The paragraph is an introduction to NVIDIA Corporation's Fourth Quarter Earnings Call for fiscal year 2025. Christa, the conference operator, begins by explaining the call format, including the question and answer session. Stewart Stecker then welcomes attendees and mentions that NVIDIA executives Jensen Huang and Colette Kress are present. The call is webcast live and will be available for replay until the first quarter of fiscal 2026 earnings results are discussed. Attendees are reminded that the call's content is proprietary and cannot be reproduced without permission. The call includes forward-looking statements that come with risks and uncertainties, and attendees are directed to official filings for more information. Statements are current as of February 26, 2025, and there is no obligation to update them unless legally required. Lastly, non-GAAP financial measures will be discussed.

In Q4, NVIDIA reported record revenue of $39.3 billion, surpassing expectations and contributing to a fiscal 2025 revenue of $130.5 billion, marking a 114% increase from the previous year. The data center segment saw significant growth, with fiscal 2025 revenue at $115.2 billion, more than doubling from the prior year. The Blackwell product ramped up rapidly, delivering $11 billion in Q4, and became the fastest product launch in NVIDIA's history due to high demand and quick supply expansion. The data center compute revenue increased 18% sequentially and more than doubled year on year as customers scaled infrastructure for advanced AI models. NVIDIA's infrastructure and software are in demand for AI model customization, with companies like Hugging Face hosting numerous derivatives from foundational models.

The paragraph discusses the massive demand for compute in post-training and model customization, especially for reasoning AI, which requires significantly more resources than one-shot inferences. Blackwell is highlighted for its ability to enhance reasoning AI models, offering substantial improvements in token throughput and cost efficiency compared to current technologies. NVIDIA's full-stack inference platform, including tools like TensorRT and Triton inference server, is helping companies across industries significantly boost performance and reduce costs. The Blackwell architecture addresses the entire AI market, enhancing AI models and supporting over 4,400 applications, ensuring investments remain relevant in evolving markets. Additionally, there has been a 200% reduction in inference costs over the past two years, driven by NVIDIA's innovation.

In the article, NVIDIA Corporation reports significant growth and advancements in its data center and AI capabilities. The company's full-stack optimizations have improved customer economics, with large cloud service providers (CSPs) doubling their data center revenue year over year by deploying NVIDIA's GV200 systems globally to meet increasing AI demand. Coreweave launched a large-scale GV200-based instance, and consumer internet revenue grew threefold due to expanding AI applications. Additionally, companies like XAI and Meta are leveraging NVIDIA's technology for AI model training and advertising, respectively. Enterprise revenue nearly doubled due to rising demand for model fine-tuning and AI workflows. NVIDIA also introduced the Llama Numitron model family to support AI agent deployment in various applications.

Leading AI platforms like SAP and ServiceNow are utilizing new models alongside healthcare leaders like IQVIA, Lumenon, and the Mayo Clinic, all using NVIDIA's AI technology to enhance drug discovery and genomic research. NVIDIA's infrastructure is being adopted for robotics and autonomous vehicles, with a significant portion of its automotive revenue, expected to reach $5 billion this year. Hyundai Motor Group and Uber are advancing their AV and robotics projects with NVIDIA's technology. NVIDIA's new Cosmos World platform aims to revolutionize physical AI, just as language models have for language AI. Global demand for compute infrastructure is increasing, with significant AI investments coming from the US, France, and the EU.

The paragraph highlights NVIDIA Corporation's performance and strategic actions in various sectors. Data center sales in China have remained low due to export controls, while networking attached to GPU compute systems has been robust, driven by transitions to newer technologies like NVLink 72 and Spectrum X, which has seen significant success and adoption by companies like Microsoft Azure and Cisco. Gaming revenue decreased both sequentially and year-on-year, but annual revenue experienced growth of 9%. Despite supply constraints affecting Q4 shipments, strong demand is expected to drive sequential growth in Q1, supported by the release of the new GeForce RTX 50 series GPUs designed for gamers, creators, and developers.

The paragraph discusses NVIDIA's advancements in GPU technology with the introduction of the Blackwell architecture, offering significant performance improvements and enhanced AI-driven rendering capabilities. It highlights the launch of GeForce Blackwell laptop GPUs, which are set to improve battery life with Max-Q technology, available from March. NVIDIA's professional visualization revenue increased due to demand from industries like automotive and healthcare. The company's automotive sector also showed strong growth, driven by the adoption of NVIDIA's technologies in autonomous vehicles and partnerships with major automakers like Toyota. The paragraph also mentions the deployment of driverless trucks powered by NVIDIA's Drive platform.

The paragraph discusses NVIDIA Corporation's recent developments and financial performance. Their autonomous vehicle platform, DRIVE Hyperion, has passed significant industry safety assessments, marking it as the first to receive comprehensive third-party evaluations. The company also reports GAAP gross margins of 73% and non-GAAP gross margins of 73.5%, with anticipated fluctuations as they ramp up their Blackwell AI infrastructure. Blackwell's introduction has led to increased operating expenses, and they have returned $8.1 billion to shareholders. For the first quarter, NVIDIA expects total revenue of $43 billion and predicts sequential growth in both data center and gaming sectors, with GAAP and non-GAAP gross margins expected to be around 70.6% and 71%, respectively.

The paragraph outlines financial expectations and upcoming events for a company, including projected GAAP and non-GAAP operating expenses and tax rates for fiscal year 2026. It announces participation in several conferences in March, including the TD Cowen Healthcare Conference, Morgan Stanley Technology, Media, and Telecom Conference, and the annual GTC conference, where Jensen will deliver a keynote. The earnings call for the first quarter of fiscal 2026 is set for May 28, 2025. During a Q&A, C.J. Muse from Cantor Fitzgerald asks Jensen about the impact of reinforcement learning and test time compute on the future of inference-dedicated clusters.

Jensen Huang discusses the different scaling laws affecting NVIDIA Corporation and its customers. He highlights three areas: pretrained scaling laws, post-training scaling laws using reinforcement learning, and test-time compute or reasoning, which involves significant computational demands. Huang emphasizes that future models will require exponentially more compute power. He notes the complexity of designing data center architectures to accommodate these evolving needs, highlighting why NVIDIA's adaptable architecture is widely favored.

The paragraph discusses the advancements and capabilities of Blackwell, a system designed for reasoning AI models. It highlights the superior performance and speed in training and inference tasks, with a focus on flexibility and easy configuration for various data center needs. The conversation shifts to the GV200 and NVLink 72 platforms, with Jensen Huang expressing increased enthusiasm due to successful scaling and high demand. The company has ramped up production significantly, delivering $11 billion in revenues last quarter, and plans to continue scaling to meet customer demand for Blackwell systems.

The paragraph discusses the widespread adoption of Grace Blackwell Systems, with several companies, including Coreweave, Microsoft, and OpenAI, successfully implementing them. Colette Kress addresses a question about gross margins, stating that during the Blackwell ramp, margins will be in the low seventies, with efforts focused on expediting manufacturing to meet customer demand. Once Blackwell is fully ramped, margins are expected to improve to the mid-seventies later in the year. The systems are complex and customizable, offering opportunities for further margin improvement. The immediate focus, however, is on delivering the products to customers quickly.

In the paragraph, Jensen Huang discusses the future of data centers and AI, highlighting the shift towards machine learning-based software and accelerated computing architectures. He mentions the strong demand for computing infrastructure driven by innovative start-ups and outlines short, mid, and long-term indicators of growth, such as capital investment in data centers and the evolution of software from CPU-based to AI-driven systems. Huang anticipates future advancements in AI, including agentic AI for enterprises, physical AI for robotics, and Sovereign AI for regional ecosystems, noting that these developments are still in their early stages but show great potential.

The paragraph discusses the launch and ramp-up of the next generation Blackwell Ultra set for the second half of the year. Matt Ramsay queries about managing demand and supply chain dynamics given the overlap with the current Blackwell's ramp-up. Jensen Huang confirms they are on track for Blackwell Ultra's launch, noting the recovery from delays with the initial Blackwell. He explains that the system architecture remains the same between Blackwell and Blackwell Ultra, simplifying the transition compared to the previous challenging transition from Hopper to Blackwell. This continuity should facilitate a smoother integration for customers and the supply chain.

The paragraph discusses NVIDIA's approach to custom ASICs and merchant GPUs, emphasizing their unique offerings. NVIDIA's architecture is described as general-purpose and versatile, capable of handling various models and applications due to its comprehensive software stack and rich ecosystem. This makes NVIDIA a preferred choice for innovative algorithms. Unlike custom ASICs, NVIDIA's solutions provide an end-to-end framework from data processing and training to inference, positioning them as a flexible and holistic option for diverse computing needs.

The paragraph discusses the advantages of the company's adaptable architecture, which can operate across various cloud environments and on-premise, making it appealing for startups. It highlights the architecture's superior performance, offering 2x to 8x improvements in performance per watt, translating to increased revenues for data centers due to faster token throughput. This is particularly valuable for AI-driven applications where rapid return on investment is crucial. The paragraph also points out the complexity of the software stack required for such an architecture, noting that building and deploying chips is challenging. Despite many chips being designed, not all get deployed, emphasizing the importance of a robust ecosystem.

The paragraph discusses the business decision to deploy a new, advanced engine or processor in an AI factory, highlighting the technological superiority and rapid deployment capabilities of their product. Ben Reitzes from Melius Research asks about geographic demand shifts, particularly the increase in the US market despite potential regulations in other regions. Jensen Huang responds by noting that while China's involvement has decreased due to export controls, AI's widespread use in various services has led to its mainstream adoption.

AI is now a core component across various sectors, from education and healthcare to fintech and climate tech, marking its mainstream integration into software and services. This era represents a shift from decades of traditional computing to a future dominated by AI and machine learning. This transition is just beginning, modernizing computers for AI integration, and is poised to impact a larger portion of the global economy than any previous technology or software tool. This perspective requires reevaluating how we measure growth and scale.

The paragraph is an excerpt from a conference call involving several individuals, discussing the growth of enterprise and hyperscale purchasing within the data center market. Christa introduces participants such as Aaron Rakers, Mark Lipacis, and Marshall Pappas, who pose questions to Colette Kress and Jensen Huang. Colette confirms that enterprise data center growth doubled year-on-year during the January quarter, similar to hyperscalers. She notes that both enterprises and CSPs (Cloud Service Providers) are integral, working on aspects like large language models and inference. Jensen clarifies that CSPs constitute about half of their business, with a split between internal and external consumption, influencing their service development.

The paragraph discusses the optimization of workloads for companies with extensive NVIDIA infrastructure, highlighting the versatility and cost-effectiveness of using this technology for AI, video, and data processing. It emphasizes the potential growth of enterprises, particularly in industrial sectors not fully served by the computer industry. Using a car company as an example, the paragraph illustrates how both employees and manufactured cars benefit from AI: employees use agentic AI for productivity, while cars require AI for training and data collection. The future envisions cars as robotic systems with AI factories alongside traditional manufacturing, indicating the involvement of multiple computers in this ecosystem.

The paragraph discusses the emergence of "physical AI," a new type of artificial intelligence required for understanding real-world effects like friction and inertia, beyond just language. This concept is linked to the development of agentic AI, predicted to revolutionize company operations. The paragraph further mentions the sequential progression from agentic AI to physical AI, and then to robotic systems, indicating these advancements are in their early stages. It highlights the potential impact on industries and global GDP. The paragraph transitions into a Q&A segment, where Aaron Rakers from Wells Fargo inquires about the infrastructure and replacement cycle in relation to the advancements in AI technology. Jensen Huang responds by noting that older AI technologies, like Voltas, Pascals, and Amperes, are still in use.

The paragraph discusses the use of CUDA's programmability for data processing and curation, specifically in improving AI models for certain tasks like analyzing circumstances presented to vision language models. By generating multiple examples using domain randomization, these models can be better trained. The process involves using different CUDA-compatible architectures, such as Amperes for initial processing and Hoppers for training. The conversation then shifts to a question from Atif Malik regarding gross margins, where Colette is asked about expectations for margin improvements throughout the fiscal year, considering various factors like NVLink 72 and Ethernet mix. Malik notes that margins would need significant improvement in the second half of the year to reach the projected range.

Colette Kress discusses the complexity and opportunities in improving gross margins for semiconductor products, particularly focusing on configurations within the Blackwell system. She mentions the uncertainty surrounding the impact of tariffs, which depends on future U.S. government actions. The discussion concludes with Jensen Huang expressing gratitude and highlighting the growing demand for Blackwell and the evolution of AI from perception to reasoning. He notes that reasoning models like OpenAI's Grok 3 and DeepSeq R1 require significantly more computational power, with the latter generating global excitement due to its innovation.

The paragraph discusses the advancements and future of AI, emphasizing the significance of open-sourcing a reasoning AI model known as R1, which has influenced AI developers to enhance performance using new scaling laws. These include post-training scaling with reinforcement learning and inference time scaling, which require significantly more compute resources. It introduces "Blackwell," a platform designed for efficient AI model processing, offering substantial improvements over previous technologies. Blackwell is in full production and represents a significant leap in AI infrastructure. The text anticipates significant growth in AI sectors by 2025, predicting data centers will focus heavily on AI capabilities and suggesting a future where every company maintains AI infrastructure.

The paragraph invites readers to attend the upcoming GTC event, where discussions will cover topics like Blackwell Ultra, Rubin, and other new advancements in computing networking, reasoning AI, and physical AI products. The paragraph ends with a closing remark from Christa, indicating the conclusion of a conference call.

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