J.P. Morgan expects cumulative AI infrastructure spending to reach $5-$7 trillion by the end of 2030, making it the largest infrastructure buildout yet. James Ebert’s job is to figure out how to finance it. Increasingly, his job is about electricity.
“Power is the number one potential constraining item to the AI infrastructure build out,” said Ebert, executive director of the bank’s Security and Resiliency Initiative, speaking last week at the Great Transformation distributed energy conference in Bend, Oregon.
AI workloads are growing at two to three times the rate of everything else, he said. Data centers draw roughly 3-4% of U.S. grid capacity today. By decade’s end, J.P. Morgan sees that share reaching the mid-teens.
The strain on the energy supply is showing. Ebert pointed to PJM Interconnection, which has cited a 6-GW shortfall over the next 1 to 2 years.
Shortfalls are emerging across the grid, he said. That’s pushing more projects behind the meter.
Not surprisingly, energy is where he now spends most of his time, as AI infrastructure spending approaches $1 trillion this year alone.
“Nuclear is not distributed”
An audience member asked how much of the initiative’s capital would flow to distributed energy resources.
At first, Ebert listed gas generation, grid-enhancing technologies like reconductoring and superconductors, nuclear, solar, and renewables. The questioner cut in: “I have to say nuclear is not distributed.” He also noted that batteries can be installed more quickly than nuclear. “You could finance millions of batteries that could be installed in a year —rather than 10 years — with just people in this room.”
Ebert agreed. “Absolutely,” he said, noting that battery storage is an area where the bank is spending significant time — along with evaluating the supply chain risk behind it. “There’s an 80 to 90% dependency from China. If that gets cut off, what does that look like?”
His bottom line: “I think the data center build-out [is] going to require an all power approach.”
Much of his work, he said, lies in Washington: pairing private capital with public programs like the Department of Energy’s Energy Dominance Fund, the Office of Strategic Capital at the Department of War, and the CHIPS program at Commerce, seeking “a multiplier effect on public capital through private capital solutions.”
A 40-year-old grid meets a 10-year timeline
Ebert pointed out that lack of transmission in the US appears to be an AI bottleneck. (It’s also one of the reasons data centers are using more distributed energy.)
The average transmission line in the US is 40 years old; in China it’s 10 years old. “China does not have the same power problem that we do,” he said.
The very big numbers
Ebert walked through the bank’s bottoms-up math for the AI buildout. Roughly $1.5 trillion flows from hyperscaler cash. Another $800 billion or so arrives as equity or preferred equity. Investment-grade bond markets supply about $1.5 trillion more.
That still leaves $1.5 trillion to $3 trillion that must come from structured or leveraged finance. There, Ebert said, “there’s a real capital gap.”
He isn’t worried the money will fail to materialize. “The question we’re asking ourselves isn’t whether or not the AI infrastructure buildout is going to get financed. It’s going to be around what are the terms,” he said.
Those terms are tightening. “We’re not going to build ahead of the demand,” Ebert said — a shift, he noted, from where banks stood 24 months ago. Long-term offtake agreements now top the pile. The bank then scrutinizes development risk: fixed-price construction contracts, permitting, interconnection and power. Projects with investment-grade or hyperscaler tenants can secure 70-80% loan-to-value financing in the best cases. Neocloud tenants get 50-60%, with more equity required.
How long does the AI supercycle last?
That, Ebert said, “is a question we are asking ourselves daily at this point.”
His answer, for now, is that it will be at least several more years before the market reaches a more normalized state. He cited data center equipment backlogs and an application layer that remains very early.
“I think a year ago we would’ve said only a handful more years, but really we’re more bullish than ever,” he said.
