TheChek News Flow

Oracle on Thursday sought to reassure investors that its business of renting out servers packed with Nvidia chips to OpenAI, Meta and other artificial intelligence developers will become far more profitable as revenue surges.

Oracle co-CEO Clay Magouyrk told financial analysts Thursday that the company’s AI data center business, which is mostly made up of rentals of Nvidia graphics processing units, will eventually generate a gross profit margin of 30% to 40%. The Information last week reported that the firm’s gross margin from GPU rentals over the past five quarters was around 16%.

Without addressing The Information’s reporting, Magouyrk noted that Oracle’s AI data center business goes through a “ramp-up” period when Oracle incurs expenses before revenue starts coming in.

But Oracle has found it challenging to generate a gross margin of more than 25% from renting out Nvidia chips that came out one or two years ago, according to a new internal Oracle document that hasn’t been previously reported.

The document suggests some of the difficulties Oracle may face as it tries to narrow the gap between its projection and its margins today, even excluding the ramp-up period Magouyrk mentioned. Oracle could be betting it can negotiate better pricing with Nvidia over time, or could charge customers a higher rate.

Magouyrk also raised Oracle’s long-range revenue projections for the cloud business to $166 billion in revenue for the 12 months ending May 2030, which was 15% higher than the projection Oracle gave for that period just a month ago. That’s up from around $10 billion in revenue in the 12 months that ended in May this year. Oracle stock rose 3% from Wednesday’s close.

Magouyrk emphasized that while OpenAI is a major Oracle customer, it isn’t the only one. He pointed out that in one 30-day period during the quarter, Oracle signed $65 billion worth of contracts from four cloud customers, none of which was OpenAI. One was Meta, he said.

Oracle’s presentation was aimed at defusing investors’ concerns about the impact on its profits of its expansion into the cloud. Oracle’s highly profitable software business accounts for most of its revenue, but the company’s long-term projections imply that most of its revenue by 2030 will come from GPU rentals, which will drag down its overall profit margins.

Oracle executives on Thursday encouraged investors to focus on its projected revenue growth rather than its margin percentages, because the growth will help the firm generate more profits overall, even if margins are lower.

One thing working in Oracle’s favor is that it has a wide array of other, higher-margin business units that help bring up its overall profitability. Even within its cloud unit, it has four different business lines, some of which have higher margins than AI server rentals to OpenAI and Meta, such as its cloud contracts with enterprise and database customers including Uber, Zoom and Palantir.

The company didn’t break down revenue from each segment. Magouyrk said the revenue projections include all of the cloud business. OpenAI’s GPU rentals contracts could dwarf contracts from other customers, however, as it plans to spend hundreds of billions of dollars with Oracle through the early part of next decade.

Magouyrk said the projected 30% to 40% gross margin range for AI server rentals covers the entirety of a customer contract, including the period when “there’s a cost where we’re paying for something, but we’re not making revenue yet,” he said.

Magouyrk presented a slide that depicted a hypothetical AI infrastructure deal with $60 billion in revenue over six years. The example showed Oracle might pay $570 million before the contract starts, which would result in a negative margin. But according to the slide, these costs are minimal compared to the $10 billion in revenue, on $6.4 billion of expenses, that it could generate per year once the data center is ready for a customer.

Much remains uncertain, however. Cloud customers typically agree to the price they are willing to pay for GPU rentals per hour over a period of time, meaning if a cloud provider’s costs change, its margins could fluctuate. However, Nvidia hasn’t determined prices for its next-generation chips, and other factors such as energy and additional data center equipment could impact Oracle’s costs.

Executives at some of Oracle’s rivals—namely Microsoft—have expressed some skepticism about Oracle’s ability to pull off that kind of growth, The Information reported Thursday. Hitting those numbers could theoretically put Oracle on par with Microsoft in cloud server rental revenue by early next decade. (Microsoft is No. 2 behind Amazon Web Services.)

Oracle’s cloud rivals are less dependent on Nvidia chips because most computing workloads still involve servers powered by central processing units rather than GPUs for AI. Higher-margin CPU servers can smooth over the drag that Nvidia’s pricey chips put on gross margins. But the major cloud providers are also trying to figure out how to minimize that impact, given Nvidia’s dominant position, which gives the chip designer pricing power. (Nvidia’s operating margin has risen 40 percentage points in recent years.)

Amazon, for instance, is benefiting from using its own AI chips to power some of its cloud services rather than Nvidia chips. Oracle doesn’t develop its own AI chips, however.

As Oracle reinvents itself, it suddenly finds itself on the bleeding edge of AI data centers. That means it is one of the biggest buyers of Nvidia’s expensive GPUs and is trying to string the chips together quickly in facilities so customers can utilize them and they don’t have a negative effect on profit margins.

In the August quarter, Oracle generated around $900 million from rentals of Nvidia chips and recorded a gross profit of $125 million—meaning it kept 14 cents for every $1 of sales. That number blended several types of Nvidia GPUs, ranging from old generations that came out years ago to Nvidia’s newest AI chips, which just began rolling out and are more expensive or aren’t highly utilized.

Oracle lost nearly $100 million from rentals of the newest chip, Blackwell, which dragged down its overall GPU gross margin. (Including all depreciation costs, Oracle’s overall GPU gross profit margin was in the single digits due to the Blackwell chip losses, the document suggested.)

Nvidia CEO Jensen Huang said in response to The Information’s story last week that such losses can be expected in the initial rollout of new hardware.

“When you first ramp up a new technology, there’s every possibility that you might not make money in the beginning,” Huang said in an interview with Jim Cramer.

“But over the life of the system, they’ll be wonderfully profitable.”

That claim remains to be seen, however. About half of Oracle’s GPU rental revenue in the August quarter came from Hopper chips, which became generally available on Oracle Cloud two years ago, and the company generated a gross margin of 26% from those chips in the August quarter.

Oracle also generated a 19% gross margin from Hopper chips that came out a year ago, known as H200s, and represented nearly 20% of its revenue in the August quarter. Oracle previously had trouble finding customers to use them, but it signed a new deal with OpenAI to boost their usage, the document shows.

A spokesperson for Oracle didn’t have a comment on the figures.

The internal document also could help answer a looming question in the industry: how long Nvidia chips will hold their value. That’s become an especially worrisome issue for cloud providers like Oracle and Amazon as Nvidia releases a new GPU almost every year, leaving the companies relatively little time to get chips up and running before they need to order and set up the next generation of chips.

According to internal data, Oracle loses money on some older-generation Nvidia chips that became generally available on Oracle Cloud five years ago. The exact reason isn’t clear, but it could be because the company can no longer charge much money for them. It’s also possible Oracle is heavily discounting those GPUs to win customer contracts.

Oracle also plans to rent out more non-Nvidia chips to AI developers, which could help lower its costs or give it negotiating power with Nvidia. In recent months, Oracle has rented out more AMD chips, and earlier this week it announced a deeper commitment to buying AMD GPUs. OpenAI has separately said it could eventually use 6 gigawatts of AMD chips to power its AI, and that usage could happen through cloud providers like Oracle.

For now, margins on the AMD chips don’t seem great. In the August quarter, Oracle lost money on AMD GPUs, which became generally available in fall 2024, after properly accounting for depreciation, the document shows. The reason for the loss couldn’t be learned.

AMD chips represented only 3% of Oracle’s total GPU rental revenue in the quarter, however.

Other GPU cloud providers, such as CoreWeave, don’t have other major revenue streams to help boost their margins. At the end of the August quarter, Oracle rented out around 270,000 GPUs of various types. CoreWeave says it has more than 250,000 GPUs, though it generated $200 million more in revenue than Oracle Cloud in its most recent quarter.