Nvidia: Blackwell delivery volume will exceed previous expectations (FY25Q3 conference call)

NVIDIA (NVDA.O) released its Q3 FY2025 financial report (for the period ending October 2024) after U.S. stock market hours on November 21, Beijing time. The details are as follows:

Below is the summary of NVIDIA's Q3 FY2025 earnings conference call. For the financial report interpretation, please refer to NVIDIA is still the backbone, just reaching peak firepower.**

1. $NVIDIA(NVDA.US) Core information review of the financial report:

2. Detailed content of NVIDIA's financial report conference call

2.1 Key information presented by executives:

1. Business progress

Data Center Business

Revenue situation:

- Revenue was $30.8 billion, up 17% quarter-over-quarter and up 112% year-over-year.

- Demand for NVIDIA Hopper is strong, with NVIDIA H200 sales significantly increasing to several billion dollars, and overall ownership costs reduced by 50%.

Driving factors:

- Cloud service providers account for about half of data center sales, with revenue more than doubling year-over-year.

- NVIDIA H200 is widely deployed for AI training and inference, covering deployment needs for tens of thousands of GPUs.

- NVIDIA GPU regional cloud revenue doubled year-over-year, with increased NVIDIA cloud instances and sovereign cloud construction in North America, EMEA, and Asia-Pacific.

Technical progress:

- Rapid advancements in the company's software algorithms have improved Hopper inference performance by 5 times and reduced inference latency by 5 times. The upcoming NVIDIA NIM will further enhance Hopper inference performance by 2.4 times.

- The new Blackwell GPU has a performance improvement of 2.2 times, with significant cost efficiency improvements. 13,000 GPU samples have been delivered to customers.

Regional and application expansion:

- NVIDIA cloud instance revenue in North America, Europe, the Middle East, and Asia-Pacific doubled year-over-year; data center revenue in the Chinese market achieved quarter-over-quarter growth due to the shipment of export-compliant products.

- Customer companies are leveraging Hopper infrastructure to support next-generation AI models, generative AI content, and other applications, with related revenue doubling year-over-year Future Outlook

- With the comprehensive deployment of Blackwell, the data center business is expected to continue strong growth in the fourth quarter.

Networking Business

Revenue Situation:

- Revenue increased by 20% year-on-year, with some revenue from InfiniBand and Ethernet switches, SmartNICs, and BlueField DPUs showing quarter-on-quarter growth, strong and continuously growing network demand is expected to grow quarter-on-quarter in Q4.

Networking Platform Performance:

- CSPs and supercomputing centers are using and adopting the NVIDIA InfiniBand platform to power new H200 clusters.

- NVIDIA Spectrum - X Ethernet for AI revenue increased more than threefold year-on-year, with multiple CSPs and consumer internet companies planning large-scale cluster deployments, and pipeline construction continues.

- Utilizing Spectrum-X, the Hopper supercomputer achieved up to 95% data throughput.

Gaming Business

Financial Data:

- Gaming business revenue reached $3.3 billion, up 14% quarter-on-quarter and up 15% year-on-year.

Key Drivers:

- Strong demand for the RTX series, driven by back-to-school and holiday seasons.

- The newly launched GeForce RTX AI PC supports Microsoft Copilot+ features, enhancing gaming and creative application scenarios.

Professional Visualization

Financial Data: Revenue of $486 million, up 7% quarter-on-quarter, up 17% year-on-year.

Business Performance:

- RTX workstations are widely used in professional graphic design, engineering, and generative AI model prototyping.

- Generative AI further drives content creation demand, covering the media and entertainment sectors.

Autonomous Driving

Financial Data: Revenue reached a new high of $449 million, up 30% quarter-on-quarter, up 72% year-on-year.

Main Drivers:

- Increased demand for the NVIDIA Orin autonomous driving platform.

- Volvo launched a fully electric SUV based on NVIDIA DriveOS.

International Performance

China:

- Data center revenue increased quarter-on-quarter due to export-compliant products, but still below pre-export control levels.

- Market competition is fierce, and the company will continue to comply with export control regulations.

India:

- Collaborating with Tata Communications and Yotta Data Services to build an AI factory based on NVIDIA GPUs, with GPU deployment expected to increase tenfold by the end of the year.

Japan:

- SoftBank is building a supercomputer supported by NVIDIA DGX Blackwell and Quantum InfiniBand, and promoting the transformation of 5G networks based on the NVIDIA AI Aerial platform

2. Financial Performance

Revenue Side

Q3 revenue was $35.1 billion, an increase of 17% from the previous quarter and a year-on-year increase of 94%, far exceeding the expected $32.5 billion.

Gross Profit Side

GAAP gross margin was 74.6%, and non-GAAP gross margin was 75%, a quarter-on-quarter decline mainly affected by changes in the data center product mix.

With the full rollout of Blackwell products, the gross margin may temporarily decline to around the low 70s, but is expected to rebound to 75% in the medium to long term.

Expense Side

GAAP and non-GAAP operating expenses were $4.8 billion and $3.4 billion, respectively.

Q4 Guidance

Total revenue for Q4 is projected to be $37.5 billion, with a fluctuation of 2%.

GAAP and non-GAAP gross margins are expected to be 73% and 73.5%, respectively, with a fluctuation of 50 basis points.

GAAP and non-GAAP operating expenses are expected to be approximately $4.8 billion and $3.4 billion, respectively.

GAAP and non-GAAP other income and expenses are expected to be about $400 million in income (excluding gains and losses from non-affiliated investments).

GAAP and non-GAAP tax rates are expected to be 16.5%, with a fluctuation of 1%.

2.2 Q&A Analyst Q&A

Q: Has the expansion of large language models reached a bottleneck? How does your company help clients address these challenges? Additionally, given that many clusters have not yet adopted the Blackwell architecture, does this further drive demand for Blackwell?

A: The pre-training expansion of foundational models is still ongoing. Although this is a rule of thumb rather than a physical law, existing evidence suggests that the expansion trend remains. However, we find that relying solely on pre-training is insufficient to meet demand. Currently, there are three expansion methods being developed in parallel:

Post-training expansion: The initial post-training method was based on reinforcement learning from human feedback, which has now evolved into reinforcement learning based on AI feedback, combined with generated synthetic data. These technologies greatly assist in the expansion during the post-training phase.

Inference-time expansion: An important advancement is OpenAI's o1 model (such as Strawberry and ChatGPT o1), which introduces inference-time expansion (Test-Time Scaling). During inference, the longer the model thinks, the higher the quality of the output. This method involves techniques such as chain thinking and multi-path planning, similar to the process of deep thinking humans engage in before answering questions.

Pre-training expansion: This method remains core but is combined with post-training and inference-time expansion, significantly enhancing overall expansion capabilities.

With the combination of these expansion methods, the market demand for NVIDIA infrastructure has also significantly increased. The current generation of foundational models requires about 100,000 H100 GPUs during the training phase, while the next generation will require 100,000 Blackwell GPUs at the start, clearly reflecting the industry's progress in expansion capabilities. In addition, the demand for inference is also growing rapidly. NVIDIA is currently the largest inference platform in the world, benefiting from a vast installed base of devices. From Amperes to Hoppers to Blackwells, this infrastructure not only trains foundational models but also leaves a strong reserve of computing power for inference tasks.

The demand for generative AI applications among enterprises is also growing rapidly, driving the overall expansion of the AI market. Overall, the training of foundational models, the expansion of inference, the rise of AI localization companies, and the widespread application of AI by enterprises are collectively driving the sustained strong demand for NVIDIA's hardware and technology.

Q: Can NVIDIA execute the product roadmap proposed this year as planned, including the Ultra to be released next year and the Rubin transition plan in 2026? What is the company's response, and can it explain its ability to ensure timely execution?

A: The production of Blackwell is progressing comprehensively. This quarter's Blackwell delivery volume will exceed previous expectations. The supply chain team is working closely with suppliers to continuously expand Blackwell's capacity and plans to further increase supply next year.

Current demand significantly exceeds supply, which was expected in the early stages of the generative AI revolution, especially against the backdrop of the rapid development of the next generation of foundational models. These models can achieve inference and long-term thinking, particularly making breakthroughs in the field of physical AI, enabling AI to understand the structure of the physical world. The demand for Blackwell is very strong, and the company's execution is performing well.

We are conducting complex engineering integrations with multiple global partners, including cloud service providers (CSPs) such as Dell, CoreWeave, Oracle, Microsoft, and Google. These partners are accelerating the launch of their respective Blackwell systems; for example, Oracle has launched its system, and Microsoft is about to preview its Grace-Blackwell system. Although Blackwell is a full-stack and full-infrastructure product, we still need to break it down and integrate it into the architectures of customized data centers around the world.

This integration process is based on the experience accumulated from our multiple generations of products. Although complex, the company is very skilled at this. From the various systems currently deployed, the progress of Blackwell is good. Meanwhile, this quarter's delivery plan has exceeded our previous expectations.

In terms of executing the product roadmap, the company adheres to an annual iteration rhythm to ensure significant performance improvements of the platform. This performance enhancement not only reduces the costs of training and inference, making AI technology more accessible, but also directly increases customer returns. In data center environments with limited power supply (expanding from tens of megawatts to hundreds of megawatts and even gigawatts), the optimal performance-to-power ratio means customers can achieve the highest returns. The company is always committed to creating value for customers through technological innovation, and all current projects are progressing as planned, staying on track.

Q: Regarding supply chain constraints, is it a problem caused by multiple components, or is it mainly related to HBM (High Bandwidth Memory)? What is the current situation of supply chain constraints? A: Regarding the supply chain, the Blackwell system requires seven types of custom chips, supporting various configurations, including air cooling or liquid cooling, NVLink 8, 36, and 72 connections, as well as x86 or Grace architecture. The complexity of integrating these systems into global data centers is nothing short of miraculous. This quarter, the delivery volume of the Blackwell system surged from zero to billions of dollars, showcasing an astonishing growth rate. In this process, almost all major supply chain companies globally participated, including TSMC, Amphenol, Vertiv, SK Hynix, Micron, Foxconn, Quanta, Dell, reflecting a strong partner network.

Q: What is the growth trajectory of Blackwell this year? Is it still a correct expectation to surpass Hopper's shipment volume in the April quarter next year? Will the product transition in the April quarter peak the gross margin pressure?

A: Regarding gross margin, with the initial rollout of Blackwell, we will focus on providing the best experience for our customers. Due to the launch of various configurations and chips, the gross margin is initially expected to be in the low 70% range. However, as the rollout progresses, we expect the gross margin to quickly rebound to the mid-70% range in subsequent quarters.

The demand for Hopper will continue next year, especially in the first few quarters. Meanwhile, the shipment volume of Blackwell will increase quarter by quarter. Next quarter's shipment volume will exceed this quarter's, and subsequent quarters will see further growth. This indicates that we are at the beginning of two fundamental shifts in the computing field.

The first is the transition from traditional CPU-based programming to machine learning running on GPUs, a shift that has been widely adopted across industries and is the foundation of generative AI. Data centers and computing systems worth trillions of dollars globally are undergoing modernization for machine learning.

Secondly, a whole new capability, namely AI, will emerge on top of these systems. The advent of generative AI makes these data centers akin to "AI factories," not only processing data but also generating AI services around the clock, similar to how power plants provide electricity to a large number of consumers. These AI factories will operate long-term, providing continuous support to large-scale customers.

These two trends—modernization of data centers and the creation of a new AI industry—have just begun, and we expect this growth and transformation to continue in the coming years.

Q: Regarding the possibility of restoring gross margins to 70% - 75% in the second half of 2025, and how to face the digestion phase in the hardware deployment cycle?

A: It is possible to reach 70% - 75% gross margins in the second half of 2025, but it depends on the growth situation.

Regarding the digestion phase, it may not occur until the global data center modernization is completed. Data centers are still primarily based on manually written applications running on CPU architectures. This approach is no longer suitable for current demands. In the coming years, as the company expands data centers, it will be more inclined to adopt new architectures focused on machine learning and generative AI, while existing traditional data centers will gradually be replaced by modernized upgrades From the trend perspective, assuming that global IT spending on data centers grows at an annual rate of 20%-30%, by 2030, the global computing-related data center market could reach approximately $20 trillion. In this process, data centers will gradually transition from a traditional coding-based model to a machine learning model, which is the main task of the first phase.

The second phase is the development of generative AI, which represents a brand new capability and market opportunity. Similar to the emergence of the iPhone, generative AI does not replace existing technologies but opens up entirely new application scenarios. For example, OpenAI has not replaced any traditional business but has created a new category of intelligent services. With the development of generative AI, more "AI-native" companies will emerge, such as Runway, which focuses on digital art intelligence, Harvey, which specializes in legal intelligence, and others in digital marketing intelligence. These companies are akin to "cloud-first" or "mobile-first" enterprises of the internet era, seizing new opportunities brought about by platform transformation.

In summary, the coming years will continue to drive IT modernization transformation and the construction of AI factories. These AI factories are not only part of data center modernization but also serve as a new industrial pillar for generating artificial intelligence. This trend is expected to bring long-term growth and innovation opportunities.

Q: What is the definition of a low 70% gross margin, and what is the situation with Hopper and Blackwell in Q4?

A: Regarding the first question about gross margin, the low end typically refers to below the median. We expect the gross margin to fluctuate between 71% and 72.5%, and it may also be slightly higher. As production efficiency and product yield improve, we will gradually enhance this throughout the year, ultimately restoring the gross margin to around 75%, which is the median.

As for the second question, the demand for H200 has grown significantly, showing strong momentum in both order volume and deployment speed. As the fastest-growing product currently, its market performance is remarkable. We will continue to sell Hopper in the fourth quarter, covering various configurations, including those tailored for the Chinese market. At the same time, customers are actively deploying Blackwell, so both will proceed simultaneously in the fourth quarter. Whether Hopper's revenue will continue to grow based on the third quarter depends on market dynamics.

Q: Will the growth rate of the inference market exceed that of the training market in the next 12 months? Overall, what is your view on the development of the inference market?

A: Our vision is that one day, inference can be widely applied across all industries globally. This marks the true success of AI—when every company applies inference around the clock in its marketing, forecasting, supply chain, legal, engineering, and programming departments. We hope that thousands of AI-native startups will create value for users in every computer interaction, from Outlook to Excel, through generative AI models. Even when reading PDF files, there will be an inference generation process. A noteworthy case is Google's launch of the NotebookLM application, which provides users with new experiences by interacting with documents and PDFs The next step in AI development is Physical AI, a new field following language models. Physical AI understands the structure and logic of the physical world and can predict and simulate short-term changes in the future. This capability is crucial for industrial AI and robotics, and has spurred the rise of numerous AI-native companies and Physical AI enterprises. NVIDIA has developed the Omniverse platform, specifically designed to cultivate and train Physical AI, enhancing AI learning capabilities by generating synthetic data and physical feedback instead of manual feedback. This enables the efficient development of Physical AI and accelerates the rapid advancement of new reasoning technologies.

Moreover, the technical challenges of reasoning are extremely high. On one hand, high accuracy is required, while on the other hand, high throughput is needed to reduce costs, along with achieving low latency. Building a computing system that simultaneously possesses high throughput and low latency is highly challenging, especially as the context length of applications and the complexity of models continue to grow. With the rise of multimodal AI, reasoning technologies are continuously innovating, and this demand is driving the unique advantages of NVIDIA's architecture.

NVIDIA's CUDA ecosystem allows developers to innovate rapidly and ensures compatibility and reliability of technology. The innovative achievements of developers can be quickly deployed across NVIDIA's computing platforms globally, from data centers to edge devices and robotic systems. It is this comprehensive deployment capability and ecological advantage that enables NVIDIA to maintain a leading position in the reasoning field and drive industry development.

Q: What is the situation regarding the network business and confidence in Spectrum - X?

A: From the perspective of the network business, we have shown significant year-on-year growth. Since the acquisition of Mellanox, we have consistently focused on deep integration with the data center business. The network business plays a crucial role in this, and we have continuously enhanced our ability to combine it with data center systems, performing well.

The quarter-on-quarter decline this quarter is merely a temporary adjustment, and we expect to see growth restored in the future. As preparations for Blackwell and more systems progress, not only is the demand for existing network technologies continuing to grow, but many large systems will gradually adopt our newly developed network solutions, further driving business development. We are confident that Spectrum - X will reach the previously mentioned billion-dollar scale.

Q: You mentioned that the sovereign demand scale has reached a low double-digit billion-dollar level. Is there any recent progress in this regard? Additionally, can you explain the supply constraints in the gaming business?

A: Regarding sovereign AI, this area is one of the key drivers of growth, especially after the rise of generative AI. Countries are committed to building foundational models that align with their own languages and cultures and applying them to local enterprises. We also mentioned several relevant examples in this conference call. These sovereign AI projects and future pipelines remain stable, and the construction of regional cloud deployments and AI factories provides more growth opportunities in this area. The growth of sovereign AI is not limited to Europe but is gradually expanding in the Asia-Pacific region. As for the gaming business, we are currently working hard to coordinate the supply of various products. Due to the rapid growth in market demand for gaming products, the supply this quarter is relatively tight, but as the calendar year transitions into a new phase, supply capacity will gradually return to normal. Although there is significant supply pressure this quarter, we expect to restore the supply rhythm soon.

Q: Will the increase in Blackwell supply accelerate growth again, and what about the situation in the first half of next year? How will the change of the U.S. government and the situation in China affect the business?

A: The company provides quarterly guidance and is currently working to meet the supply demand for Blackwell this quarter; regarding the change of the U.S. government and the situation in China, the company will support the new government, comply with relevant regulations, and do its utmost to meet customer needs and compete in the market.

Q: What is the computational allocation and growth focus on pre-training, reinforcement learning, and inference in the AI ecosystem?

A: Computational resources are primarily concentrated in the pre-training phase of the foundational model, as later training (including reinforcement learning and inference-related technologies) is still in its early stages. Although inference costs can be reduced through optimization in both pre-training and later training, certain tasks require real-time thinking and contextual responses, making the computational demands during the inference phase unavoidable.

With the development of foundational models, especially with the introduction of multimodal foundational models, the amount of video data they need to process is enormous, measured in petabytes. Therefore, the computational demands for pre-training, later training, and inference time will continue to expand. In the future, computational capacity will continue to extend in these three directions, further validating the rationale for large-scale computational demands.

To address this trend, significant improvements in computational performance need to be continuously promoted, reducing the overall costs of AI applications through a doubling effect while enhancing efficiency, thereby accelerating the development of the AI revolution.

Risk disclosure and statement of this article: Dolphin Investment Research Disclaimer and General Disclosure