “If you can’t do that, you may have a bunch of chips sitting in inventory that I can’t plug in.” I asked Microsoft whether this was a widespread issue, but the company declined to comment. Likewise, Oracle declined to comment; Google never returned a request for comment.

Ahmed Mohamed, a professor of electrical engineering at the City College of New York, told me that he expects “this pressure will likely intensify through 2026 as both data center growth and broader electrification continue to accelerate.” As large energy users approach and even exceed local grid capacity, he added, it will surface broader challenges surrounding aging infrastructure and slow permitting. “The result can be localized congestion, higher costs, and reduced reliability margins, especially in fast-growing load pockets,” he said.

I asked Suman Kanuganti, the founder and C.E.O. of Personal AI, whether he’s seen evidence of the U.S. maxing out on available power. He responded that he believes it’s happening now, but that it’s actually a competitive opportunity for his company. After all, Personal operates small language models, which are significantly more energy efficient and can produce more tokens per kilowatt-hour than larger rivals.

Still, Kanuganti said, the Silicon Valley attitude of “innovation at any cost” is not sustainable. The cloud, he noted, “is not designed for high-performance G.P.U.s,” and the power to operate both training and inference in the cloud at scale “doesn’t exist.” The endgame, he believes, will involve a shift into more efficient edge computing (i.e., using local servers), while massive model training will remain cloud-based. He expects the market to “start pivoting away from just wasting money and electricity on a dead horse,” and start looking in a different direction. For him, that means combining S.L.M.s and L.L.M.s to serve the industry’s demands.

One of the biggest challenges facing the utility companies was laid out in a recent report by Siemens, which noted that “forecast inputs are often overestimated or unreliable due to developers’ optimistic projections.” AEP Ohio echoed this sentiment in a recent report, which said that rising data center energy demands have become “more disconnected from underlying economics.” In other words, the data centers are not only hoovering up large quantities of power now, but the companies are requesting access to future power, requiring infrastructure that, in many cases, does not yet exist.

Aaron Ruby, a communications director at Dominion, the largest electricity company in Virginia, told me that the company has structures in place—up-front payments, mostly—to hopefully suss out real demand from overly optimistic speculation. They also have a proposal pending before the state that would significantly increase the amount of up-front financial obligations from data center customers ($1.5 million per megawatt of requested capacity) and require new customers to enter into 14-year contracts with Dominion while paying for 85 percent of the power they’ve requested regardless of how much they end up using. AEP Ohio made a similar request to its Public Utilities Commission in May 2024, which was approved a few months ago.

Dominion’s proposal reflects the fact that it expects data center power demands in Virginia to more than quadruple in the next 10 to 15 years. Ruby called the power demand from his current customers “unprecedented,” and said that roughly 80 percent of that demand is coming from data centers. To serve this total need, Dominion will have to roughly double its generation capacity, vastly increase its transmission capacity, and continue building out new grid infrastructure, a process Ruby expects will “continue pretty much continuously for the next couple of decades, in order to serve all this growth.”

How might this impact normal people, who just want their toasters to work and televisions to turn on? Dr. Brandon Grainger, a professor of electrical engineering at the University of Pittsburgh, said that he doesn’t expect data center demands to cause a country-wide blackout. Manufacturers, he said, perform “extensive system studies to make sure that nothing happens as a result of putting” data centers on the grid, and he underscored that utility businesses are “conservative” and focused on safety. But, he added, it’s going to take time for the utilities to get the grid ready to bring all these data centers online. Of course, by the time those lights finally start flickering on, the A.I. industry might well look like a vastly different place.

Jeff Bezos, the world’s third-wealthiest man and Bo Burnham’s muse, has quietly launched a new startup called Project Prometheus (very subtle, Jeff), focused on developing A.I. tech for manufacturing, with a focus on computing, aerospace, and automotive. The project has reportedly secured $6.2 billion in funding and has around 100 employees. Bezos will serve as its co-C.E.O. [New York Times]

The European Commission has seemingly bowed to some of the anti-regulatory pressure it faced since passing its landmark A.I. Act. The commission published a proposal yesterday that, if enacted, would streamline the act’s requirements and kick its more heavy-handed rules to 2026 or 2027. [Le Monde]

That’s all for today. I’ll see you next week.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St., New York, NY 10006