NVIDIA: Sovereign AI to bring in billions of revenue (FY25Q1 earnings call)

NVIDIA (NVDA.O) released its first-quarter financial report for the 2025 fiscal year (ending April 2024) after the U.S. stock market on May 23, Beijing time:

The following is a summary of the NVIDIA 2025 first-quarter financial report conference call. For an interpretation of the financial report, please refer to " NVIDIA: The Strongest Stock in the "Universe", Non-stop Explosive Gifts "

I. Recap of key information from NVIDIA's financial report:

II. Details of the NVIDIA (NVDA.US) financial report conference call

2.1. Key points from executive statements:

1. Operational highlights:

① Data Center Business:

a. Revenue reached $22.6 billion, hitting a historical high, with a 23% increase quarter-on-quarter and a 427% increase year-on-year. Growth was mainly driven by strong demand for the NVIDIA Hopper GPU computing platform.

b. Large CSPs accounted for nearly 45% of data center revenue, while inference accounted for about 40% of data center revenue. Enterprise customer support, such as Tesla's AI training cluster expansion, is expected to make the automotive industry the largest enterprise vertical in data centers.

② Gaming and AI PC Business: Revenue of $2.65 billion, an 18% year-on-year increase. The installation base of GeForce RTX GPUs exceeded 100 million, catering to gamers, creators, and AI enthusiasts.

③ Professional Visualization Business: Revenue of $427 million, a 45% year-on-year increase. Omniverse will drive the next wave of professional visualization growth through collaborations with major industrial software manufacturers.

④ Automotive Business: Revenue of $329 million, an 11% year-on-year increase. Supporting Xiaomi in launching its first electric car, the SU7 sedan, using the NVIDIA DRIVE Orin AI car computer.

⑤ Global Markets: Sovereign investments in AI are diversifying data center revenue. Japan has invested over $740 million to build sovereign AI infrastructure. European countries like France and Italy are constructing powerful cloud-native AI supercomputers. NVIDIA-accelerated AI factories are being built in locations like Singapore.

⑥ Product Updates:a. Hopper GPU: Continued demand growth, H100's inference speed increased by 3 times, H200 is about to be produced and shipped.

b. Grace Hopper Super Chip: Providing energy-efficient AI processing capabilities for new supercomputers worldwide.

c. InfiniBand: Network growth year-on-year, supply chain issues leading to a moderate continuous decline.

d. Spectrum-X Ethernet Network Solution: AI optimization, opening up a completely new market.

1. New Product Releases:

a. NVIDIA Inference Microservices (NIM): Providing secure and performance-optimized containers to accelerate the deployment of generative AI applications.

b. Blackwell GPU Architecture: Training speed 4 times faster than H100, inference speed 30 times faster, overall cost and energy consumption reduced by 25 times.

1. Financial Highlights

2. Shareholder Returns: Returning $7.8 billion to shareholders in the form of stock buybacks and cash dividends, announcing a 10-for-1 stock split and a 150% increase in dividends.

3. Second Quarter Guidance:

1. Huang Renxun's View on the Importance of Transformation:

With the arrival of the next industrial revolution, NVIDIA is collaborating with companies and countries globally to lead the transformation from traditional data centers to accelerated computing, establishing AI factories to produce artificial intelligence. The application of AI is significantly increasing productivity across industries, helping companies reduce costs and energy consumption while expanding revenue opportunities.

As early adopters of generative AI, cloud service providers have accelerated workloads and saved capital and electricity through collaboration with NVIDIA. Tokens generated by the NVIDIA Hopper platform are driving growth in AI service revenue, while NVIDIA cloud instances are attracting a large number of developers.

The growth of the data center business is benefiting from strong demand for generative AI training and inference on the Hopper platform, a demand that continues to grow as model learning becomes more multimodal. The rapid growth of inference workloads indicates that generative AI is driving the transformation of full-stack computing platforms, which will change the way computers interact.

NVIDIA is transitioning from an information retrieval mode to a computational mode of generating answers and skills. AI will understand context and intent, possess the ability to know, reason, plan, and execute tasks. This marks a fundamental change in the way computers work and function, from general-purpose CPUs to GPU-accelerated computing, from instruction-driven software to models that understand intent, from information retrieval to executing skills, from producing software to generating tokens, and creating digital intelligenceIn addition, generative AI has expanded to consumer internet companies, enterprises, sovereign AI, automotive, and healthcare clients, opening up multiple vertical markets worth billions of dollars. The comprehensive production of the Blackwell platform provides the foundation for trillion-parameter-scale generative AI, combined with technologies such as Grace CPU and NVLink, offering a richer and more complete AI factory solution. The company believes that sovereign AI this year can bring in revenue close to several tens of billions of dollars.

Spectrum-X has opened up a new market for NVIDIA, bringing large-scale AI into data centers using only Ethernet. NVIDIA NIMs, as a new software product, provides enterprise-optimized generative AI that can run on cloud, on-premises data centers, and on RTX AI PCs through a wide ecosystem of partners.

NVIDIA is ready for the next wave of growth, from Blackwell to Spectrum-X to NIMs, the company is at the forefront of this transformation.

2.2, Q&A Analyst Q&A

Q: Blackwell commented that the product has been fully put into production, which means the product is no longer in the sampling stage. So, if the product has been put into production, what will be the delivery and shipment time in the hands of customers?

A: We will start shipping in the second quarter and gradually increase in the third quarter. Customers are expected to complete data center construction in the fourth quarter. This year we will see a large amount of revenue from Blackwell.

Q: What are the differences in deployment between Blackwell and Hopper? Liquid cooling is still a first in large-scale applications, bringing engineering challenges at the node and data center levels. Will these complexities affect the deployment transition period? How do you view the development of this process?

A: The Blackwell platform offers diverse configurations, including air cooling, liquid cooling, x86 and Grace, InfiniBand, and Spectrum-X technology, as well as the large-scale NVLink architecture showcased at GTC. Customers will be able to easily transition from Hoppers to H100, H200, and B100. The Blackwell system design ensures electrical and mechanical compatibility with existing data centers while ensuring that the software stack on Hopper can run efficiently on Blackwell.

We have been actively preparing the entire ecosystem for liquid cooling technology and have had in-depth discussions with various parties about the Blackwell platform. CSPs, data centers, ODMs, system manufacturers, and supply chain partners, including cooling and data center supply chains, have all prepared for the arrival of Blackwell and its integration with Grace Blackwell 200. The performance of GB200 is expected to be outstandingQ: How do you ensure that products are fully utilized and avoid premature procurement or hoarding due to factors such as supply shortages and competition? What mechanisms have you established in the system to ensure that monetization and shipment growth are synchronized?

A: The demand for GPUs in data centers is enormous, and we are working hard every day to meet this demand. Applications such as ChatGPT, GPT-4o, multimodal Gemini, Anthropic, and CSPs are consuming all the GPUs available in the market. In addition, there are numerous generative AI startups in fields such as multimedia, digital characters, design tools, productivity applications, digital biology, etc., all driving the demand for GPUs. Customers are urgently requesting us to deliver systems as soon as possible.

In the long term, we are completely redesigning the way computers work, which will be a major platform shift. Computers will no longer be simply instruction-driven devices, but intelligent computers that can understand user intent, reason, and provide solutions. This transformation will impact the global computing architecture, and even the PC computing stack will undergo a revolution. What we are currently seeing is just the tip of the iceberg, and the work being done in our labs and collaborations with global startups, large enterprises, and developers will bring extraordinary results.

Q: With the high demand for H200 and Blackwell products, do you expect a slowdown in demand for Hopper and H100 when transitioning to these new products? Will customers wait for these new products? Or do you believe that the demand for H100 is sufficient to sustain growth?

A: This quarter, the demand for Hopper continues to grow, and we expect demand to exceed supply for a long time during the transition to H200 and Blackwell. All parties are eager to deploy their infrastructure because they want to save costs and achieve profitability as soon as possible.

Q: Many cloud customers, while working with you, have also announced new or updated internal projects. How do you view them as competitors? Are they mainly limited to handling internal workloads, or could they potentially expand into broader areas in the future?

A: NVIDIA's accelerated computing architecture differs from competitors in many ways. Firstly, it can comprehensively handle the entire process from unstructured data processing to training preparation, structured data processing, data framework processing, training, and inference. Inference has shifted from simple detection to generation, which requires fundamentally different processing architectures. Our TensorRT-LLM has been well received as it triples the performance of our architecture on the same chip, demonstrating the richness of our architecture and software.

Secondly, the versatility of the NVIDIA platform means it can be used for all computing modes such as computer vision, image processing, computer graphics, etc., providing a sustainable solution under the constraints of computing costs and energy inflation. When general-purpose computing reaches its limit, accelerated computing is the only way forward, helping to save computing costs and energy, making our platform the choice with the lowest TCO for data centersThird, NVIDIA has a layout in all cloud platforms, providing developers with a ubiquitous development platform, whether on local, cloud, or various sizes and shapes of computers.

Finally, NVIDIA not only manufactures chips but also builds an AI factory, which is a systemic issue that requires optimizing all chips to work together as a system. Through system optimization, we have significantly improved performance, thereby creating tremendous value for customers. In today's high infrastructure costs, the highest performance solution also means the lowest TCO.

Q: Customers' competitiveness in products that are currently heavily invested in will rapidly decline, even faster than the product's depreciation cycle. Can you discuss how customers will handle their existing large installed base as they transition to Blackwell, where these products, although software-compatible, are far inferior in performance to the new generation of products? How do you view customers' coping strategies in this transition process?

A: Firstly, different stages of construction progress will bring different feelings. Currently, customers are only in the early stages of construction, so they need to advance as quickly as possible. The arrival of Blackwell will bring significant improvements, and after that, we have other new products that will be launched one after another. Our product updates follow an annual rhythm, and customers can continue to advance according to their construction progress.

Secondly, customers need to continue investing to achieve performance averaging. They need to profit immediately while saving costs, and time is extremely valuable to them. For example, the first company to reach a new milestone will announce a breakthrough AI technology, while the following company may only announce a 0.3% performance improvement product. Therefore, it is crucial to be a company that continues to provide breakthrough AI technology.

Third, we can quickly advance and optimize the technology stack because we build and monitor the entire data center. We can accurately identify bottlenecks and optimize them, rather than relying on guesswork. The systems we deliver perform exceptionally at scale because we build the entire system.

Furthermore, we can deconstruct the built AI infrastructure and integrate it into their data centers according to customer needs to ensure optimal performance. Our deep understanding of the entire data center scale and the ability to build each chip from scratch enable us to ensure that each generation of products maximizes utility.

Q: Currently, the workloads that are driving demand for your solutions are mainly neural network training and inference, which seems to be a relatively limited workload type, possibly more suitable for customized solutions. So, are general-purpose computing frameworks facing greater risks? Or is the diversity and rapid evolution of these workloads sufficient to support traditional general-purpose computing frameworks?

A: NVIDIA's accelerated computing is versatile but should not be seen as general-purpose computing. For example, we are not good at executing spreadsheet tasks designed for general-purpose computing. The control loop code of the operating system is not the best choice for accelerated computing. Our platform can accelerate a variety of applications, but they share common characteristics in parallel processing and multithreading, where 5% of the code may account for 99% of the runtimeOur platform's versatility and system-level design have been the reasons why many startups have chosen us in the past decade. These companies' architectures are often more fragile, and when faced with emerging technologies such as generative AI or fusion models, they need solutions that can adapt to the entire field, rather than just a single model. As AI advances, these solutions need to be able to adapt to the continuous development and expansion of software.

We believe that in the next few years, these models will expand by millions of times, and we are prepared for this. The versatility of our platform is crucial, and overly fragile or overly specific solutions, such as FPGA or ASIC, while potentially more effective for specific tasks, do not have the broad applicability of general-purpose computers.

Q: Could you share your thoughts on the launch of the H20 product in China and its impact on demand in a supply-constrained environment? How do you balance customer demand between H20 and other Hopper products? Additionally, could you elaborate on the potential impact this may have on sales and gross margins in the second half of the year?

A: We are committed to respecting and serving every customer well. Despite a decrease in our business in the Chinese market compared to the past, and with increased market competition due to technological limitations, we will still make every effort to meet customer demands. Our overall view of the market is that demand continues to exceed supply, especially for H200 and Blackwell products, which will be particularly evident towards the end of the year.

Q: The GB200 system is currently showing strong market demand. Looking back, NVIDIA has sold a large number of HGX boards and GPUs, while the system business has been relatively small. What are the reasons for the current growth in system demand? Is it only related to TCO, or are there other factors at play?

A: We have maintained our sales strategy for the GB200. We reasonably deconstruct components and integrate them with computer manufacturers. This year, the Blackwell platform will launch 100 different computer system configurations, far exceeding the peak period of Hopper. You will see various versions including liquid cooling, air cooling, x86, Grace, and more. These systems are provided by our many excellent partners. The Blackwell platform significantly expands our product line, integrates CPUs, and increases computing density. Liquid cooling technology will save a significant amount of power costs for data centers while improving efficiency, making it a better solution. We provide more data center components, allowing data centers to achieve higher performance, including network switches and NICs. We now offer Ethernet, enabling customers who are only familiar with Ethernet operations to adopt NVIDIA AI on a large scale because they have the corresponding ecosystem.

Overall, the Blackwell platform is more comprehensive in many aspects, allowing us to provide customers with a richer range of products and services.

Q: Despite Intel and AMD providing high-quality x86 solutions as excellent partners, can NVIDIA also offer unique advantages in the emerging AI workload field, especially when other competitors are facing more challenges?A: With the joint efforts of our x86 partners, we have built many outstanding systems. However, the Grace CPU allows us to achieve what existing system configurations cannot: the memory system between Grace and Hopper is collaborative and interconnected, almost forming a super chip, with interface speeds reaching several terabits per second. Grace uses the first data center-grade low-power LPDDR memory, which significantly saves power on each node.

Furthermore, we can independently design the architecture of the entire system, creating products with a vast NVLink domain, crucial for the inference of next-generation large language models. For example, the GB200 has a 72-node NVLink domain, equivalent to 72 Blackwell connected into a massive GPU. Therefore, out of necessity for architecture, software programming, and system design, we developed Grace Blackwell.

As you saw yesterday, Satya announced the next generation PC - Copilot-Plus PC, which runs excellently on NVIDIA RTX GPUs, already applied in laptops. At the same time, it also perfectly supports ARM architecture, providing new opportunities for PC system innovation.

Q: Faced with increasing competition from GPUs and custom ASICs, how do you view NVIDIA's future innovation speed? What challenges do we need to address in the next decade?

A: I can confirm that we have another chip after Blackwell, and we maintain an annual update pace. In terms of networking technology, we also iterate rapidly, just announcing Spectrum-X for Ethernet. We are excited about the future development of Ethernet and have strong ecosystem support from partners like Dell who will bring Spectrum-X to market.

We are committed to developing three key technologies: NVLink as a computing structure for a single computing domain, InfiniBand as a computing structure, and Ethernet as a network computing structure. All three aspects will rapidly evolve, bringing new switches, NICs, features, and software stacks. We will see a large number of chip updates, including new CPUs, GPUs, network NICs, and switches.

Most importantly, all these new technologies will run CUDA and our entire software stack. This means that investments in our software stack today will automatically appreciate over time, improving performance. Similarly, investments in our architecture will expand as it is applied in more cloud platforms and data centers.

I believe that our pace of innovation will enhance technological capabilities and reduce TCO. NVIDIA's architecture will support this new era of computing and drive a new industrial revolution, where we will no longer just manufacture software, but massively manufacture artificial intelligence tokens

Risk Disclosure and Disclaimer for this Article: Dolphin Investment Research Disclaimer and General Disclosure