Internal Oracle Data Show Financial Challenge of Renting Out Nvidia Chips
The Takeaway
- Previously undisclosed data on Oracle AI cloud business points to industrywide challenge
- Oracle is banking on bigger revenue to offset profit margin decreases
- Oracle lost nearly $100 million from rentals of Blackwell chips in most recent quarter
Oracle became the best-performing megacap stock of 2025 after its executives said last month that the once-sleepy database firm will generate an astonishing $381 billion in revenue from renting out specialized cloud servers to OpenAI and other artificial intelligence developers over the next five fiscal years.
But internal documents show the fast-growing cloud business has had razor-thin gross profit margins in the past year or so, lower than what many equity analysts have estimated. That could raise questions about whether the AI cloud expansions undertaken by Oracle and its rivals will affect profitability and sustain investors’ expectations.
In the three months that ended in August, Oracle generated around $900 million from rentals of servers powered by Nvidia chips and recorded a gross profit of $125 million—equal to 14 cents for every $1 of sales, the documents show. That’s lower than the gross margins of many nontech retail businesses.
As sales from the business nearly tripled in the past year, the gross profit margin from those sales ranged between less than 10% and slightly over 20%, averaging around 16%, the documents show.
In some cases, Oracle is losing considerable sums on rentals of small quantities of both newer and older versions of Nvidia’s chips, the data show. A spokesperson for Oracle, which doesn’t publicly disclose the AI cloud server unit’s financials, did not have a comment on those figures.
The 14% gross margin figure appears to take into account the labor, power and other direct costs of running Oracle data centers, including depreciation expenses for some of the equipment. Other unspecified depreciation expenses would eat up another 7 percentage points of margin, the documents show. Equity analysts such as Gil Luria from D.A. Davidson say they tend to account for depreciation of chips and other data center equipment when calculating a cloud provider’s gross margin, as the equipment is a key cost of selling cloud services.
The internal documents offer a rare glimpse into Oracle’s dramatic transformation from a weak link of the cloud industry to an AI data center powerhouse. It also shows that being in the business of buying and renting out Nvidia graphics processing units isn’t easy.
If Oracle’s revenue from the server rental business eventually eclipses its traditional software businesses, such as databases and enterprise resource planning tools, the weak margins on server rentals mean its overall gross profit margins will drop considerably from the level they’ve been at in recent years—roughly 70%. Those margins have already sunk from around 80% a decade ago as the company has increasingly turned to selling cloud services.
The previously undisclosed Oracle figures aren’t comparable to what other AI cloud server rivals such as publicly traded CoreWeave and privately traded Lambda report to investors. Each company appears to calculate its margins differently.
All cloud firms face the same financial challenge: the high price of Nvidia’s chips. The cost of powering Nvidia server chips far exceeds that of using traditional servers, which were the backbone of the cloud computing industry before the rise of ChatGPT. AI computing also requires other specialized hardware, such as networking equipment to connect the servers.
And to win big deals with AI customers, Oracle and other cloud providers have offered hefty discounts on GPU rental prices, compared to prices they list for regular customers, according to several people with knowledge of these deals. That further diminishes the profitability of those contracts.
Several big cloud providers have said their spending on such chips has hurt their margins in recent quarters.
Building AI-focused data centers has “put pressure on Azure gross margins,” Microsoft Chief Financial Officer Amy Hood said at a Morgan Stanley event in March, in reference to the company’s cloud server rental business. But she said that pressure would lessen as Microsoft made improvements to run the servers more efficiently.
“I feel better about margins [this year] than I felt when we started in June” last year, Hood said.
Pricing ‘Aggressively’
Bigger cloud providers have been able to offset the margin hit from renting out Nvidia GPUs to some extent because most of their businesses don’t rely on that expensive hardware. Amazon Web Services reported a 33% net operating profit margin and Google Cloud reported a 17% net operating margin in the quarter that ended in June, for instance. (Neither company discloses a gross profit margin or breaks out financials for its GPU server rentals.)
Oracle is taking a more direct hit to its margins by renting out GPUs. Oracle reported around $10 billion in sales from renting out cloud servers in the fiscal year that ended in May, with around 20% of that revenue coming from GPU servers, according to the internal documents. In the most recent quarter, the percentage of GPU cloud server sales rose to 27%.
Oracle’s public disclosures imply its GPU cloud business could equal the company’s noncloud revenue as soon as 2028.
Aside from whether Oracle can build enough data centers to generate the revenue it has promised investors, another challenge is how reliant its GPU cloud business could become on a single customer.
“Oracle said, ‘OK, would you rather us not sign these deals and keep the margins higher, or would you rather us sign these deals and [see] the margins go down, but we have incremental profit?’” said Guggenheim Partners’ John DiFucci.
Oracle could still generate healthy free cash flow in the long run, even on the relatively low-margin business of renting out GPUs, if it can sustain rapid revenue growth. One way to do that could be through raising prices, something Luria reckons Oracle may attempt.
“[T]hey are pricing their offering very aggressively” right now, Luria said.
‘Not Our Specialty’
Cloud industry financial analysts largely believe AI computing is the lowest-margin business in the field, given the high cost of Nvidia GPUs. Still, Oracle’s margin on the business in recent quarters is lower than what analysts such as DiFucci and KeyBanc Capital Markets’ Jackson Ader had previously estimated.
Oracle’s margins suffer whenever it installs Nvidia’s latest GPUs in its data centers. In recent months, for instance, the GPU cloud gross margin dipped to the low teens from more than 20% due to the rollout of new Nvidia chips in facilities in Abilene, Texas, that Oracle rents out to OpenAI, the internal documents show.
Oracle could be taking an outsize hit to its margins because it doesn’t own the data centers its customers use, unlike AWS and Google Cloud, which own the majority of theirs. Instead, Oracle primarily leases its data centers from third parties, such as Crusoe.
Owning properties is “not really our specialty,” Oracle Executive Vice Chair Safra Catz said on the company’s earnings conference call last month.
Aside from whether Oracle can build enough data centers to generate the revenue it has promised investors, another challenge is how reliant its GPU cloud business could become on a single customer. Virtually all of the $317 billion worth of cloud deals it signed in the three months that ended in August came from OpenAI.
Blackwell Bleeding
Already, Oracle’s top five AI cloud customers make up roughly 80% of that business: ByteDance, Meta, xAI, OpenAI and Nvidia itself, which uses cloud-based GPUs for its own research and development. Other GPU cloud providers, including CoreWeave, Nebius Group and Lambda, face a similar risk of concentrating most of their cloud revenue in a small number of customers.
One silver lining in Oracle’s GPU business is the amount of revenue it is generating from older generations of Nvidia chips, such as the Ampere chips that came out in 2020. Those chips appear to be helping Oracle’s margins, while newer versions of Nvidia chips strain them.
That dynamic seems to contradict Nvidia CEO Jensen Huang’s recent comment that the release of newer chips essentially wipes out demand for older ones.
Oracle’s margins are also affected by how many of its servers customers are actually using—and paying for. Utilization of Oracle’s GPU cloud servers ranges between 60% and 90%, depending on the type of Nvidia chip that powers them, according to internal documents.
In the three months that ended in August, Oracle lost nearly $100 million from rentals of Nvidia’s Blackwell chips, which arrived this year. That’s partly because there is a period between when Oracle gets its data centers ready for customers and when customers start using and paying for them, the documents show. It’s not clear what causes the gap or how Oracle plans to shorten it.