Oracle just posted the biggest backlog number in company history, and investors sold the stock anyway. That’s the whole story in one sentence, and if you’ve been following the AI infrastructure trade, you already know why: a record backlog doesn’t pay the data center bill. It just promises that someone will, eventually.
Here’s what actually happened, what it means, and why the gap between “biggest number ever” and “stock fell 7-10%” tells you more about where AI infrastructure spending is headed than the headline figure does.
The Numbers, Straight Up
Oracle reported fiscal fourth-quarter 2026 revenue of $19.2 billion on June 10, up 21% year over year. Total cloud revenue rose 47% to $9.9 billion, with Oracle Cloud Infrastructure revenue jumping 93% to $5.8 billion and cloud applications revenue up 10% to $4.1 billion. Non-GAAP earnings per share came in at $2.11, up 24% — or $2.03 excluding one-time investment gains, still a 20% jump. Remaining performance obligations, Oracle’s term for contracted future revenue, hit $638 billion at the end of the fiscal year, up from roughly $138 billion twelve months earlier. That’s a 363% increase in twelve months. Most companies don’t grow revenue 363% in a decade, let alone backlog in a year.
So why did shares drop? Despite the record numbers, Oracle’s stock fell 7-10% in after-hours trading as capital spending worries overshadowed the headline figure. Everything below explains that disconnect.
What RPO Actually Means (And Why $638 Billion Is a Bigger Deal Than It Sounds)
RPO isn’t revenue. It’s a promise of future revenue contracts signed, not cash collected. Think of it as a massive order book that hasn’t shipped yet. The reason analysts obsess over it for cloud companies is that it’s the clearest signal of demand before that demand shows up on the income statement.
Oracle’s $638 billion in cloud commitments now exceeds both Alphabet’s and Microsoft’s RPO figures — Microsoft sits at $627 billion, Alphabet at $460 billion. Oracle, a company that for most of the last two decades got lumped in with “legacy enterprise software,” just out-backlogged the two biggest cloud players on the planet. That’s not a small flex. It’s also not free money sitting in a vault.
Here’s the part that gets buried in most coverage: over half of that backlog won’t convert into revenue for more than three years, with only 12% expected to land within the next 12 months and another 34% landing between 13 and 36 months out. So when you see “$638 billion,” mentally split it: a little over a tenth shows up as cash relatively soon, a third shows up over the next couple of years, and more than half is a multi-year bet that has to survive contract renegotiations, customer solvency, and whatever AI demand looks like in 2029.
I’ve sat through enough of these earnings calls over the years to know the pattern: a headline backlog number gets the press release treatment, and the maturity schedule gets buried three paragraphs into a footnote. That maturity schedule is the actual story.
Why the Stock Dropped on a Record Quarter
This is the part that confuses people who only read the headline. Oracle posted record revenue, record EPS growth, and a record backlog and the stock still fell. The market wasn’t reacting to demand. It was reacting to cost.
Oracle’s fiscal-year capital expenditures came in at $55.7 billion, alongside roughly $23.7 billion in negative free cash flow, with plans to raise another $40 billion during fiscal 2027. Put plainly: Oracle is spending tens of billions building data centers faster than the contracted revenue can reimburse it, and it’s borrowing to cover the gap. That $40 billion raise comes on top of $48 billion in debt and equity Oracle already raised during fiscal 2026.
There’s a nuance worth knowing here, because it changes how scary this actually is. A meaningful chunk of the recent backlog growth came from large AI contracts where customers either prepaid Oracle for GPU capacity or supplied their own hardware, and that prepaid-plus-customer-supplied portion now totals $75 billion. That’s customers effectively co-funding Oracle’s buildout which reduces how much Oracle itself has to borrow, even as it complicates the accounting (some of that capex shows up on Oracle’s books due to timing and ownership structure even though customers fronted the cash). Reported capex is actually running $20 billion to $25 billion higher than it otherwise would because of these customer prepayments and timing effects. CFO Hilary Maxson, who joined Oracle in April after a stint leading finance at a capital-intensive infrastructure business, told analysts on the call that Oracle doesn’t expect to take on additional debt funding for the remainder of calendar 2026 — though she flagged that fiscal 2027 gross margin will step down as new data centers ramp up before hitting full contractual revenue.
The honest read: this isn’t a company in trouble. It’s a company spending at a pace that makes even bullish investors nervous, funded by a mix of debt, prepayments, and a bet that the AI compute shortage doesn’t ease up before the bills come due.
The OpenAI Factor
You can’t talk about this backlog without talking about where most of it came from. Most of the recent backlog growth traces back to a small number of enormous AI contracts, including a reported $300 billion, five-year agreement with OpenAI signed last year. That single deal is doing a lot of the heavy lifting in that $638 billion figure.
Here’s the part that should make anyone tracking this nervous: OpenAI reportedly remains unprofitable, and the company filed confidential paperwork for an initial public offering just days before Oracle’s earnings report. Oracle has effectively bet a third of a trillion dollars in future revenue on a company that’s still burning cash and is only now testing public markets. If OpenAI’s growth stalls, or its IPO underwhelms, or compute demand from ChatGPT plateaus, Oracle’s “biggest backlog in the industry” starts looking like concentration risk instead of a moat.
This is the trade-off nobody puts in the headline: massive backlog growth and customer concentration risk are often the same fact, viewed from two angles. A $300 billion contract with one customer is either the best deal of the decade or the biggest single point of failure in the balance sheet and right now, nobody outside Oracle’s finance team really knows which.
Where All This Compute Is Actually Going
Step back from Oracle for a second and ask the bigger question: why does anyone need this much compute? The honest answer is that AI workloads stopped being chatbots a while ago. Most of what’s eating GPU capacity now is autonomous agents software that plans, acts, and loops on its own instead of waiting for a single prompt and reply.
If you’ve played with frameworks like Agent Zero, you’ve felt this firsthand. A single autonomous agent running multi-step research, browsing, and tool calls can burn through more compute in an hour than a thousand basic chat queries. That’s a big part of what’s actually driving demand for facilities like the ones Oracle is racing to build. If you want to see what that workload looks like in practice, our Agent Zero AI guide walks through how these agents actually consume resources, and our Docker installation walkthrough for Agent Zero shows how fast a local setup scales once you start running multiple agents in parallel.
The job market is reacting to this shift too — and fast. Demand for agentic AI roles has surged right alongside this infrastructure spending, because companies need people who can actually design and supervise autonomous systems, not just prompt a chatbot. On the security side, the rise of always-on agents that can take real-world actions has created an entire new category of AI red team jobs, focused specifically on stress-testing what happens when an autonomous agent gets bad instructions or compromised tools.
It’s not just chat-style agents either. Compute-hungry niche applications are scaling fast machine learning models built for sports betting are a good example of a use case most people don’t associate with “AI infrastructure,” but that quietly consumes enormous GPU cycles running constant in-game probability recalculations. Multiply that across dozens of similar verticals fraud detection, logistics routing, real-time trading models and you start to see why $638 billion in contracted compute doesn’t feel like an exaggeration anymore.
If you’re trying to actually build on top of this wave instead of just reading about it, the practical path runs through learning how to set up and prompt these systems properly. Ou rguide to setting up Neo as an autonomous AI agent and our broader tutorial on building autonomous AI agents are the two most practical starting points we’ve published. And because none of this works well without good instructions, our advanced prompt engineering techniques for 2026 breaks down what’s actually changed in how you need to prompt agentic systems versus old-school single-turn chat. If you want examples specific to agent frameworks, the Agent Zero prompts guide is worth bookmarking.
What Could Go Wrong From Here
A few honest risks worth flagging, because most coverage of this earnings report skipped them:
Component cost inflation is real and Oracle said so directly. During the analyst Q&A, CEO Clay Magouyrk addressed rising memory and storage costs, explaining that Oracle uses fixed-price contracts where input costs are predictable, but switches to cost-pass-through mechanisms when supply chain uncertainty is too high to lock in a price. That’s a reasonable hedge but it also means some of Oracle’s margin protection depends on customers absorbing future cost spikes, which isn’t always a popular ask once contracts are already signed.
Margin compression is coming regardless. Oracle told investors that fiscal 2027 gross margin will step down because new data centers take time to reach full contractual revenue once they’re built there’s a lag between “we built it” and “it’s generating the revenue we contracted for.” That’s normal for capital-intensive buildouts, but it means near-term profitability metrics may look worse before they look better.
And then there’s the concentration risk we already covered: a meaningful share of that $638 billion sits with a small number of mega-customers, OpenAI chief among them. Oracle itself disclosed that it booked $67 billion in new AI infrastructure contracts in the most recent quarter alone which is genuinely impressive demand, but it also means the backlog can grow or shrink fast depending on a handful of decisions made by a handful of companies.
What to Actually Watch Next
If you’re tracking Oracle stock, skip the backlog headline next quarter and go straight to three numbers: free cash flow (is the gap between spending and collections narrowing or widening), the RPO maturity schedule (is more of it shifting into the “within 12 months” bucket, meaning it’s actually converting), and customer concentration disclosures (is OpenAI still the dominant driver, or is the contract base diversifying).
If you’re building on Oracle Cloud Infrastructure or evaluating it against AWS, Azure, or Google Cloud for an AI workload, the 93% OCI growth rate matters more to you than the backlog total that’s the number that tells you Oracle is actually shipping capacity, not just signing contracts for it.
And if you’re just trying to understand where the AI infrastructure race stands in mid-2026: Oracle didn’t win this quarter because it has the best technology story. It won because it was willing to take on more debt and more risk than Microsoft or Alphabet to chase the same demand. Whether that bet pays off depends entirely on whether the agentic AI buildout — the agents, the autonomous workflows, the compute-hungry verticals we walked through above — keeps accelerating at the pace Oracle’s balance sheet is now betting on.
