Oracle's GPUs Are Sold Out — and Its Customers Are Funding the Next Batch

Oracle's GPUs Are Sold Out — and Its Customers Are Funding the Next Batch

I run servers for a living, so when Oracle says its GPU fleet is 97.9% utilized, that is not a stock-market story to me. That is a capacity story. It is the closest thing to "sold out" that exists in this business.

Oracle reported its fiscal first quarter on Thursday. Record revenue of $19.3 billion, up 30%. Cloud infrastructure revenue — the part that rents GPU capacity to AI developers — hit $7.4 billion, up 121%, the ninth consecutive quarter of accelerating growth in that segment. The stock had fallen 22% this year and sat about 55% below its September 2025 peak. It popped 4% after hours and kept climbing Friday.

That is the version everybody ran. Here is the version that matters if you own hardware, because the second story is sitting in the cash flow statement.

The Numbers That Actually Mattered

Scale first. Total cloud revenue was $11.6 billion, up 62%. Applications grew 10%. The legacy software business — the one Oracle lived on for thirty years — shrank 3% and missed. Non-GAAP operating income was $8.2 billion, up 31%, with operating margin flat at 42%. Earnings per share came in at $1.92 against $1.74 expected.

Then the backlog. Remaining performance obligations reached $664 billion — up $209 billion year over year and $26 billion in a single quarter. Oracle booked more than $30 billion of new AI cloud contracts, delivered 850 megawatts of new data centre capacity, and shipped more than 300,000 GPUs: nearly triple the previous quarter, and 73% of everything it delivered in the prior fiscal year. For the first time in its history, Q1 revenue grew sequentially instead of dropping.

Now the number that quietly gutted the bear case on this whole sector.

The Number Nobody Expected: Used GPUs Sold at a Premium

"Our gross margin did decline as expected," the CFO said, and everybody nodded, because that is what happens when you pour $28.5 billion of capital expenditure into one quarter and the revenue arrives later. Fine. That is not the line.

The line came from co-CEO Clay Magouyrk. When Oracle has capacity come up for renewal, he said, "we actually can achieve higher prices to the order of 20%." Read that again. GPUs that came up for renewal were renewed or resold at a premium to the prior contract. Not a discount. Not a write-down. A premium.

I have been told for two years that these chips are depreciating assets on a two-to-three year clock, that the hardware is worthless before it is paid off. Oracle just published a data point that says otherwise — at least for capacity a customer is already sitting in, with the network, the power and the operations already wired around it. Scarcity of qualified capacity is doing the pricing, not the age of the silicon.

If you are still depreciating GPU nodes over twenty-four months, you are pricing yourself out of your own margin.

Now the Other Reading — Follow the Cash, Not the Revenue

Same quarter. Different story. Oracle generated record operating cash flow of $23 billion, up 184%. But $11.4 billion of it — roughly half — came from customer prepayments carrying what the filing calls a "significant financing component." Customers paid in advance for GPU time they have not used yet, and that money funded construction of the capacity those same customers will eventually rent back.

Capital expenditure was $28.5 billion, against $8.5 billion a year earlier. Free cash flow was negative $5.4 billion, versus negative $362 million a year ago. Net cash capex after those prepayments was $18 billion. Total debt sits at $125 billion, and interest expense rose 55% to $1.4 billion. On top of all of it, Oracle sold $20 billion of new common stock through an at-the-market programme during the quarter, with the diluted share count up about 3% to three billion shares.

So the beat was real. So was the dilution. So was the $5.4 billion outflow. All three showed up at the same party.

The Bottleneck Nobody Wants to Say Out Loud: This Buildout Is Paid For By Its Own Participants

Here is the piece that changes how you underwrite everybody in this chain. Oracle's co-CEO said the $30 billion of new AI contracts were signed "without requiring additional capital from Oracle." That sounds like strength. Look at the mechanics. The funding structures behind those deals are customer prepayments, bring-your-own-hardware arrangements and supplier financing — and Oracle says customers are prepaying or supplying hardware for $75 billion of its largest AI contracts.

Then look at who is lending to whom. On September 2, HPE announced that Oracle would deploy its Juniper routing and switching gear across Oracle's AI data centres globally. As part of the deal, HPE issued Oracle a warrant to buy up to 4,156,466 HPE shares at one cent each — a warrant dated July 2, according to HPE's quarterly filing. HPE did not cut the price. It handed over equity. HPE stock then rose 9% to 11% on Friday as the market re-rated its quarter.

Follow the circle. The customer prepays the cloud. The cloud spends the prepayment on network gear. The network vendor pays the cloud in its own stock, at a penny a share. Everybody books a win. Nobody's cash leaves the family.

I am not calling it fraud, because it is not fraud. This is how scale gets financed when demand genuinely runs ahead of supply and the debt market charges you a premium for it. But see it for what it is: the last layer of this buildout is funded by the participants themselves, it never appears as a line item on an income statement, and it is the layer that snaps first if one participant flinches.

What This Actually Means for Independent Hosting Providers

First, re-underwrite your depreciation schedule. A 20% renewal premium on aging GPU capacity is a direct challenge to the two-year write-off. If your model calls the asset scrap in 24 months and Oracle is re-leasing the same class of asset at a higher price than the first time, your model is costing you money. Extend the life, reprice the capacity, and stop giving margin away.

Second, understand that the biggest clouds are effectively full. 97.9% utilization is not a marketing number — it is the point where a provider starts queuing customers. When supply is that tight, prices go up. Magouyrk said it out loud: "in a world where demand exceeds supply, typically prices don't go down, they do go up… and we have to charge more money for them so that we get compensated." That is your window. Do not leave rates where they were last year.

Third, copy the prepayment structure if you have an anchor tenant. Oracle pulled $11.4 billion of working capital out of its own customers in one quarter. If you have a customer who needs guaranteed capacity for three years, a prepayment-for-capacity contract is the cheapest capital you will ever raise, and it does not dilute you or put a lien on your hardware.

Fourth, watch payer quality, not backlog size. A $664 billion backlog looks like a fortress until you work out that roughly half of it traces back to one counterparty — OpenAI — which is not yet profitable and has contracted to buy about $300 billion of Oracle capacity over five years from 2027, averaging $60 billion a year against reported annualized revenue of around $40 billion. S&P cut Oracle to BBB- in July, one notch above junk, for exactly that reason. Underwrite the customer, not the headline.

Fifth, price against the dilution tax. Anybody building AI capacity with equity pays for it twice — once in capex, once in share count. Twenty billion dollars of new stock in one quarter is a real cost of capacity. If you are not diluting, that is a structural advantage. Use it.

The Counter-Argument — "Demand Is Real, So Who Cares Who Paid First?"

Fair enough, and the bulls have a case. Oracle delivered 850 megawatts in a quarter, has sites running ahead of schedule, says New Mexico and Wisconsin will not touch guidance, and is taking delivery of Vera Rubin systems performing better than expected. Abilene alone delivered 131,000 GPUs and trained GPT-6 Astra. Full-year guidance went up to at least $90 billion. None of that is a house of cards.

But the capex has to be serviced no matter who fronted the cash. Interest expense up 55% does not care that the prepayment came from a customer. And if the strategy depends on one customer's IPO landing, then it depends on the IPO, not on the demand. Magouyrk gave us the honest version himself: "if their plan relies on 100% achievement of every one of their deliverables, we have a term for that. It's called a bad plan." He is right. He also just told you every deliverable in this plan is load-bearing.

The Bottom Line

Oracle proved the demand is real, and it proved that qualified AI capacity commands a premium even after it has been used. Anybody who has spent 2026 telling you the buildout was already dead should go read that renewal number twice.

It also showed you the bill: $125 billion of debt, negative $5.4 billion of free cash flow, twenty billion dollars of new shares pushed into the market in ninety days, and a cash flow statement where half the operating inflow is the customer paying in advance to build the thing they are going to rent back.

I have run infrastructure long enough to know the difference between a backlog and a bank balance. Backlog is a promise. Cash flow is a fact. Oracle has more promises than anybody in this industry right now — and it just got the market to fund half of them.

— Allan Ali, Founder

This article was produced with AI-assisted research and editorial support. Sources: Oracle Q1 FY2027 results (Oracle Investor Relations / PRNewswire, September 10, 2026), Oracle Q1 FY2027 earnings call transcript (The Motley Fool, September 11, 2026), Reuters, Bloomberg Television, Forbes, Diginomica, SiliconANGLE, TechTimes, Futurum Group, Investing.com, HPE quarterly filing, S&P Global Ratings, Council on Foreign Relations commentary.