CES showcases the intersection of technology and humanity, transforming challenges into opportunities for a more connected, sustainable, and dynamic future through groundbreaking innovations.

The CEO of the Consumer Technology Association introduces NVIDIA and its CEO, praising their significant contributions to global innovation across various industries, highlighting their pioneering work in AI and accelerated computing.

Intelligence is redefined through a novel factory concept, generating tokens as the foundational elements for artificial intelligence, transforming how AI is constructed and understood.

Tokens unlock a world of endless opportunities, transforming various aspects of life from knowledge and art to safety and healthcare, demonstrating their pivotal role in advancing human experiences and technology.

Numbers are highlighted as essential tools that bring meaning to our world, enabling us to comprehend our surroundings, forecast potential dangers, and discover treatments for internal threats.

Tokens are portrayed as tools that revive lost capabilities and facilitate advancement, enabling collective leaps forward after recovering what was once lost.

The Nvidia founder and CEO appears at CES in Las Vegas, sporting an unexpected jacket choice, and encourages the audience to embrace the change.

Starting in 1993 with NV1, NVIDIA embarked on a journey to enhance computer capabilities, introducing the programmable GPU in 1999 and CUDA six years later, significantly advancing computer graphics and algorithmic processing.

The dialogue traces the significant advancements in AI from initial challenges in understanding concepts like Cuba to the transformative impact of AI, particularly through the development of Transformers in 2012, which revolutionized computing by enabling machine learning on GPUs. This evolution has led to the ability to understand and generate various modalities of information, changing every layer of the technology stack and fundamentally altering how applications are built and computing is performed.

The speaker discusses the breakthrough in ray tracing technology, enabled by AI, which allows for the efficient rendering of high-quality graphics by computing fewer pixels and using AI to predict the rest, leading to the announcement of the next-generation RTX Blackwell family.

NVIDIA introduces the GeForce RTX 50 series, featuring the Blackwell Architecture with 92 billion transistors and significant advancements in AI and ray tracing capabilities, promising unparalleled performance and efficiency in both desktop and laptop GPUs.

The industry is rapidly scaling artificial intelligence, driven by the scaling law which states that more data, larger models, and increased compute lead to more capable AI. With the internet doubling its data production annually, AI can harness this vast, multimodal data for foundational knowledge. Two additional scaling laws have emerged, reflecting the intuitive growth in AI capabilities.

The dialogue discusses the evolution of AI through scaling laws, emphasizing post-training and test time scaling techniques that enhance AI capabilities. It highlights the significant computational demands of these advancements and the pivotal role of advanced computing hardware, exemplified by the Blackwell chip, in driving AI innovation and addressing complex problems.

The discussion highlights the development and capabilities of the MV Link system and Blackwell GPUs, emphasizing their role in generating AI tokens for applications like ChatGPT. The system, comprising 600,000 parts and 1.5 tons, offers a four-fold improvement in performance per watt over the previous generation, significantly enhancing computational efficiency and capacity in data centers. Notably, a single Blackwell system is equivalent to the world's largest supercomputer, achieving 1.4 exaflops of AI floating point performance, and processing data at a rate comparable to global internet traffic.

The necessity for massive computational resources is discussed in the context of training and operating increasingly complex AI models, emphasizing the need for higher token generation rates and reduced costs to maintain high-quality, affordable AI services.

The dialogue highlights NVIDIA's strategy to facilitate the development of agentic AI through collaboration with software developers and the IT ecosystem. Key initiatives include Nvidia Nims, AI microservices optimized for various tasks; Nvidia Nemo, a system for onboarding, training, and evaluating AI agents; and a suite of open models based on the Lalama framework, tailored for enterprise use. These efforts aim to significantly enhance AI's capabilities in understanding, interacting with users, and performing complex tasks across industries.

Nvidia's AI technologies are transforming various sectors including industrial AI, software coding, and knowledge work, through partnerships with companies like ServiceNow, SAP, Siemens, and Cadence. The focus is on developing AI agents as digital workforces for tasks such as document ingestion, software security, drug discovery, and video analysis, aiming to enhance productivity and innovation. Additionally, the vision includes integrating AI directly into PCs through generative APIs, marking a shift from traditional computing models to AI-assisted environments.

The integration of Windows WSL 2 and Nvidia Nimo aims to transform Windows PCs into world-class AI platforms, enabling developers to leverage AI models for various applications including image generation guided by 3D assets.

NVIDIA introduces Cosmos, a world foundation model designed to advance physical AI by understanding the physical world, including dynamics, spatial relationships, and cause and effect, using auto-regressive and diffusion-based models.

NVIDIA Cosmos, the world's first world foundation model, trained on 20 million hours of video focusing on physical dynamics, enables AI to understand the physical world, facilitating synthetic data generation for robotics and enhancing large language models through accurate video captioning.

The Cosmos platform, now open-licensed and available on GitHub, features an autoregressive model for real-time applications, a diffusion model for high-quality image generation, an advanced tokenizer, and an AI-accelerated data pipeline designed for extensive data processing.

The integration of Cosmos, an open-world foundation model, with Omniverse, a physics-based simulation system, creates a physically grounded multiverse generator, mirroring the success of Llama 3 in enterprise AI. This combination ensures AI generation is based on ground truth, enhancing robotics and industrial AI capabilities.

NVIDIA presents a three-computer system strategy for robotics and industrial applications, comprising an AI training computer (dgx), an AI deployment computer (agx), and a digital twin for simulation and refinement. This approach targets the software-defined future of the $50 trillion manufacturing industry, focusing on automation and robotics integration. Partnerships with leading companies like Keyon and Accenture aim to create innovative solutions for warehouse automation and digital manufacturing.

Kon, Accenture, and Nvidia are utilizing AI and digital twin technology to optimize warehouse logistics and autonomous vehicle operations, addressing challenges like variable demand and workforce availability through simulation and predictive analytics.

A significant advancement in the autonomous vehicle industry is announced with the introduction of 'Thor', the next generation processor designed to power self-driving trucks and cars, highlighting the industry's rapid growth and potential to become a multi-trillion dollar robotics sector.

The robotics computer, Thor, boasts 20 times the processing capability of the previous generation, enabling it to handle massive sensor data for autonomous vehicles and robots. Additionally, the system is certified to the highest standard of functional safety for automobiles, marking a significant achievement in safety and AI technology.

The autonomous vehicle industry utilizes Nvidia's DGX, Omniverse, and Drive AGX for AI model training, simulation, and generating synthetic data. Omniverse constructs digital twins from AI and sensor logs, creating photorealistic 4D simulations for enhanced training data, while Cosmos scales synthetic datasets to billions of effective miles, improving autonomous driving safety and capability.

The use of synthetic data generation, based on physically grounded simulations, promises to exponentially increase the training data for AI in autonomous vehicles, accelerating industry progress.

The breakthroughs in enabling technologies are poised to rapidly advance general robotics, making possible the deployment of robots in existing human environments without requiring special adaptations. Key areas of focus include agentic robots, self-driving cars, and humanoid robots. A major challenge is collecting sufficient demonstration data for humanoid robots, which is labor-intensive. Nvidia Isaac Groot addresses this by offering a platform for synthetic motion generation, enabling the creation of massive datasets from a small number of human demonstrations through simulation workflows for imitation learning. This process, involving teleoperation, data multiplication, and domain randomization, accelerates the development of general-purpose robot models, heralding the robotics era powered by AI.

A project initiated a decade ago, named Project DGX, transformed AI development by providing researchers and startups with an out-of-the-box AI supercomputer. Initially designed to simplify access to AI computing power, DGX 1 has evolved to make AI integral to everyday computing, catering to software engineers, artists, and all computer users.

NVIDIA has developed a new AI supercomputer, codenamed Project Digits, running the entire NVIDIA AI stack and utilizing the GB110 chip, the smallest Grace Hopper variant. This supercomputer can function as a workstation or cloud-based system, featuring a CPU collaboratively built with Mediatek and connected to a Blackwell GPU via chip-to-chip MvLink. Expected to be available around May, it showcases significant advancements in AI computing capabilities.

A breakthrough in technology is announced, featuring new supercomputers and the world's first physical AI foundation model, aimed at advancing industries like robotics and self-driving cars.