FOLLOW US

Behind the breathtaking numbers is a financial paradox: the institutions underwriting AI’s future may also be exposed to its disappointment.

As a rule of thumb, the money comes before the proof

It could start in a location that's more a construction site than a lab: a fenced, Texas field, an awaiting power transmission line, and a financing plan that runs in the billions of dollars.

AI models require servers in order to answer a question. Now before those servers can be put to use, they require chips, cooling, fiber connections, water, permits and electricity. Until someone borrows the money and then you build it, none of that can bring in any revenue.

There, Wall Street has stepped in. In 2026, the five biggest U.S. hyperscalers, which are large cloud and technology firms, will be spending $697 billion on capital expenditures, according to J.P. Morgan. In what's becoming a trend in the AI industry, Nvidia, the maker of much of the computing infrastructure, is among a group of companies—including Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR—that are developing platforms to mobilize over $500 billion for AI infrastructure.

The terminology used in the deals is quite impressive. “In AI, compute is revenue” said Jensen Huang, Nvidia's CEO. “Modern computing is a ‘scarce, mission-critical asset class” says Apollo President Jim Zelter. J.P. Morgan global co-head of Investment Grade Finance, John Servidea, said that AI financing is the “biggest secular theme in our professional lifetimes”

The potential is vast. And the question that looms over it: What if the infrastructure is constructed before the profits come in?

A new type of asset, old financial mechanism

It's a complicated financial structure, but the idea is quite familiar. The flexibility of corporate debt is that the company can borrow it off their balance sheet and use the money for any purpose: for example, data centers, energy systems, networking and computing. Project-level debt is more focused. It is fixed to a certain asset and is typically based on visible cash flows, long term contracts or an agreed tenant.

A recent financing for Hut 8's Beacon Point data center in Texas illustrates the way investors are seeking to eliminate uncertainty. J.P. Morgan led the $4.25 billion bookrun for the senior secured notes. It was reported that the facility was 100% pre-leased under a triple-net lease for 15 years with a highly rated tenant. The transaction was 95% loan to cost, which is a high ratio of debt to the cost of the project.

But the funds are no longer flowing solely from the pockets of technology firms' cash reserves. According to Goldman Sachs estimates, AI financing now makes up almost a quarter of gross investment-grade U.S. issuance, and investment in AI is nearing $600 billion this year, reported Fortune. The same report outlined a number of security firms looking for long-duration assets to meet long-term obligations, including insurers, pension capital and private-credit funds.

That is the paradox: a growth engine with a lot of corporate money is getting a lot of credit from loans from banks, asset managers, insurance portfolios and retirement funds.

Scott Wilcoxen, J.P. Morgan's global head of Digital Infrastructure Investment Banking, highlights the tangible nature of the abstract financial discussion. “Data doesn’t move in a vacuum, it moves over a medium” he said.

The physical world does not expand as quickly as a spreadsheet. But with capital available, projects can be delayed for years by power connections, equipment, workers, water and cooling systems, and permits.

“Accelerating time to power is a top priority for many of our clients, but the grid infrastructure isn’t there” Wilcoxen said. “There’s a lot of work to do.”

The resulting bottleneck poses a financial risk: A project might be completely financed and yet not be operational due to technological advances and the time that interest runs up.

Jane Sydenham, a senior investment manager at Rathbones, captured the concern in a question to the BBC: “The worry is that more and more money is going into these projects. Are they all going to earn the right return for the future?”

The risk is now off the balance sheet

The most critical aspect of the narrative might be the one that is not obvious to investors. The combined leases for hyperscalers' data centers, research facilities, offices and equipment increased to $1.5 trillion, from about $200 billion five years ago, according to CNBC. Approximately $1 trillion was referred to as “uncommenced” or not yet accounted for in accounting statements.

Leases may work well, but they can also complicate leverage comparisons. Even if the economic burden is the same, it may seem different on paper between a firm with an owned facility and a firm that is leasing the facility under a joint venture.

The Bank of England has said this lack of transparency is important. It projected that the 75 investment-grade corporate bonds that are most closely associated with the AI revolution already represented 14.5% of the JULI index of investment-grade corporate bonds, ahead of 11.5% in 2020. It also estimated that new general lending to AI companies in leveraged finance and private-market debt had risen to almost $100 billion year to date in 2025. Some of that credit can be passed down to banks and non-bank investors via “collateralized loan obligations”.

The risk can lie with the arranging bank, a private credit fund, an insurer, a pension vehicle or a bond investor if a large AI project goes wrong. However, a complex chain may have a terminal exposure that is not easily identified.

The hedge-fund fiasco serves as a warning but not an indicator of a general industry meltdown. Situational Awareness had about $10 billion in assets after losses on "concentrated, leveraged" positions, CNBC reported. According to JPMorgan CEO Jamie Dimon, the level of margin debt is “pretty high.”

However, it is necessary to hear the counter argument. “While elevated earnings expectations are the biggest threat, not leverage alone, it is clear that this is an issue that will need attention” said Sahil Mahtani with Ninety One. There was no evidence to suggest that leverage was an inherent systemic risk for hedge funds, the Alternative Investment Management Association said.

Servidea presented a similar representation of the whole financing market. He adds investors are not “blindly buying everything” as such. They're looking at credit quality, relative value, pricing and downside protection. “The red lights would start flashing if it felt like there was blind euphoria” he said. “That’s not what we’re seeing”

Bubble, revolution—or both?

The response could be based less on the authenticity of AI as a fact than on how fast the financing of AI evolves. The technology has the potential to deliver real productivity benefits, despite the loss of money by some projects and lenders. Transformations do not ensure that the assets created in their name bring a pleasing return, nor can there be a bubble in a revolution.

According to the Bank of England, AI firms have captured 67% of year to date S&P 500 returns in 2025 and 44% of the market capitalisation as of the end of 2022. It also estimates that AI infrastructure spending could cost trillions of dollars over the next five years, with Morgan Stanley Research predicting $2.9 trillion between 2025 and 2028, of which 1.5 trillion USD will be from outside the U.S., with only $1.4 trillion from within the U.S. These are not definite results, which are still to come.

A disappointment would result in lower technology valuations, wider bond spreads, margin calls, lower data-center revenues and private-credit losses and higher costs for banks.

But it's possible that the opposite occurs. Today's debt could finance infrastructure as significant as the previous networks, factories and railways, if the systems evolve rapidly, if the companies can charge for the use of their systems, and if the new facilities are filled with demand. KKR's Joe Bae and Scott Nuttall explained that it's easy to dream of becoming a great player, but it's difficult to actually do it.

That's the litmus test for the banks. They are making assumptions about customers, productivity and cash flow in the future, power and chips. If this is true, the capital is able to establish a new economic base. The 2 failures combined may spread losses in a rapid manner with the same equipment.