Oracle's $638 billion backlog just met its bill

Oracle reported the strongest quarter in its history on Wednesday, beat expectations on both lines, guided fiscal 2027 revenue up 34 percent, and watched its stock fall as much as 10 percent in extended trading. The number everyone will repeat from this report is $638 billion — the company's remaining performance obligations, the contracted future revenue sitting in its backlog, up 363 percent in a year. The number that actually moved the stock is smaller and appears later in the call: another $40 billion of debt and equity to be raised this fiscal year, on top of the roughly $48 billion Oracle raised in the year just ended. The market heard both numbers and decided the second one was the news. It was right.

Spend a minute on what was genuinely good, because it was genuinely good. Fourth-quarter revenue came in at $19.2 billion, up 21 percent, against expectations of $19.1 billion. Adjusted earnings of $2.03 a share beat the $1.96 consensus. Total cloud revenue grew 47 percent to $9.9 billion, and the part that matters most to the AI story — Oracle Cloud Infrastructure, the GPU-stuffed data-center business — grew 93 percent to $5.8 billion. Operating cash flow for the full year reached $32 billion, up 54 percent. There is no demand problem visible anywhere in this report. Anyone framing Wednesday's selloff as the market doubting AI demand is reading the wrong line of the filing.

What a backlog is, and what it is not

The $638 billion needs translating, because remaining performance obligations have quietly become the headline metric of the AI trade, and almost nobody who repeats them prices what they are. An RPO is a contract: a customer's promise to buy services in the future, often spread across five or more years. It is not revenue, not cash, and not collateral. It converts into money only if the customer still exists, still needs the capacity, and still can pay at the pace the contract imagined — and Oracle's grew by $85 billion in a single quarter, an acceleration the company attributes largely to enormous AI training contracts, including arrangements in which customers prepay for GPUs or even supply the chips themselves.

That last detail deserves more attention than it got on the call. When a customer prepays for GPUs, or shows up with its own, the backlog inflates by the full value of the arrangement while the underlying economics — Oracle's actual margin for hosting someone else's silicon — are thinner and lumpier than a classic software contract. The composition of a backlog matters as much as its size. A dollar of contracted database subscription and a dollar of pass-through GPU capacity sit in the same disclosure and are not the same dollar.

Then there is the concentration, which the filing does not itemize but the reporting around it does: a very large share of Oracle's AI backlog is widely reported to trace to a single customer — OpenAI, whose multi-year cloud commitment to Oracle was reported last year at around $300 billion. Oracle's future, in other words, is now substantially a bet on one counterparty's ability to keep raising and spending money at historic rates. OpenAI is currently burning billions a year, just filed confidentially for its own IPO, and this week was reported to be weighing significant cuts to its token prices to fend off Anthropic. Hold that thought.

A backlog is a promise about someone else's future cash. Capex is a certainty about your own.

The funding gap, in cash

Here is the mechanism the selloff priced. To convert that backlog into revenue, Oracle has to build the data centers first, and the building is brutally front-loaded. The company's net cash capital expenditure in fiscal 2026 was about $48 billion — against $32 billion of operating cash flow. For fiscal 2027 it guided to roughly $70 billion net, with gross reported capex expected around $90 to $95 billion once customer prepayments and equipment timing wash through. Put plainly: for the second year running, Oracle will spend roughly double its operating cash flow on construction, and the difference does not appear by magic. It is borrowed, or it is sold.

Both, in fact. The fiscal year just ended was financed with about $43 billion of new debt and $5 billion of equity. The new plan adds $40 billion more, including the $20 billion at-the-market share sale announced earlier this spring. The arithmetic of the AI build-out at Oracle now looks like this:

• Operating cash flow, fiscal 2026: about $32 billion
• Net cash capex, fiscal 2026: about $48 billion; guided to about $70 billion for fiscal 2027
• Raised in fiscal 2026: roughly $43 billion in debt plus $5 billion in equity
• Planned for fiscal 2027: about $40 billion more in debt and equity, including a $20 billion share sale
• Backlog those numbers are chasing: $638 billion, heavily weighted to a handful of AI customers

Management's defense is a return number: the chief financial officer told analysts the infrastructure business earns a return on invested capital in the high twenties at steady state. Taken at face value, that is an excellent return and a rational reason to lever up. But note the load-bearing phrase. "At steady state" assumes the GPUs stay rented at today's prices for their accounting lives, that the customer mix holds, and that the depreciation schedule — five to six years, on chips whose economic half-life is the subject of open argument across the industry — describes reality. Every one of those assumptions is also an assumption about the counterparty on the other side of the backlog.

The equity sale is the tell

Companies with Oracle's cash generation do not sell $20 billion of their own stock for fun; equity is the most expensive money a confident management can raise, which is precisely why issuing it carries information. After $43 billion of new debt in a single year, Oracle's leverage was already the most aggressive among the hyperscalers relative to its size, its credit spreads have been a running topic in the bond market, and the rating agencies' patience is not infinite. Choosing dilution over more leverage says the debt capacity is closer to its ceiling than the growth story implies. The stock falling on the announcement is not the market doubting Oracle's engineering. It is the marginal buyer recalculating how much of the next decade's backlog conversion was just sold to pay for the construction.

It is worth situating this in the wider trade, because Oracle is not an outlier so much as the leading edge. AI-related borrowing across the sector has roughly doubled year on year, the investment-grade index now carries more AI-linked paper than bank paper, and the build-out as a whole has shifted from being funded out of cash flow to being financed in the capital markets. Oracle is simply the company doing it with the smallest cash-flow cushion and the largest single-customer exposure — which makes it, for better or worse, the cleanest expression of the thesis available on a public exchange.

The price of a token sets the value of the backlog

Now connect the week's two Oracle-adjacent stories, because the market certainly did. On Wednesday evening, the Wall Street Journal reported that OpenAI is weighing significant cuts to what it charges for tokens, anticipating a price war with Anthropic as both approach the public markets. Token prices are the revenue side of the GPU rental equation: what an AI company can charge for inference is, one step removed, what it can afford to pay Oracle for the computers that produce it. If the unit price of intelligence is about to fall — and competition plus IPO-window pressure says it is — then the revenue per GPU-hour that underwrites a 'high twenties' return on $70 billion of annual construction is a moving target, and it is moving in the wrong direction.

This is the quiet structural risk in the whole AI-infrastructure trade: the landlords are adding leverage at exactly the moment the tenants are discussing rent cuts. Nothing about that is fatal. Falling token prices can expand the market enough that volume swamps price, which is the genuinely bullish case and has respectable precedent in every computing cost curve since the mainframe. But that case requires demand elasticity to show up on schedule, every year, for the life of the debt — and the debt, unlike the thesis, has coupons with dates on them.

What breaks if the story is merely late

The downside case for Oracle does not require AI to fail. It only requires the backlog to convert slower than the construction bill arrives. A counterparty that stretches its commitments, a renegotiated contract, a year in which token prices fall faster than usage grows — any of these widens a funding gap that is already being bridged with shareholder dilution, while $70 billion of partially-built data centers depreciates on the balance sheet regardless of whether the GPUs inside are earning. The asymmetry is the point: if everything goes right, equity holders get a utility-grade growth story they paid a software multiple for. If the timeline merely slips, they get the dilution, the leverage, and the depreciation, and someone else gets the contracts rebid at lower prices.

The company and the trade, as ever, are separable. The company had a superb quarter, sells something the world demonstrably wants, and is run by people who can read a bond covenant. The trade is a leveraged, concentrated position on one customer's fundraising calendar and the future price of a token, financed in part by selling the equity of the company making the bet. Wednesday's after-hours tape was not a verdict on the quarter. It was the sound of the market noticing that the backlog everyone repeats comes, like the data centers behind it, with a bill attached — and that for the first time, the bill is being mailed to the shareholders.