Video Summary: NVIDIA CEO Jensen Huang Keynote at CES 2025 by Jensen Huang

Recommendation

Since its inception in 1993, Nvidia has fueled a revolution in computing and artificial intelligence (AI) by introducing pioneering technologies, including the NV1 graphics accelerator and the programmable GPU. CEO Jensen Huang’s AI-enabled keynote at CES 2025 showcases Nvidia’s latest leap: the GeForce RTX 50 series of Blackwell GPUs, delivering double the power of their predecessors at a fraction of the cost.

The keynote also highlights Nvidia’s transformative Blackwell platform for AI scaling, Cosmos for physical AI, and Isaac GROOT for robotics, as well as Nvidia’s plans for AI PCs, autonomous vehicles, and enterprise AI. The company’s innovations promise to revolutionize industries and workflows in a fully AI-integrated future that Huang sees as just around the corner.

 

Take-Aways

• Since 1993, Nvidia has helped bring about a revolution in computing and artificial intelligence (AI).
• Nvidia’s pioneering GeForce GPUs made the growth of AI possible, and now AI is transforming GeForce.
• The new Blackwell platform offers higher performance at lower cost.
• Nvidia offers systems, blueprints, and models to support the implementation of agentic AI.
• The company will soon offer AI-powered Windows PCs that put sophisticated AI tools in every user’s hands.
• Nvidia’s Cosmos and Omniverse platforms are set to revolutionize physical AI.
• The company is accelerating the development of autonomous vehicles (AVs) by offering technology for training, data generation, and vehicle control.
• Nvidia’s Isaac GROOT platform generates massive datasets to support the training of humanoid robots.
• Nvidia’s new Project DIGITS packs a supercomputer in a small box.

Summary

Since 1993, Nvidia has helped bring about a revolution in computing and artificial intelligence (AI).

Nvidia’s pioneering efforts in artificial intelligence began in 1993 with the development of the NV1 graphics accelerator, which enabled gaming on PCs. The groundbreaking programmable graphical processing unit (GPU) followed in 1999. These innovations laid the foundation for modern computer graphics and GPU-powered computing. In 2006, Nvidia introduced compute unified device architecture (CUDA), enabling the programmability of GPUs and revolutionizing the processing of algorithms. A breakthrough for AI came in 2012 when AlexNet used CUDA for processing.

“Every single layer of the technology stack has been completely changed — an incredible transformation in just 12 years.”

In 2018, Google’s Transformer model, BERT, redefined AI and fundamentally changed computing by advancing machine learning. With the advent of machine learning-powered models every layer of the technology stack has transformed. AI now extends across multiple modalities — images, text, sounds, and even amino acids and physics — and it can generate new knowledge across these domains. Today, AI is embedded at the core of applications across industries, pushing the boundaries of computing and creating endless possibilities for future advancements. In recent years, the company’s focus has shifted from perception AI — such as speech recognition and medical imaging — towards generative and agentic AI. The next step — physical AI — waits just around the corner.

Nvidia’s pioneering GeForce GPUs made the growth of AI possible, and now AI is transforming GeForce.

GPUs now incorporate the use of programmable shading and ray-tracing acceleration to create high-quality images. And ultra-high performance rendering is achieved through deep learning super sampling (DLSS), where AI generates additional pixels and frames through inference. Innovations like neurotexture compression and neuromaterial shading enable extraordinary graphics.

These processes require a new level of compute power, and Nvidia’s GeForce RTX 50 series of Blackwell GPUs represents a transformative leap. These units feature 92 billion transistors, 4 petaflops of AI power, G7 memory, and 1.8 TB/second memory bandwidth, enabling them to deliver unprecedented rendering quality. The RTX 5070 offers the same performance as its predecessor, the 4090, at a fraction of the cost — Nvidia is offering it for $549. The top-of-line RTX 5090 doubles the 4090’s performance, at a price tag of $1,999.

“The future of computer graphics is neural rendering — the fusion of artificial intelligence and computer graphics.”

AI-driven computing efficiency and the resulting energy efficiency make it possible for these high-performance GPUs to fit into an ultra-thin laptop.

The new Blackwell platform offers higher performance at lower cost.

Nvidia has introduced the new Blackwell platform to support more powerful data centers and advanced AI development. According to the traditional, empirically established pre-training scaling law, more training data, larger models, and additional compute power will translate into correspondingly greater intelligence. Now, two new scaling laws have emerged. According to the post-training scaling law, an AI system can improve itself through techniques such as reinforcement learning, human feedback, and the use of synthetic data. And according to the test-time scaling law, an AI system can decide for itself, dynamically, the amount of computation to use as it works to produce an answer and can reason about how best to approach a problem.

“We need the token generation rates to go way up, and we also have to drive the cost way down simultaneously, so the quality of service can be extraordinary, the cost to customers can continue to be low, and AI will continue to scale.”

The techniques involved in these new types of scaling require tremendous amounts of computing power, and the Blackwell chips help meet the demand for it. A wide variety of Blackwell systems are now being assembled in 45 factories worldwide. The flagship GB200 NVL72 GPU integrates 72 Blackwell GPUs into a single system that acts as a unified “chip.” The system, weighing 1.5 tons, incorporates 130 trillion transistors and 2,592 Grace CPU cores, and operates at 120 kW. It delivers 1.4 exaflops of AI floating-point performance and 14 terabytes of memory with 1.2 petabytes/second bandwidth — equivalent to processing the entire internet’s traffic in real time. Compared to previous models, the GB200 improves performance per dollar by 3X and performance per watt by 4X.

Nvidia offers systems, blueprints, and models to support the implementation of agentic AI.

In agentic AI, systems composed of multiple models can interact with users and with one another, reason, retrieve information, and generate responses. Unlike traditional AI, which offers simple Q&A interactions, agentic AI can break down complex problems, call on multiple models, and utilize tools such as calculators or internet resources to deliver solutions. AI agents can perceive, reason, plan, and act.

Nvidia offers three key frameworks for developing and deploying agentic AI:

• Microservices — Nvidia NIM consists of prepackaged AI microservices, supporting models for vision, language understanding, speech, animation, digital biology, and emerging areas like physical AI. They’re already available in every cloud.
• A “digital HR” framework — Nvidia’s NeMo framework functions as a “digital HR system” for onboarding and training AI agents, allowing companies to tailor AI agents to specific workflows, language, and business processes. NeMo enables feedback, evaluation, and guardrails for AI behavior. Nvidia also offers open-source blueprints for specific ecosystems and applications, allowing businesses to customize AI agents to their needs.
• Language foundation models — Nvidia now offers a suite of Llama Nemotron language foundation models, consisting of fine-tuned versions of Meta’s Llama 3.1 models, for enterprise use in tasks such as chat, instruction, and retrieval.

“In a lot of ways, the IT department of every company is going to be the HR department of AI agents in the future.”

AI agent use cases include, for example, knowledge assistants, climate modeling, software security, drug discovery, and industrial analytics. Nvidia is partnering with a host of companies, such as ServiceNow, SAP, Siemens, Cadence, Perplexity, and Synopsys, to develop tools for these use cases. The world has a billion knowledge workers and over 30 million software engineers, and agentic AI is poised to revolutionize their productivity in tasks like coding.

The company will soon offer AI-powered Windows PCs that put sophisticated AI tools in every user’s hands.

Nvidia envisions a future where PCs become AI-powered personal assistants, revolutionizing workflows and creativity. Key to this vision is Windows Subsystem for Linux (WSL2), a dual-operating system environment optimized for CUDA and cloud-native applications. This makes WSL2 a perfect foundation for running Nvidia’s AI tools, such as NIM, NeMo, and generative AI blueprints.

“Our focus is to turn Windows WSL2 Windows PC into a target, first-class platform that we will support and maintain for as long as we shall live.”

Nvidia’s models and systems will enable PCs to handle a variety of tasks, such as creating visuals based on simple 3D layouts. Nvidia is collaborating with leading PC manufacturers to ensure that hundreds of millions of Windows PCs are AI-ready.

Nvidia’s Cosmos and Omniverse platforms are set to revolutionize physical AI.

Nvidia Cosmos is a platform for building the “first world foundation model” — akin to a language model — designed for physical AI. Trained on 20 million hours of video, it’s designed to help AI understand and simulate the physical world, allowing the AI to interpret physical dynamics like gravity, friction, and object permanence, as well as to understand spatial relationships and cause-and-effect.

Cosmos incorporates autoregressive and diffusion-based world foundation models, advanced video tokenizers, and AI-accelerated pipelines to handle large datasets. It works by ingesting prompts such as text, images, and videos to generate videos that represent virtual world states. By using Omniverse, another Nvidia tool, developers can create physically and spatially accurate simulations and then output the results to Cosmos to generate realistic synthetic data. Cosmos is open source and available on GitHub.

“Cosmos world foundation model, being open, we really hope will do for the world of robotics and industrial AI what Llama 3 has done for enterprise AI.”

Nvidia envisions every factory and warehouse having a digital twin, where simulated operations allow AI systems to learn and practice the management of high-performance logistics. Already, KION, a warehouse automation provider, and Accenture, which has a leading digital manufacturing practice, are working with Nvidia to optimize warehouse logistics. KION is using an Omniverse blueprint called Mega to simulate warehouse operations in real time, allowing robot brains to reason about the contents of the digital twin setting before implementing their actions in the real warehouse.

The implementation of robotic systems will depend on the so-called three-computer solution: one computer for AI training (called the DGX computer), one for deployment (the AGX computer), and a third dedicated to refinement, where the AI can practice in a digital twin of the real setting, created through use of Cosmos and Omniverse.

The company is accelerating the development of autonomous vehicles (AVs) by offering technology for training, data generation, and vehicle control.

Nvidia is contributing to the autonomous vehicle industry by providing all three components in the three-computer solution: DGX training systems; omniverse simulation and synthetic data generation systems; and Drive AGX, the in-car supercomputer. The company is collaborating with major car manufacturers including Tesla, BYD, JLR, Mercedes, Lucid, Rivian, and Aurora, and is announcing a new partnership with Toyota to build Toyota’s next generation of AVs. Nvidia’s business in AVs is already worth $4 billion.

The company is also announcing Thor, its next-generation robotics processor, destined for AV use. Thor, now in full production, is based on the Hyperion 9 platform and offers 20X the processing power of its predecessor, Orin. Nvidia is also announcing that its Drive OS AI computer has been certified ASIL D, the highest functional safety standard for automobiles.

“I predict that (AV) will likely be the first multitrillion-dollar robotics industry.”

Omniverse and Cosmos are accelerating the drive toward AVs by generating massive amounts of synthetic data for training the vehicles’ AI systems. Omniverse and Cosmos tools create realistic 3D environments and high-fidelity simulations, enabling Nvidia’s engineers to scale training data from thousands of recorded actual drives to billions of varied training miles.

Nvidia’s Isaac GROOT platform generates massive datasets to support the training of humanoid robots.

Training robots, especially human-like ones, requires collecting large amounts of data to support imitation — a costly and labor-intensive task, because it depends on gathering data from human demonstrations. To solve this problem, Nvidia’s Isaac GROOT platform provides a complete technology stack that includes robot foundation models, data pipelines, simulation frameworks, and the Thor robotics computer. The platform incorporates a blueprint for synthetic motion generation that enables developers to create massive datasets from just a few human demonstrations.

“The ChatGPT moment for general robotics is just around the corner.”

The workflow starts with the GROOT-Teleop tool, where skilled human workers control digital twins of robots via Apple Vision Pro, capturing motion trajectories without the risk of physical damage to the real robot. GROOT-Mimic and GROOT-Gen, using the multiverse simulation engine of Cosmos and Omniverse, then produce an exponentially larger dataset from the collected data for use in training robot policy. Isaac Sim, another part of the platform, provides a virtual testing environment that reduces the risks and costs associated with real-world experimentation.

Nvidia’s new Project DIGITS packs a supercomputer in a small box.

In 2016, Nvidia released the DGX1, the world’s first “out of the box” supercomputer, intended for researchers and start-ups, and delivered the first unit to OpenAI. The company’s newest “out of the box” AI supercomputer is dubbed Project DIGITS.

“Every software engineer, every engineer, every creative artist, everybody who uses computers today as a tool, will need an AI supercomputer.”

Project DIGITS is powered by the Grace Blackwell GB110 chip, combining a Blackwell GPU and a Grace CPU, and integrates Nvidia’s full AI software stack. Project DIGITS is wireless and can serve as a workstation or cloud platform. Project DIGITS is in full production and set for release in May 2025.

About the Speaker

Jensen Huang is co-founder, president, and CEO of Nvidia, the world’s largest semiconductor company, specializing in high-end graphics processing units and platforms supporting AI development and implementation.