AI Cloud Spending Is Splitting Companies Into the Funded and the Stretched

AI infrastructure spending has been booming over the past couple of years, and the biggest question for tech sector investors isn't how much revenue these companies you hear about have booked, but who bears the risk while they borrow billions.

Oracle closed the first quarter of its fiscal 2027 with $664 billion in remaining performance obligations, a backlog bigger than its own market cap. It also had $1.4 billion in interest expenses that quarter as its total notes payable and other borrowings reached $125.3 billion.

A backlog larger than the whole company and a debt bill growing faster than revenue point to the real question behind every AI infrastructure stock: When a company signs a 15-year contract to build data centers for a hyperscaler, who's fronting the cash to pour the concrete, and what happens if the contract pays out slower than planned?

The AI build-out has split the companies financing it into two camps. One group is borrowing against decades of cash flow and established business lines with broad customer bases. The other is borrowing against contracts that didn't exist 18 months ago, and the debt is often collateralized by the GPUs and data centers themselves.

Borrowing against decades of cash flow

Oracle is using its long-established enterprise software business as collateral for its AI bet. Its RPO grew on multiyear cloud deals with OpenAI and other AI labs, and the company now borrows tens of billions of dollars a year to build capacity.

Its interest expense grew more than 50% year over year, and spending is outpacing cash inflows: Oracle generated $23.1 billion in operating cash flow last quarter, but capital expenditures of $28.5 billion to build AI data center capacity left it free-cash-flow negative, per Business Insider. The honest risk isn't solvency. It's execution. Oracle has acknowledged that much of its $664 billion backlog won't convert into revenue for years; shareholders are betting it can deliver the capacity before the market's patience runs out.

Borrowing against contracts, concentrated in a few hands

Neoclouds CoreWeave and Nebius built their substantial backlogs almost from scratch over the past two years, largely by signing deals with a handful of very large customers. CoreWeave's revenue backlog reached $104 billion last quarter, built on commitments from Microsoft, OpenAI, and Nvidia.

But Microsoft alone accounted for roughly 67% of CoreWeave's 2025 revenue, and as of June 30, CoreWeave carried $35.1 billion in total debt, including $3.7 billion in non-recourse debt secured directly against GPU collateral. If a major customer renegotiates, lenders holding that collateral get paid before shareholders see anything.

Nebius shows a similar pattern with a sturdier cushion. Its RPO stood at $33.6 billion in the first quarter, with management citing more than $40 billion in additional contracted revenue from investment-grade customers such as Microsoft.

For both, the contracts are real and typically take-or-pay. The risk is that a handful of hyperscalers now hold outsize leverage over smaller providers that are dependent on them for most of their revenue.

Beware, though: Both tickers have been selling off in recent days as investors focus on their high-leverage, debt-funded build-outs, and rising financing costs.

Borrowing against the data centers themselves

Applied Digital and Iren both pivoted from crypto mining into leasing data center capacity to AI hyperscalers, financed heavily through non-recourse debt tied to their physical assets. Applied Digital has built roughly $36.2 billion in contracted revenue over 15-year base lease terms, structured as take-or-pay, non-cancellable agreements with CoreWeave and investment-grade hyperscalers. Long-term debt climbed from roughly $5 billion to $6.3 billion in a single quarter this year, per Reuters. For a company with a market cap under $7 billion, this is significant leverage, even with billions in leasebacks backing it.

Iren signed a roughly $9.7 billion, five-year contract with Microsoft for GPU capacity at its facility in Texas, plus a $3.4 billion arrangement with Nvidia. It has financed its data center build-out partly through investment-grade-rated, non-recourse GPU financing and convertible notes, a structure that protects Iren's broader balance sheet if one deployment goes wrong. For its fiscal 2027, which began in July, it's guiding for $25 billion to $30 billion in capital spending. All of this points to the need for lots of external capital for success.

So are these companies buys

So, what does all this mean for people who own these tickers or are considering buying them? In short, there's no reason to dismiss these companies' backlogs. The compute capacity shortage in AI is real, and hyperscalers don't sign contracts this large casually or without full intent to execute on them. But booked revenue isn't cash, and where these five companies differ is in how they are financing the gap.

Oracle has its database and applications businesses to fall back on. CoreWeave and Nebius have binding contracts, but their small customer bases leave them exposed if even one major buyer renegotiates. Applied Digital and Iren built their backlogs rapidly, borrowing against the data centers and GPUs they need to deliver.