Oracle’s big AI dreams are freaking out Wall Street

Oracle’s AI “power” play depends on everyone believing that the lights will stay on long enough for the money to show up. The company wants to be the one keeping the world’s biggest AI models plugged in — racing to build out infrastructure that OpenAI, Meta, and others say they need — and Oracle has the contracts, the backlog, and the talking points to prove it.

But this quarter’s earnings, for the second quarter of the 2026 fiscal year, just made it painfully clear that the grid is wired to a very nervous bond market — and that the AI boom Oracle is selling is being financed on credit first, cash later.

On the surface, Oracle’s demand story looks spectacular. Cloud revenue is growing in the mid-30% range. Infrastructure is up about 68%. And remaining performance obligations (RPO) — the IOU pile that’s supposed to justify this entire super cycle — exploded 438% year-over-year to $523 billion in the quarter, driven by long-term AI infrastructure commitments from Meta, Nvidia, and more. A growing share of that backlog is supposedly going to show up sooner rather than later; Oracle has said 40% is now expected to convert into revenue within 12 months, up from 25% a quarter ago.

On paper, this is the kind of order book people were promising when they said AI would pay for itself.

So all good, right? Wrong.

Oracle stock fell about 16% in early-morning trading Thursday after it reported earnings late Wednesday, wiping out something in the ballpark of $70 billion from the company’s market value as investors instead saw a funding story they didn’t love. The same earnings that touted the massive backlog also showed that, over the quarter, Oracle has burned more than $10 billion of free cash, even as AI-branded cloud sales surge.

In other words, the build is outrunning the cash.

The company is busy telling the market that its AI contracts are real and enormous — but the capacity has to exist now, while the money that’s supposed to justify it drips in over years and years. So the company is plugging the gap with its balance sheet. Its debt load has swelled to roughly $100-$112 billion — after a recent $18 billion bond sale and what looks like close to $38 billion of additional loans and structured financings tied to the company’s Stargate data-center build — pushing its debt-to-equity ratio to a level that looks more like a utility or project-finance vehicle than a traditional software vendor.

For now, at least the company’s income statement still looks OK. Revenue grew 14% to about $16 billion, just below Wall Street expectations. And Oracle beat the Street’s earnings expectations with an adjusted earnings per share of $2.26, although that’s largely because of a $2.7 billion gain on the sale of Oracle’s stake in Ampere to SoftBank. Executive chairman and chief technology officer Larry Ellison used the sale to tell everyone on the earnings call that Oracle is “chip neutral” and happy to deploy whatever CPUs and GPUs customers want. The cosmetic message is that Oracle is surfing the AI wave on skill and flexibility. The financial message is that a one-time asset sale is flattering the quarter at the exact moment the core business is bleeding cash to keep the GPU arms race going.

Underneath Oracle’s “strong quarter with a small revenue miss” storyline lies negative free cash flow, narrowing margins, a sharply higher capital-spending plan — full-year capex bumped to roughly $50 billion from about $35 billion just three months ago — and a balance sheet that’s doing most of the work to keep the AI story turning the pages. That massive 438% RPO jump is only comforting if the company can build the infrastructure without overstretching its credit.

Oracle’s AI story looks less like “software margins forever” and more like “utility economics with a whole lot of leverage.”

Moody’s has already moved Oracle’s outlook from stable to negative, tying that to the AI buildout. The firm expects debt to rise faster than EBITDA and emphasizes that free cash flow is likely to stay negative “for an extended period before reaching breakeven.” Credit analysts now put Oracle’s debt-to-equity ratio in the 450%–500% range, compared with something closer to 30%–50% for Microsoft and Amazon. Oracle doesn’t have Microsoft or Google’s cash engines or net-cash balance sheets. So if this is how stretched the numbers look at a company with a $523 billion backlog and marquee AI names on the hook, smaller clouds, GPU lessors, and sovereign-AI hopefuls are going to have a harder time convincing bond investors that their own capex binges will pay off.

Oracle matters because it’s about as clean a test case as public markets are going to get. It has a cloud business that’s growing like it means it, big AI partners, and a backlog with a staggering anchor: more than $300 billion of committed compute spending from OpenAI over five years, starting in 2027. (Ignore for a second that Oracle is already underwater on the deal.) If any non-hyperscaler should be able to show a self-funding AI build, it’s this one. But it just showed the opposite.

That’s why this quarter reads like a dress rehearsal for an “AI bubble — bond market first” scenario. On one side, there’s a half-trillion-dollar backlog, double-digit cloud growth, and big customers signing big contracts that stretch out to the end of the decade. On the other side, there’s free cash flow that has gone deeply, deeply negative, a debt stack marching toward 12 digits, and a credit rating agency already running the math on how long that combination can stay investment-grade.

The generous read is that Oracle is simply early, effectively front-loading the infrastructure so it can lock in tenants for a decade. The harsher read is the one that has traders and analysts circling AI as the purest AI-debt proxy. If this quarter is what the numbers look like for a supposed AI winner, the rest of the field has to lean just as hard on its balance sheet to chase GPUs — or admit that the AI infrastructure supercycle is going to be smaller, slower, and less predictable than the trillion-by-2028 pitch decks have been promising.