Accelerating computation, global data centers are set to double again (NVIDIA 4QFY24 Summary)

On February 22, Beijing time, NVIDIA (NVDA.O) released its 2024 fiscal fourth-quarter earnings report (as of January 2024) after the U.S. stock market closed.

For a detailed summary of NVIDIA's 2024 Q4 earnings call, please refer to the article "NVIDIA: AI Dominance, the True King of Chips."

Key Highlights of the Earnings Report:

• Total revenue reached $22.1 billion, a 22% increase MoM and a 265% increase YoY, surpassing analysts' expectations of $20.41 billion. Net profit was $12.3 billion, a 765% YoY increase, with adjusted earnings per share of $5.16, higher than the expected $4.59. The adjusted profit margin was 76.7%, exceeding analysts' expectations of 75.4%. Data center revenue was $18.4 billion, surpassing the market's expectation of $17.21 billion.

Detailed Content of the Earnings Call:

Executive Statements:

FY24Q4 performance significantly exceeded expectations. Revenue was $22.1 billion, a 22% MoM increase and a 265% YoY increase, higher than analysts' expected $20.41 billion. Net profit was $12.3 billion, a 765% YoY increase, with adjusted earnings per share of $5.16, higher than the expected $4.59. The adjusted profit margin was 76.7%, surpassing analysts' expectations of 75.4%. Data center revenue was $18.4 billion, exceeding the market's expectation of $17.21 billion.

FY25Q1 guidance also significantly exceeded expectations. a) Total revenue is expected to be $24 billion, with a fluctuation of 2%. The expected MoM growth in data centers and professional editions will be partially offset by the seasonal decline in gaming. GAPP and non-GAAP gross margins are expected to be 76.3% and 77%, with a fluctuation of 50 basis points. Similar to the fourth quarter, the first-quarter gross margin benefits from favorable component costs. After the first quarter, we expect the gross margin to return to around 75% for the remaining time. GAPP and non-GAAP expenses are expected to be around $3.5 billion and $2.5 billion, respectively. b) GAPP and non-GAAP operating expenses for the 2025 fiscal year are expected to increase by around 30%. GAPP and non-GAAP other income is expected to be around $150 million, excluding gains or losses from non-associated investments. GAPP and non-GAAP tax rates are expected to be 17%, plus or minus 1% (excluding any discrete items).

Accelerated computing and generative AI have reached a tipping point. Demand from enterprises, industries, and countries worldwide is surging, including large cloud service providers, specialized GPU providers, enterprise software, consumer internet companies, automotive, finance, and healthcare industries. Last year, about 40% of data center revenue came from AI inference, with over half of data center revenue coming from large cloud providers.

The demand for next-generation (chip) products far exceeds the supply.

We have started supplying alternative products to the Chinese market in small quantities. The latest quarterly data shows that the revenue share from the China region's data centers has dropped to single digits, and it is expected to remain in a similar range in the first quarter.

Overall supply is improving, but still in short supply. Supply constraints will continue throughout the year, and demand cannot be met in the short term.

Supply shortages will lead to the shipment of the next-generation chip B100 later this year.

2.2, Q&A Analyst Q&A

Q: What changes have there been in the expectations for the data center business in 2024 and 2025 in the past three months? In the data center business, what are the company's views and expectations on emerging areas such as software, sovereign AI, etc.? There have been recent articles mentioning that NVIDIA may participate in the ASIC market. How does the company plan to compete and develop in this market in the coming years?

A: The conditions for continued growth in the data center business from 2024 to 2025 and beyond are excellent. NVIDIA is in a leading position in accelerated computing, which significantly improves LNG efficiency and data processing costs while maintaining incredible speed. With the rise of generative AI, a new type of data center has emerged - the AI generation factory. Each region has its own language, knowledge, history, and culture, so there is a desire to train and create digital intelligence using their own data. This sovereign AI infrastructure is being built globally, in countries like Japan, Canada, and France. There was no direct answer to whether NVIDIA is involved in the ASIC market. NVIDIA's inference business segment has grown significantly, estimated to have grown by about 40%. NVIDIA is diversifying into new markets, including large CSPs (cloud service providers) that are still expanding, and a new category has emerged - GPU-specialized CSPs that specifically provide NVIDIA AI infrastructure. In addition, enterprise software platforms are also deploying AI services, such as ServiceNow, Adobe, SAP, etc.

Q: Regarding the data that 40% of the total revenue comes from inference, how has this proportion grown? The growth of large language models (LLMs) in inference? How reliable is the measurement of inference and training usage on GPUs?

A: Recommendation systems play a crucial role on the internet, responsible for recommending content such as news, videos, music, and products to users. With trillions of content on the internet and limited space on mobile screens, recommendation systems need to compress and display all this information to users. In the past, recommendation systems mainly relied on GPU and CPU methods. However, with the development of deep learning, especially the rise of generative AI, recommendation systems now directly depend on GPU acceleration. This includes steps like embedding, nearest neighbor search, reordering, and generating enhanced information. Now, GPUs are an indispensable part of recommendation systems, and almost every large company needs to run these large recommendation systems.When it comes to various generative models like ChatGPT and Mid Journey, as well as collaborations with companies such as Getty and Adobe's Firefly, these models have all been brand new in the past year, generating a large amount of content for consumers.

Q: Expectations for the next generation of products to be affected by supply constraints are based on what reasons? Why, even as Hopper supply gradually loosens, are there still supply constraints? How long is it expected for these supply constraints to last? When will the research on these supply constraints be conducted?

A: NVIDIA's overall supply chain is improving, including the supply of components from wafers, packaging, memory to various power regulators, transceivers, network equipment, and cables. NVIDIA's Hopper GPU is extremely complex, containing 35,000 parts, weighing 70 pounds, and being rightfully called an AI supercomputer. The back-end systems of data centers are very complex, including the cable layout of network systems, which is the densest and most complex seen in the world. InfiniBand business has grown fivefold year-on-year, and the support of the supply chain is excellent. Therefore, overall supply is improving, and demand is expected to remain strong. Whenever a new product is launched, there is a growth process from zero to mass production. This cannot be achieved overnight and requires a gradual increase in production. Currently, the new generation products H200 and Spectrum-X are being launched. In the short term, due to production ramp-up, it may not be possible to meet all demands. Spectrum-X is a brand-new product designed to enter the Ethernet market, and compared to InfiniBand, it is more suitable as an expansion system. InfiniBand will become NVIDIA's infrastructure dedicated to AI, while Spectrum-X will be an optimized AI network. The demand for all new products exceeds supply, which is a common feature of new products. NVIDIA's supply is steadily increasing.

Q: Faced with significantly higher demand than supply, how does NVIDIA consider the allocation of products? How does NVIDIA and its companies consider fairness in allocating their products across industries and customers (many of which are competitors)? How do they balance the interests of the company and the industry?

A: NVIDIA has been cooperating with its customers for many years, assisting them in setting up GPU instances in the cloud and internally. Many providers have multiple products to meet the various needs of their end customers and their internal needs. Providers anticipate the new clusters needed and continuously discuss the Hopper architecture and future technology trends with NVIDIA in advance to understand and predict future demands. This is an ongoing process involving the products that providers will purchase, the products being built, and the products ultimately used by end customers. The relationships NVIDIA has established with its customers and their understanding of complexity help NVIDIA in product allocation and improving communication with its customers. NVIDIA maintains transparency with its cloud service providers (CSPs), sharing product roadmaps and transition information, giving CSPs confidence in knowing when, where, and how many of which products to place.The company is committed to fair product distribution, avoiding waste, and seeking opportunities to connect partners and end-users.

Q: How to effectively convert backlogged orders into actual revenue? NVIDIA has significantly shortened its product delivery time, does this mean a corresponding reduction in financial commitments to suppliers (such as inventory purchase commitments and prepayments)? If we combine inventory, purchase commitments, and prepayments, it seems to have decreased overall, how should we interpret this?

A: NVIDIA's strategy in inventory management is that once new inventory enters, they will make efforts to deliver it to customers immediately according to their allocation strategy. This indicates NVIDIA's emphasis on meeting customer delivery schedules. Purchase commitments consist of multiple components, including components for manufacturing and potential capacity that may need to be procured. The procurement deadlines for these components and capacities vary, some may be for the next two quarters, while others may be for several years. NVIDIA's prepayments are to ensure reserved capacity at several key manufacturing suppliers. These prepayments are also designed based on different terms. Although the total amount of NVIDIA's inventory, purchase commitments, and prepayments seems to remain unchanged, in reality, they each have different terms. Sometimes, NVIDIA needs to purchase components for long-term supply or capacity for their construction, which results in different terms for these payments.

Q: Colette Kress has mentioned that the gross margin should return to around 75%, what are the reasons and basis behind this forecast? The new products contain high compute density (HCM) content, is this a major factor affecting the gross margin forecast, and besides HCM content, what other factors may drive this forecast?

A: NVIDIA's gross margin is relatively high in Q4 and Q1 mainly due to the optimization of component costs in the supply chain, involving various aspects of computing and networking businesses, and multiple stages of their manufacturing process. For the upcoming fiscal year, NVIDIA expects the gross margin to return to around 75%, consistent with their levels before Q4 and Q1. NVIDIA believes that the balance of product portfolio will be their biggest driver in the next year. This means they will adjust their product line according to market demand to maximize profits.

Q: When considering the one million times improvement in GPU computing over the past decade and similar expected improvements in the future, how do customers view the value of the company's current investments in long-term use? In other words, how will their current training clusters transform into impactful clusters in the future? How does NVIDIA view this trend?

A: The platform has two main features: acceleration and programmability, which are the reasons they can achieve significant performance improvements. The company is the only one that has continuously supported various deep learning architectures from the very beginning (such as the early stages of CNN and AlexNet), including RLS, Transformers, Vision Transformers, Multi-modality Transformers, etc.The company has been able to invent new architectures and technologies, such as their intensive courses, Transformer engine, new numerical formats, and processing structures, all of which have been implemented in various 10th generation courses. The company not only supports its installed infrastructure but also continuously integrates new software algorithms and inventions into these infrastructures. They can provide software support and continuous improvement for the installed infrastructure, as well as create new technologies like the Hopper Transformer engine and implement them in future products. The company's future products may bring astonishing breakthroughs in large language models, some of which will be achieved through CUDA software and provided to the installed infrastructure.

Q: What specific products are the alternative solutions being shipped in the Chinese market? Why should we not expect other alternative solutions to enter the market?

A: Due to restrictions imposed by the U.S. government, the company needs to reconfigure its products to adapt to the specific requirements of the Chinese market. These restrictions mainly involve the latest accelerated computing and AI technologies. The company understands these limitations and has redesigned its products in a non-software way. This process takes time, so the company has reset its product supply to the Chinese market. They have started sending samples to customers in China and plan to compete and succeed in this market. Despite the significant decline in the company's business in the last quarter due to the suspension of shipments in the market, they have resumed shipments. Although they have started shipping in the market, performance for this quarter is expected to be similar to the previous quarter. The company hopes to compete fully and succeed in the market in the future, but specific results remain to be seen.

Q: What are the different components of the software business? Where does the growth of this business come from?

A: Success in the software business stems from the growth of accelerated computing in the cloud, especially through collaborations with cloud service providers. The company collaborates with cloud service providers to help operate and manage their businesses and provide extensive technical support when issues arise. Software is crucial in accelerated computing, which is the fundamental difference between accelerated computing and general computing. Generative AI is driving every enterprise and software company to adopt accelerated computing. This is necessary because relying solely on general computing is increasingly challenging to sustain throughput improvements. We are responsible for managing, optimizing, patching, tuning, and installing base optimizations for the software stacks of these companies, containerizing them into the NVIDIA AI Enterprise stack. NVIDIA AI Enterprise is described as a runtime environment, similar to an operating system, specifically designed for artificial intelligence. The company charges $4,500 per GPU per year, and it is expected that all enterprises and software companies globally deploying software in public cloud, private cloud, and on-premises environments will use NVIDIA AI Enterprise, especially those using NVIDIA GPUs. This business is expected to become a very important business in the future.Huang Renxun's Summary: The computer industry is undergoing two platform shifts simultaneously (accelerated computing and generative AI). The foundation of data center installations worth trillions of dollars is shifting from general computing to accelerated computing. Each data center will be accelerated so the world can meet computational demands by increasing throughput while managing costs and energy. NVIDIA has achieved incredible acceleration, introducing a new computing paradigm, generative AI, where software can learn, understand, and generate any information from human language to biological structures and the 3D world. We are now at the beginning of a new industry where AI-specific data centers process vast amounts of raw data, refining it into digital intelligence. Similar to the transition from steam power plants in the last industrial revolution, NVIDIA's AI supercomputers essentially serve as the AI power plants of this industrial revolution. Every company and industry is fundamentally built on its proprietary business intelligence and, in the future, its proprietary generative AI. Generative AI has initiated a new investment cycle to build the infrastructure for the next $1 trillion AI generative factory. We believe that these two trends will double the global data center infrastructure installations in the next 5 years, representing annual market opportunities of hundreds of billions of dollars. This new AI infrastructure will open up a whole new world of applications that are currently impossible. We have embarked on the AI journey with hyperscale cloud providers and consumer internet companies. Now, industries ranging from automotive, healthcare, financial services to industrial telecommunications, media, and entertainment are all involved. NVIDIA's full-stack computing platform with industry-specific application frameworks and a vast developer and partner ecosystem provides us with speed, scale, and influence, helping every company in every industry become an AI company.

At the upcoming GTC in San Jose next month, we have a lot to share with you, so be sure to join us. We look forward to updating you on the latest developments next quarter.

End of Document

Risk Disclosure and Disclaimer of this Article: Dolphin Research Disclaimer and General Disclosure