A：Tokens are the building blocks of AI, acting as a new kind of factory generator that turns data into knowledge, reason, and action. They enable the identification of patterns in complexity, help mirror cities to keep people safe, and facilitate robots to learn from and work alongside humans. Tokens also extend the reach of AI to places humans cannot, providing assistance and bridging the gap between hope and healing.

A：AI has become an integral part of Taiwan's rapidly growing economy, with the annual GDP projected to grow almost 10%. The ecosystem in Taiwan has expanded significantly, touching all aspects from the software stack and developer ecosystem to the supply chain and extending to data centers and end-users. The presence of a rich ecosystem has contributed to substantial business growth, with Nvidia seeing an increase in demand for AI computing power and a booming ecosystem that includes contributions from 70 different conferences simultaneously.

A：AI has significantly impacted the software development industry by introducing agentic AI, which has made software coding more accessible and efficient. With AI now being a profit generator and GDP creator, the demand for AI development has surged. This has led to a substantial increase in the number of commits in platforms like GitHub and a marked rise in productivity, with AI generating $9 trillion worth of output from $3 trillion of input. The industry has seen a net increase in the number of software engineers, as AI has created more jobs rather than reducing them.

A：The new computing pattern facilitated by AI is centered around agents, which are intelligent applications that can understand, observe, reason, act, and use tools. These agents are composed of large language models within a harness that manages memory and orchestrates the use of various tools. Unlike traditional applications, agents can engage with a vast array of computing resources, including CPUs, GPUs, and specialized processors, to execute complex tasks. This computing model is characterized by its disaggregated, distributed, and heterogeneous nature, which differs significantly from previous computing paradigms where resources were more centralized.

A：Nvidia's CUDA libraries are crucial to the new AI computing model as they provide the tools and algorithms for agents to use effectively. These libraries have been instrumental in solving some of the world's most complex problems across various fields of science and engineering. They will continue to be relevant as the computing pattern shifts to support more sophisticated use of tools by AI agents, and they will now include skills that teach the AI how to utilize them. The CUDA libraries' ability to interact with different computing resources, such as CPUs and GPUs, and orchestrate tasks, including memory management, makes them indispensable in the new heterogeneous computing landscape.

A：Vera Rubin is a comprehensive system from Nvidia that includes GPUs, Nvidia nvlink 72, CPU orchestration, storage systems, encryption for security, and the software stack called doka. It is designed to be secure and to enable the construction of AI factories.

A：Nvidia's goal with its AI infrastructure is to provide customers with a complete stack similar to GPUs and nvlink, which allows them to build amazing AI infrastructure.

A：The AI factories being built today are incredibly complex, involve multiple layers of computing hardware, and must be designed together. They are also built to be energy-efficient and flexible, able to operate with the grid and adjust power usage in real-time.

A：Nvidia Dsx is a blueprint and reference design for building and operating AI factories at maximum efficiency and profitability. It is used for designing, validating, provisioning, operating, monitoring, and remediating AI factories before they are even built.

A：Benefits of Nvidia Dsx for AI factories include a digital twin for planning and validation, more efficient use of power and cooling, automated operations, dynamic power allocation, grid coordination, and the potential to reduce power costs and increase revenue generation.

A：Nvidia's ecosystem is different from the computing ecosystem of the past in that it integrates Nvidia's computing layers, software, and computing stacks directly into its own platforms, unlike in the past where these were integrated into third-party platforms and libraries.

A：When building AI factories, economic factors to consider include the high cost of capital, the immense complexity of the infrastructure, and the revenue generation potential. Each AI factory can cost between 30 to 100 billion dollars and must be carefully designed to ensure profitability and efficient use of capital.

A：Nvidia offers several advantages to customers building AI factories, including an integrated stack of hardware, software, and libraries, rapid time to market, world-class throughput per watt, high reliability, and the ability to quickly adapt to new technological developments.

A：Compute power is important in AI factories because it directly translates into revenue generation. Higher compute power means more profitable AI operations and a greater ability to generate revenue from the data centers.

A：Nvidia's approach to AI infrastructure is unique due to its end-to-end integration, including designing, simulating, and building the entire system. This integrated approach ensures a high throughput per watt, fast time to market, and a long useful lifetime of the AI infrastructure.

A：Nvidia has created a supply chain for Vera Rubin that is twice as large as for Grace Blackwell, with an assembly time reduced from two hours to five minutes. This enables the company to meet the high demand for Vera Rubin and supports rapid deployment and scaling.

A：Key components of Vera Rubin include TSMC's 7nm chips, HBM4 memory from Micron and SK Hynix, and an 18 trillion-transistor MVL72 chip for prompt and context understanding, reasoning, and planning. The system also features a new modular compute train with streamlined PCB midplane design, hot-swappable MVL72 switch trans, and high-efficiency manifolds and liquid-cooled buzz bars.

A：Vera Rubin is described as an 'agent-centric system' because it is designed not just to run AI but also to run AI agents, making it the most advanced computer of its kind, tailored to the needs of AGI (Artificial General Intelligence).

A：Vera Rubin improves upon previous systems with its cableless design, streamlined architecture, and high-reliability features. It also integrates an advanced networking system with the world's first CPOE (Centralized Patch Panel Over Ethernet) switch, which is revolutionary for the AGI era.

A：The Vera CPU features extremely low latency, high single-threaded performance, and high bandwidth for data movement, all built with a new CPU architecture designed specifically for agents. It has the highest instructions per clock (IPC) in the world, offers top bandwidth per core, and overall high bandwidth around the chip, making it highly energy efficient for embedding in AI factories.

A：Nvidia Vera's CPU design supports the needs of AGI by focusing on instructions per clock (IPC), bandwidth per core, and total bandwidth around the chip, alongside energy efficiency. Its unique features include the highest IPC in the world, top bandwidth capabilities, and a new CPU architecture optimized for the AGI era.

A：Nvidia Vera Bluefield four Sts is going to be a major growth driver for Nvidia, powering context, memory, and AI storage. It is positioned as the CPU for the age of agents, with every data center, cloud, enterprise, and company that works with Nvidia on AI having already qualified Grace and Vera. Vera is expected to become the most optimized agentic CPU in the world.

A：Nvidia Vera has demonstrated significant performance improvements, such as running SQL three times faster than before and processing real-time stream data 6 times faster than previous technologies, thanks to high bandwidth and single-threaded instruction execution capabilities.

A：Nvidia Vera CPU signifies the opening of a new market, which is tailored for agents and agentic systems that have different properties from traditional CPUs. Vera is being adopted by Taiwan's ODMs and computer makers, and the market for such CPUs is anticipated to be larger than the previous markets because there will be many more agents than people, and these agents are very impatient.

A：The Nvidia Toolkit for agents includes models, harnesses, tools, and a runtime, representing the operating system of the modern enterprise. It comprises open models, such as the new Nemotha, and is designed to help companies build agents as a service or to operate agents. The toolkit is intended to allow for customization and improvement of the models and includes secure, open-source components that can be run in various environments.

A：Nvidia and Cadence are partnering to build chip design agents using AI to optimize the chip design process. The agents use models, harnesses, accelerated libraries, and tools to speed up chip design verification cycles by over 40 times, reduce the time needed from weeks to hours, and enable the design of AI factories that power the world's frontier AI models.

A：Nemotha is the next open model announced by Nvidia, designed to be used by developers to create their own agents. It is a smart model that comes with all the data used for training, allowing developers to modify and improve it. Nemotha is trained on a large suite of data and is intended to be a robust foundation for developers to build proprietary agents that incorporate their own expertise.

A：The new PC reinvention by Microsoft and Nvidia includes a personal AI that operates on the user's PC, providing assistance with tasks like helping in research, reading files, and assisting with work. The new operating system is described as the old operating system plus large language models. The new PC is also characterized by the presence of an autonomous agent that understands the user, responds to prompts, and is integrated with computer vision and AI capabilities, offering modernized input and output functionalities.

A：Innovations included in the RTX Spark consist of an RTX GPU with 6144 CUDA cores and 1 petaflop of AI performance, a custom 20 core Grace CPU, 128 GB of unified memory, TSMC's 3 nm process technology with 70 billion transistors, and a Windows platform for agents. This GPU is designed to support a broad range of applications from digital biology to astrophysics, Nvidia's entire software stack, and comes with a Windows platform optimized for running AI agents on PCs.

A：The significance of the RTX Spark lies in its ability to run a broad range of applications from various fields such as digital biology, seismic processing, astrophysics, AI, and computer graphics due to its powerful Nvidia GPUs and comprehensive software stack. It supports AI integration, including local AI models and cloud-based models like nemotha 3 ultra or Claude. The device is designed to be agent-friendly, offering capabilities that excite developers and create an ecosystem of powerful and intelligent laptops that are ready for the next era of computing.

A：An AI agent integrated with a PC assists in tasks like house design by using the tools on a laptop to run AI models such as Claude from the cloud, which can help users by modeling the site, shaping terrain, and suggesting building forms optimized for cost, comfort, and quality. The agent can propose and refine designs, place structural elements automatically, and even export models from Rhino into Blender for further fine-tuning. This integration simplifies a complex workflow and allows for faster and more intuitive design processes.

A：The new line of Windows machines represents a complete reinvention of PCs with AI integration and performance. These machines are designed to be 100% Windows compatible, 100% CUDA compatible, and to support Nvidia AI Tensor Core. They signify the first major re-engineering of PCs in 40 years and include a range of desktop, laptop, and workstation models that are powerful enough to run AI models, create agents, and support developers working with large language models. They are positioned to facilitate the creation of AI systems across various sectors like self-driving cars, satellites, and even home-based supercomputers.

A：The future PC is expected to evolve significantly from current perceptions, possibly becoming indistinguishable from a home theater, a tool for launching applications, or a device for clicking and typing. AI agents could play a central role in this evolution by running continuously in the background, assisting users with various tasks such as booking travel, managing household functions, and interacting with other devices in the user's home. As AI agents become smarter and more prevalent, they could become as integral to daily life as household devices such as stereos and game consoles, shaping the way people think of PCs in the coming decade.

A：Cosmos is an open frontier, multimodal model for physical AI, built on a new mixture of transformers, architectures, and various forms of data such as pixels, action, sound, and language. It is designed to observe and understand the physical world by generating physics-accurate synthetic video and watching the world, describing scenes, and flagging important events. Cosmos functions as a world model that closes the loop for policy, training, and evaluation, predicting the future frame by frame and can generate actions for robots.

A：Nvidia's open model for self-driving cars is called Alpha Mayo 2. The company is working with car companies across the world, representing about 80% of the world's cars, to integrate this model. Nvidia is also connected with approximately 97% of the world's mobility services through the Nvidia Hyperion system, which will allow deployment of Alpha Mayo on the Hyperion runtime with the Haloes operating system, connecting to these services globally.

A：Alpa Mayo is described as the world's first reasoning autonomous vehicle. It is unique because it talks to itself continuously, which is essentially thinking and making decisions. The technology applied to Alpa Mayo also has potential applications for humanoid robots.

A：The Nvidia Isaac Group is an open development platform for humanoid robots, including an integrated software stack for simulation, training, and data generation. It features a fully integrated humanoid robot with 25 degrees of freedom in each hand and 31 degrees of freedom on the robot itself, designed to operate on Nvidia's Thor system. The platform includes the entire software stack and is intended to simplify the creation of humanoid robots for researchers and developers.

A：According to the speaker, the computer industry has changed dramatically in the last six months due to the realization of agents and their convergence with the latest AI models. This has made it possible for AI to now perform useful work. The computing pattern is now centered around the agent, which includes a model, a harness that utilizes tools and skills, and a runtime that varies depending on the environment such as cloud, on-prem, or a robot. Nvidia is now considered an infrastructure company offering tools to maximize revenue and profits for businesses adopting this new computing pattern.

A：Nvidia has introduced a new computing pattern with agents, which will result in the development of CPUs tailored for agents, not just people. This 'agent-centric' computing pattern is expected to have its own set of requirements and will be implemented in various devices including PCs, robots, satellites, and base stations. This pattern is anticipated to significantly alter the way personal computers are thought of in the future.