Forecast gross margin 30%-40%! Oracle reassures investors, don't believe that "AI only makes noise and doesn't make money"?

In response to external doubts about the profitability of its AI data center business, Oracle made a clear commitment to investors: this business, primarily consisting of NVIDIA chip rentals, will ultimately achieve a gross margin of 30% to 40%. This statement aims to quell market concerns about its declining profit margins and to justify its aggressive cloud computing expansion strategy.

On Thursday, Oracle Co-CEO Clay Magouyrk stated at the annual investor conference in Las Vegas that the aforementioned gross margin forecast covers the entire customer contract cycle, including the initial investment phase. Previously, technology media The Information reported that Oracle's gross margin for GPU rental business over the past five quarters was only about 16%, far below the company's overall gross margin level of approximately 70%, raising market concerns about its AI business profitability.

Oracle also raised its revenue forecast for cloud business in fiscal year 2030 to $166 billion, 15% higher than the forecast made a month ago. This implies explosive growth from approximately $10 billion in fiscal year 2025. The company expects total revenue in fiscal year 2030 to reach $225 billion, exceeding the average analyst expectation of $198 billion compiled by Bloomberg, with adjusted earnings per share expected to be $21.

NVIDIA CEO Jensen Huang endorsed Oracle in an interview with CNBC, stating that it is normal for new technologies not to be profitable in the early stages: "They will achieve excellent profitability over the entire lifecycle of the system." Oracle's stock price closed up 3.1% at $313 that day, but fell about 2% after releasing long-term financial forecasts in after-hours trading.

Early Losses Weigh on Current Performance

Oracle's profit pressure is evident in the latest leaked internal documents. According to documents obtained by The Information, the company generated approximately $900 million in revenue from NVIDIA chip rentals in the quarter ending in August, with a gross profit of only $125 million, resulting in a gross margin of 14%.

The performance of the latest generation Blackwell chips was particularly dismal. In the quarter ending in August, Oracle incurred nearly $100 million in losses on chip rentals, pulling down the overall GPU gross margin. When accounting for all depreciation costs, Oracle's overall GPU gross margin fell to single digits due to losses from Blackwell chips.

This confirms the "climbing period" phenomenon mentioned by Magouyrk. He presented a hypothetical case in his demonstration: A six-year, $60 billion AI infrastructure contract, where Oracle might pay $570 million before the contract starts, resulting in a negative gross margin. However, once the data center is operational, the company could generate $10 billion in revenue annually, with costs of $6.4 billion, bringing the overall gross margin to 35%.

Even with the Hopper chips, which have been on the market for two years, Oracle struggles to meet its long-term goals. Internal documents show that Hopper chips contributed about half of the GPU rental revenue in the quarter ending in August, but the gross margin was only 26%. The H200 chip, launched a year ago, had a gross margin of 19%, accounting for nearly 20% of the quarterly revenue. For even earlier NVIDIA chips launched five years ago, Oracle even incurred losses in rentals

NVIDIA's Pricing Power Becomes a Core Variable

Whether Oracle can fulfill its gross margin commitment largely depends on its bargaining power with NVIDIA. Internal documents show that even with NVIDIA chips launched one to two years ago, Oracle has found it difficult to achieve a gross margin exceeding 25%.

NVIDIA's dominance gives it strong pricing power. According to Bloomberg Intelligence analyst Anurag Rana, NVIDIA's operating profit margin has increased by 40 percentage points in recent years. Cloud service providers must bear the drag on gross margins from expensive chips, and NVIDIA has yet to determine the pricing for its next-generation chips, while other factors such as energy and additional data center equipment may also impact Oracle's costs.

Oracle is at a disadvantage compared to competitors like Microsoft and Amazon Web Services. These companies are less reliant on NVIDIA chips, as most computing workloads still use central processing units (CPUs) rather than GPUs, and high-margin CPU servers can smooth out the drag from NVIDIA's expensive chips. Amazon has also reduced costs by using self-developed AI chips, but Oracle does not develop its own AI chips.

According to The Information, some executives at Microsoft have expressed skepticism about Oracle's ability to achieve such growth. If these goals are met, Oracle's cloud server leasing revenue could theoretically match Microsoft's in the early part of the next decade. Microsoft is currently second only to Amazon Web Services.

Diversified Contracts Provide Buffer Space

Oracle emphasizes that its cloud business does not solely rely on GPU leasing. Magouyrk pointed out that during a 30-day period this quarter, Oracle signed contracts worth $65 billion with four cloud customers, none of which were OpenAI, one of which was Meta.

The diversification within the cloud business provides Oracle with some leeway. This division includes four different business lines, some of which have higher gross margins than the AI server leasing provided to OpenAI and Meta, such as cloud contracts with companies and database clients like Uber, Zoom, and Palantir.

Oracle's high-margin software business still accounts for a large portion of its revenue and also helps enhance overall profitability. However, the company's long-term forecast suggests that by 2030, most revenue will come from GPU leasing, which will drag down overall profit margins. In this regard, Oracle executives encourage investors to focus on revenue growth rather than margin percentages, as growth will help the company generate more overall profit.

Oracle is also trying to reduce its dependence on NVIDIA. In recent months, the company has increased its leasing of AMD chips and announced this week a deepened commitment to purchasing AMD GPUs. OpenAI has indicated that it may ultimately use 6 gigawatts of AMD chips through cloud service providers like Oracle. However, in the quarter ending in August, Oracle incurred losses on AMD GPU leasing, as these chips will become widely available starting in the fall of 2024. AMD chips accounted for only 3% of total GPU leasing revenue for the quarter.

The cloud computing GPU leasing market remains fraught with uncertainty. Customers typically agree to pay a price for GPU leasing per hour over a period of time, meaning that if provider costs change, their profit margins may fluctuate. NVIDIA's near-annual release of new generations of GPUs also puts pressure on cloud service providers, leaving them with little time to complete the startup and operation of chips before needing to order and deploy the next generation Oracle rented out approximately 270,000 various types of GPUs in the quarter ending in August. In comparison, GPU cloud service provider CoreWeave claims to have over 250,000 GPUs, but its revenue in the most recent quarter was $200 million more than Oracle's cloud business. Professional GPU cloud service providers like CoreWeave do not have other major revenue sources to enhance profit margins, which may pose a greater challenge to their profitability