Wall Street banks are racing to finance AI data centers, as deals swell into the tens of billions, forcing a rethink of how these projects are funded.

“If you can’t invest a billion dollars, we don’t even want to talk to you,” said Adam Lewis, a managing director at Citizens, a regional lender that has emerged as a key player in the sector. Just a few years ago, a $100 million financing was a milestone; today, it’s a rounding error.

For Lewis, that billion-dollar floor reflects the rising cost of land and electricity, which has pushed these projects beyond the limits of traditional commercial real estate loans and into the realm of large-scale infrastructure finance.

As deal values surge, banks are focused on seizing what could be Wall Street’s largest-ever financing opportunity. Over the past two years, lenders including Morgan Stanley, Goldman Sachs, and JPMorgan have formed integrated teams across disciplines to become fluent in the mechanics of how data centers are actually constructed.

Citigroup estimates the buildout could require $3 trillion by 2030, according to an internal memo sent in late February by leaders of the firm’s investment banking unit. In the memo, senior bankers from across investment banking, corporate banking, and financing said that Citi would establish a dedicated AI infrastructure group to break through internal silos and evaluate “all pockets of capital” as deals grow larger and more complex.

The sheer scale of the AI buildout is beginning to exhaust the cash reserves of the world’s largest tech giants. While hyperscalers cannot afford to fall behind in the infrastructure race, the costs have become too great to carry on their own balance sheets. To Fred Turpin, the global chair of investment banking at JPMorgan, this represents the “largest investment cycle in the history of capitalism.”

To bridge that gap, Turpin helped organize a firmwide working group that pairs technology and energy experts with bankers versed in private capital markets. The approach allows the bank to jump-start projects using its own balance sheet before connecting them to “long-term” capital from sovereign wealth funds, pension funds, and dedicated infrastructure investors looking for stable, generational returns.

Integrated teams

To put together the unprecedented amount of money to build AI infrastructure, bankers are drawing on multiple sources of capital, from bank loans and bonds to private credit and institutional investors, often assembled into a single structure from the outset.

At Goldman Sachs, the shift has taken shape inside its Capital Solutions Group, a unit formed last year to bring together origination, structuring, and capital distribution as deal sizes and complexity have grown. The group pulls in bankers from across investment-grade and high-yield debt, infrastructure and real estate financing, and equity capital markets, allowing the firm to consider multiple financing options at once.

“We’re elbow to elbow with the bankers that cover sponsors so that we can ensure a direct line between our origination efforts and distribution efforts to financial sponsors,” said John Greenwood, a partner who serves as global head of the infrastructure and real asset finance group within Capital Solutions.

At Morgan Stanley, Richard Myers and William Graham, two top investment bankers, are members of a data-center-focused task force launched in 2024. Last year, Myers and his team arranged a $2.6 billion financing for CoreWeave that used Nvidia chips as collateral. They later pioneered a first-of-its-kind $27 billion bond deal for a joint venture between Meta and Blue Owl. That work increasingly requires bringing together specialists from across the bank — from power and project finance to real estate — to arrange multiple sources of capital.

And Graham, the firm’s global cohead of leveraged finance, has led a $3.2 billion senior secured note offering for TeraWulf and a $2.35 billion raise for Applied Digital — two specialized infrastructure firms that have pivoted from crypto mining to hosting the high-density power loads required for AI.

New vocabulary

Unlike traditional corporate financings, data centers sit at the intersection of real estate, energy, and technology, which means bankers have to weigh not just financial risk — but whether a project can actually be built, powered, and brought online as planned. Bankers said they’ve had to become fluent in a new language — the lexicon behind how these massive projects are built.

“We can read electrical diagrams and mechanical diagrams and understand land use permits and power configurations,” said Lewis, the managing director at Citizens, whose team of more than 30 bankers focuses on advising, structuring, and financing data center projects. Bankers are now required to understand what could delay or derail a project, and to give investors confidence that it will actually come online as planned.

“Most of us just assume it happens magically in some ephemeral thing called the cloud,” said Scott Wilcoxen, who leads digital infrastructure investment banking at JPMorgan. “But physically, what that actually means is there is effectively an unbroken physical connection between individual users and the data sources.”

This technical knowledge is ever more important as bankers say projects are increasingly constrained by limits on power, equipment, and labor. But those constraints don’t appear to be cooling demand, raising questions about how far the buildout can stretch — and what it will take to sustain it.

Goldman’s Greenwood noted that in a recent meeting with a client, someone in the room used a surprising adjective: “terrestrial.”

“I was in a meeting last week, and they were talking about terrestrial data centers,” he said, suggesting the next frontier could be “on the bottom of the sea, or in space.”