The conference call for NVIDIA's third quarter earnings is being conducted by operator JL, with Simona Jankowski as the main speaker. The call is being webcast live and will be available for replay. Forward-looking statements may be made during the call and non-GAAP financial measures will be discussed. Colette Kress will be leading the call.

In the second paragraph, Colette Kress discusses the record-breaking performance of NVIDIA in Q3, with revenue of $18.1 billion, up 34% sequentially and 200% year-on-year. The company's Data Center segment saw the most significant growth, driven by the adoption of NVIDIA HGX platform and InfiniBand networking, making it the go-to choice for AI supercomputers and data center infrastructure. The company's investments in AI infrastructure have led to strong demand for NVIDIA's accelerated computing, particularly in training and inferencing large language models, deep learning, and generative AI applications. The enterprise wave of AI adoption is also beginning, with companies like Adobe, Databricks, Snowflake, and ServiceNow adding AI copilots and systems to their platforms. Cloud service providers also contributed to the growth of the Data Center segment in the quarter.

NVIDIA has seen strong demand for its H100 Tensor Core GPU instances from hyperscale CSPs and GPU-specialized CSPs globally. The company has increased supply to meet this demand and plans to continue doing so next year. However, new export control regulations from the U.S. government will require licenses for the export of certain products to China and other countries, which could lead to a decline in sales to these destinations. The regulations aim to promote U.S. technology leadership and economic growth, and have different requirements depending on the performance level of the product.

The company is expanding its Data Center product portfolio to offer compliance solutions for different regulatory categories. They are working with customers in China and the Middle East to obtain licenses from the US government, and are also collaborating with India's government and major tech companies to boost their sovereign AI infrastructure. There is a growing need for national investment in compute capacity to support local generative AI ecosystems, and the company sees this as a multi-billion dollar opportunity. The majority of revenue in Q3 was from the NVIDIA HGX platform, with lower contribution from the Ampere GPU architecture. The company has also begun shipping the new L40S GPU and the GH200 Grace Hopper Superchip, which is expected to be a multi-billion dollar product line. Grace Hopper instances are now available at certain cloud providers and will soon be available at Oracle Cloud.

Grace Hopper is gaining traction in the supercomputing market, with shipments to major institutions like Los Alamos National Lab and the Swiss National Supercomputing Center. The UK government and German supercomputing center, Julich, have also announced plans to build powerful AI supercomputers using Grace Hopper Superchips. The combined AI compute capacity of all the supercomputers built on Grace Hopper is estimated to exceed 200 exaflops next year. Inference is a major driver of data center demand, and NVIDIA offers the best performance and cost efficiency for AI inference. The company has also announced the H200, which will be the first GPU to offer HBM3e, further increasing performance for generative AI and LLMs. Compared to the A100, the H200 offers an 18x performance increase for inferencing models like GPT-3, without any increase in latency.

The article mentions that major cloud service providers such as Amazon Web Services, Google Cloud, Microsoft Azure, and Oracle Cloud will be offering H200-based instances next year. At the Microsoft Ignite event, NVIDIA announced a collaboration with Microsoft to provide an AI foundry service for developing and tuning custom AI applications on Azure. SAP and Amdocs are the first customers of this service. Microsoft will also launch new confidential computing instances based on the H100. InfiniBand, a critical component for training large language models, saw exceptional demand and is being expanded into the Ethernet space with the launch of Spectrum-X in Q1 next year.

NVIDIA's Spectrum-X offers 1.6x higher networking performance for AI communication compared to traditional Ethernet options. Their software and services offerings are seeing strong adoption, with an annualized revenue run rate of $1 billion expected by the end of the year. The company sees growth opportunities in their DGX cloud service and NVIDIA AI Enterprise software, with recent partnerships including Gentech for AI drug discovery. Gaming revenue has also seen significant growth, with a 15% sequential increase and over 80% year-on-year increase, driven by strong demand for NVIDIA RTX and DLSS technology. The RTX ecosystem now includes over 475 games and applications, and generative AI is becoming a key application for high-performance PCs.

NVIDIA has released TensorRT-LLM for Windows, which speeds up on-device LLM inference by 4x. Their GeForce NOW cloud gaming service has surpassed 1,700 titles and their Pro Vis. Revenue has increased by 10% sequentially and 108% year-on-year. NVIDIA RTX is the preferred platform for professional design, engineering, and simulation use cases, with AI emerging as a demand driver. They have launched new workstations optimized for AI workloads and continue to make progress on their Omniverse software platform. Mercedes-Benz and Oxxon are using Omniverse for their manufacturing processes. NVIDIA has also announced two new Omniverse Cloud services for automotive digitalization.

In the fourth quarter of fiscal 2024, NVIDIA's revenue is expected to be $20 billion, driven by strong demand in the Data Center segment. Gross margin is expected to be 74.5% to 75.5%, while operating expenses are expected to increase due to higher compensation and benefits. Other income and expenses are expected to be an income of $200 million, and the tax rate is expected to be 15%. More information can be found on the company's IR website.

The company will be attending several upcoming events for the financial community, including conferences and an earnings call. They are expecting a decrease in revenue from China due to export controls, but do not have a clear understanding of the impact. The question of where the company is in the adoption curve for its products in the generative AI market is also raised.

Nvidia is working to expand their Data Center product portfolio to offer new regulation compliance solutions without a license in the coming months, but it will not significantly impact revenue in Q4. CEO Jensen Huang discusses the impact of generative AI, which has changed the way data is accessed and processed. This has led to the development of copilots and assistants, which will replace manually used tools and be integrated into teams of AIs, becoming the modern version of enterprise business software.

The transformation of software is driving changes in hardware, particularly in terms of accelerated computing which can save energy, time, and cost. This has led to the emergence of new data centers known as AI factories, which focus on training models and generating AI. These factories are being built globally and are being used by large language model start-ups, generative AI companies, and consumer Internet companies. Additionally, partnerships with enterprise software companies are being formed to incorporate AI into their platforms.

The paragraph discusses the growth and significance of GPU specialized CSPs and sovereign AI infrastructures around the world. It also mentions the increasing use of generative AI in various industries and regions. The question then shifts to the networking side of the business and its strong growth rates, with the mention of a new networking platform and the dominance of InfiniBand in large scale AI factories. The speaker predicts that the networking business will continue to grow and become even larger in the future.

The use of InfiniBand in AI factories is essential due to its high data rate, low latency, and advanced software-defined capabilities. It also serves as a computing fabric and can perform computations directly on the switch. This results in a significant difference in effectiveness and can save companies millions of dollars in infrastructure costs. InfiniBand's value proposition is undeniable for AI factories, but it is also being utilized in enterprise computing to enable companies to build their own custom AIs. For example, the company is currently creating multiple custom AI models, such as ChipNeMo, for their own use.

The company has invented a platform that extends Ethernet and allows for AI to run in an Ethernet environment. This platform, called Grace Hopper, is in high volume production and is optimized for East-West traffic. It adds to Ethernet and works with the company's Spectrum switch, allowing for some capabilities of InfiniBand. The company plans to go to market through partnerships with large enterprise partners such as HP, Dell, and Lenovo, who will offer the NVIDIA AI stack, NVIDIA AI Enterprise software stack, and Bluefield to enterprise customers. Grace Hopper is expected to expand the company's total addressable market (TAM) and will be used for different applications than traditional H100 applications.

The speaker discusses the potential success of their new data center CPU, named Grace Hopper, which has advanced capabilities in high performance computing and AI. It can simultaneously handle large amounts of fast and large memory, making it useful for tasks such as generative AI models that need to refer to large amounts of data. The speaker also mentions potential customers for the product, such as European supercomputing centers and companies looking to build their own ARM ecosystem. They express confidence in the product's potential for success.

Jensen Huang, CEO of NVIDIA, believes that the company's Data Center division can continue to grow through 2025. This is due to the expansion of their supply chain, which is necessary to produce complex supercomputers like the HGX H100. The team at NVIDIA has successfully scaled up their supply chain to meet demand, and they are also adding new customers and products. Huang is proud of their world-class supply chain and is optimistic about the future growth of the Data Center division.

Different regions around the world are creating their own GPU specialist and sovereign AI clouds to retain their country's knowledge and culture. NVIDIA is expanding into the enterprise market, offering off-the-shelf AI, proprietary AI, and custom AI services through their AI foundry. Enterprises are also building their own custom AI solutions, which NVIDIA plans to serve through their entire stack of systems and software, working with partners such as HP, Dell, and Lenovo.

NVIDIA is seeing the widespread adoption of generative AI technology, starting with start-ups and moving to larger companies. They are also working on industrial generative AI with their NVIDIA AI and NVIDIA Omniverse platforms. Colette mentioned that they will be introducing regulation-compliant products in the next few months, but the contribution to Q4 revenue will be limited. This could potentially lead to reacceleration in growth for Data Center in April and beyond. Jensen also discussed the AI foundry service announcement and how the monetization model will work.

The company is planning to move their new products and is in the process of discussing with customers. They are focused on finding the right balance for their China customers, but it is difficult to predict the outcome. The CEO, Jensen Huang, believes that there is a huge opportunity for AI foundries in the world and they have already announced partnerships with several companies. He also highlights the importance of having AI technology in a foundry, which their company has a strong capability in.

NVIDIA's business model for AI involves three key components: having the best known practices and skills for processing data and creating safe AI models, providing a platform (DGX Cloud) for building and deploying these models, and offering a license (NVIDIA AI Enterprise) for customers to run their custom models. The company's focus is on wholesale, with customers becoming the retail and creating their own monetization models on top of NVIDIA's license. This business model is off to a strong start and is expected to be a significant source of revenue for the company.

Matt Ramsay asked Jensen a two-part question about how NVIDIA is positioned for the evolution of the inference workload as we move towards larger language models, as some investors believe that AI training is NVIDIA's dominant domain and inference will become more competitive in the market.

Jensen Huang discusses the complexity of the inference workload in AI and how CUDA, the programmable GPU, enables acceleration of data processing. He also mentions the success of TensorRT-LLM, an optimizing compiler, and how it has been integrated into various frameworks. The pace of innovation and the large installed base of NVIDIA GPUs contribute to the success of their inference engine.

NVIDIA's platform stability and large installed base make it the preferred choice for software application providers. Data processing is a crucial part of machine learning and NVIDIA accelerates popular frameworks like Spark and Pandas without any additional code. This has generated a lot of excitement among users.

NVIDIA's Pandas was designed for data processing and NVIDIA CUDA is used for data science. The company is focused on ecosystem, strong partnerships and a more aggressive product cadence. This will reduce costs for customers and expand the reach of generative AI. NVIDIA is working with every cloud service provider and adapting to their unique configurations.

The complexity of NVIDIA's platform includes its compatibility with various technologies and segments, its domain specific libraries, and its end-to-end solution for data centers. This has led to market demand in various industries and a large ecosystem of developers, system makers, and distribution partners. Despite the energy required to maintain this platform, NVIDIA's decision to make everything architecturally compatible has been beneficial.

NVIDIA's strong growth is due to the transition from general purpose to accelerated computing and generative AI. This is being driven by large language models, start-ups, consumer Internet companies, and global cloud service providers. The next wave is starting with nations and regional CSPs building AI clouds, and enterprise software companies adding AI assistants to their platforms. This has created the need for a new type of data center, an AI factory, which is driving significant new investment and expanding the traditional data center infrastructure. NVIDIA's H100 HGX and AI software stack are defining the AI factory today.

NVIDIA is expanding its supply chain to meet the demand for AI and has highlighted three elements of its new growth strategy: CPU, networking, and software and services. Its first data center CPU, Grace, is in full production and ramping up to become a multi-billion dollar product line. NVIDIA's networking and software and services divisions are also seeing significant growth. The company has announced an Ethernet for AI platform for enterprises and has partnerships with major companies to provide a full generative AI solution. NVIDIA is essentially an AI foundry, offering a range of products and services for building and deploying AI applications.

This summary was generated with AI and may contain some inaccuracies.