Oracle’s $664 Billion Backlog Still Has to Be Built

Oracle‘s remaining performance obligations reached $664 billion on September 10, up from $638 billion in June and up $209 billion from a year earlier. That is the number every Oracle headline leads with now.

RPO is a promise, not a receipt. It counts contracts signed and revenue not yet delivered, and in the quarter alone Oracle booked more than $30 billion of new AI cloud commitments on top of what was already on the books. The gap between what has been promised and what has actually been billed is where this post lives.

My view, stated plainly: the backlog itself is not the risk here. The capital required to build the data centers that turn it into delivered, billable capacity, before Oracle’s margins recover from the spending, is.

The gap between promised and billed

Start with the quarter Oracle just reported. Total revenue for the three months ended August 31 was $19.3 billion, up 30% from the same quarter a year earlier, a figure that matches what Oracle told investors on its call and filed with the SEC.

Inside that number, cloud revenue, infrastructure and software combined, was $11.6 billion, up 62%. Cloud infrastructure alone, the raw computing capacity AI labs rent by the hour, rose 121% to $7.4 billion. Cloud applications, the older software-as-a-service business, grew a comparatively pedestrian 10%. Oracle’s investor relations release called it the ninth straight quarter of accelerating infrastructure growth, and the arithmetic backs that framing.

GPU utilization across that infrastructure reportedly sat near 98% in the quarter, which is another way of saying Oracle is not sitting on idle capacity waiting for customers to show up. Demand, at least right now, is not the open question.

Here is the part that gets less attention: gross margin fell to 65.8% from 70.5% a year earlier.

Building a data center ahead of demand means paying for power contracts, chips and depreciation before a single hour of that capacity gets billed. It is the same capital intensity I flagged when I looked at the AI spending cycle nobody wants to slow down; Oracle is simply running that cycle at a scale its balance sheet has never carried before. Margin is where the mismatch shows up first, and it showed up.

Thirty billion new dollars, one big customer

Oracle does not name customers in its filings, but outside reporting has done the math anyway. Independent estimates put OpenAI’s share of the RPO at roughly $300 billion, more than half the backlog, tied to a five-year cloud-computing agreement often described under the Stargate name.

That is a striking level of concentration for one counterparty, on any measure. The deal reportedly involves roughly 4.5 gigawatts of compute a year and is structured to start delivering in 2027, which implies something on the order of $60 billion a year in cloud spending from OpenAI alone once it ramps, a single customer contributing revenue on the scale of a Fortune 100 company’s entire cloud budget.

Oracle’s answer, per the same reporting, has been to grow the rest of the book faster rather than argue the concentration away: non-OpenAI backlog has reportedly more than doubled over the past year, which is presumably part of why management felt comfortable booking another $30 billion of new contracts this quarter without much fanfare on the call. I read that as a company aware of its concentration problem and actively working against it, not one hoping nobody notices.

It does not remove the exposure, though. The plainest counter-case here is straightforward: if OpenAI slows its committed spending, or its funding does not keep pace with what it has agreed to pay, Oracle is left holding infrastructure built to one customer’s specifications that cannot be instantly re-leased to somebody else. I would treat that scenario as the single biggest threat to the bull case, bigger than anything about Oracle’s own execution.

A decade ago Oracle’s cloud business was a distant follower behind Amazon, Microsoft and Google, a reputation that still colors how skeptically some investors read anything the company says about this category. The skepticism was earned once. Whether it is still earned is the actual argument this backlog number is trying to settle, and I don’t think one quarter answers it either way. It is worth remembering Micron’s experience with a memory cycle everyone assumed would repeat: concentrated, capital-heavy bets on one demand source can look brilliant for years and then reprice hard in a single quarter when that demand source stumbles.

The margin this buildout is costing

Operating margin held up better than gross margin, coming in around 33% for the trailing year, with EBIT margin near 35.9%. That tells me the pressure is concentrated in cost of revenue rather than in overspending on sales or research, which is the healthier place for margin compression to show up if it has to show up at all.

Net income for the trailing year was $17.1 billion, up from $12.4 billion. Profitability is still rising. It is just not rising as fast as revenue and the backlog headline would suggest on their own, and that gap between the two growth rates is the whole story in miniature.

A financing wrinkle in the footnotes

Buried in the same release was a line that deserved more attention than it got: Oracle said the new AI contracts came with no incremental impact on its plans to raise capital, according to ERP Today’s coverage of the filing. That is corporate language for “we are not borrowing more because of this specific batch of deals,” and it matters because much of 2026 has been spent worrying about how Oracle actually pays for all this construction.

Oracle also guided fiscal 2027 total revenue to at least $90 billion, with non-GAAP earnings per share of $8.10, both comfortably ahead of the $67.4 billion in revenue and $6.38 in trailing per-share profit it has on the books today.

I would not call any of that a settled question. One quarter’s disclosure is not a multi-year financing plan, and building $90 billion or more of annual revenue capacity is not cheap under any structure Oracle chooses. Whether that construction gets funded through operating cash flow, new debt, or the kind of off-balance-sheet vehicles other hyperscalers have leaned on, the bill still has to be paid before the backlog turns into recognized revenue, and paying it is what is compressing margin today.

What the valuation already assumes

Oracle trades around $147.61 as I write this, 54.7% below its 52-week high of $326 and 29% above its low of $114. That is a wide range for one year of trading, and it tells me the market has not settled on a story here either.

On trailing earnings the stock sits at 23.1 times, well under its own five-year average of 34.7. On a forward basis, using next year’s expected earnings, the multiple drops to 20.6. Price to sales tells a similar story: 6.3 times now against a five-year average of 7.0.

Run that forward multiple through a simple reverse discounted cash flow and the market looks like it is pricing something close to Oracle’s own guided growth, not the accelerating growth that a $664 billion backlog implies once it converts. Analysts, on average, still see more room than the stock price does: the average target sits at $254, about 72% above the current price, with 28 analysts covering the name and 86% rated buy or the equivalent.

That is an unusually wide gap between analyst targets and where the stock actually trades for a company this closely followed. Either the Street is early, or the market is right to stay cautious about how much of that backlog converts on schedule. Short interest, for what it’s worth, sits at only 1.5% of the float, so bears are not pressing this trade with much conviction either. The dividend, small at 1.35% yield and $2.00 annually, is not why anyone owns this stock right now.

Metric	Now	Comparison
Remaining performance obligations	$664B (Sep 10)	$638B in June; +$209B y/y
Cloud revenue	$11.6B, +62%	OCI +121% to $7.4B; SaaS +10%
Gross margin	65.8%	70.5% a year ago
Trailing P/E	23.1	5-yr avg 34.7
Forward P/E	20.6	Price/Sales 6.3 vs avg 7.0
Analyst target	$254 (72% upside)	28 analysts, 86% buy-rated

The ratio I would watch next

Microsoft went through a milder version of this same margin squeeze years ago while it built out Azure capacity ahead of demand, a comparison I made when I looked at Microsoft’s own multiple against its history; margin there eventually recovered once utilization caught up with the spending. Oracle’s bet is that the same pattern repeats, at a larger scale and with a far more concentrated customer base than Microsoft ever carried.

My own uncertainty sits squarely there. I do not know how fast Oracle’s non-OpenAI backlog needs to grow before customer concentration stops being the first question analysts ask on every call. Diversifying a $664 billion backlog away from one roughly $300 billion contract takes years, not quarters, and nothing in this release changes that timeline.

What I would watch next is simple. Track the backlog-to-revenue ratio each quarter, and watch whether gross margin stabilizes anywhere near 66% instead of continuing to slide toward 60%. If it holds there and non-OpenAI bookings keep compounding the way they have, the case for owning this gets easier to make. If gross margin keeps falling much past 65% while concentration stays this high, the growth is arguably costing more than it is worth for now.

Analysis and opinion only, not investment advice. Figures come from Oracle’s fiscal first-quarter 2027 results filed on SEC EDGAR and released on its investor site; backlog-concentration figures were checked against reporting from Yahoo Finance on September 22, 2026.