Nvidia CES 2025 Keynote

Key Themes and Ideas: AI Tokens as the Foundation of Intelligence: Huang introduces AI tokens as the "building blocks of AI," emphasizing their ability to transform raw data into meaningful insights and actions. "Tokens transform words into knowledge and breathe life into images...they bring meaning to numbers to help us better understand the world around us." Tokens are presented as fundamental to various AI applications, spanning text, images, videos, and even robotic movement. The entire architecture is designed around improving the token generation speed and lowering its cost. NVIDIA's Historical Role and the Evolution of AI: The keynote traces NVIDIA’s journey from the creation of the GPU, CUDA, to its pivotal role in the AI revolution. "GeForce enabled AI to reach the masses and now AI is coming home to GeForce" It highlights key moments like the AlexNet breakthrough using CUDA in 2012, and the impact of Google's Transformer architecture (BERT) in 2018 as accelerators for AI progress. NVIDIA positions itself as the driving force behind the advancements from perception AI to generative AI, to agentic AI and ultimately physical AI. The Blackwell Architecture and Its Capabilities: The RTX Blackwell family GPUs are introduced as a "beast" with significant performance enhancements (e.g., 92 billion transistors, 4,000 TOPS, 4 PetaFLOPS of AI). It emphasizes significant advancements in ray tracing performance and the integration of AI to predict and generate pixels, resulting in much higher performance. "We computed only two million pixels and have AI predict all of the other 33" The new architecture allows for the intermixing of AI and computer graphics workloads, and enables programmable shaders to process neural networks. The improvements in Blackwell are focused on improved cost per token, performance per watt and cost of training. AI Scaling Laws and Test-Time Scaling: The keynote introduces "test time scaling" as a new scaling law, where AI dynamically allocates computational resources to improve reasoning capabilities and solve complex problems. "It might reason about it, might break down the problem into multiple steps." Test time scaling is presented as a method of providing the AI with the capacity for "long thinking", allowing for more detailed and effective problem solving. This emphasizes the importance of continually training and improving AI models. Agentic AI: The Future of Digital Workforce: NVIDIA is building the tools for "agentic AI" which entails the creation of "digital workforce" systems which can work along side or in the place of human employees. NVIDIA Nim microservices are presented as packaged AI models that can be deployed anywhere, and NVIDIA Nemo provides a framework for onboarding, training, and evaluating AI agents, like you would an employee. "The IT department of every company is going to be the HR department of AI agents in the future." The company highlights several real world examples of AI agents that can enhance existing systems from software coding, scientific research, security to business integration. The concept of the AI agent is presented as a likely “multi-trillion dollar opportunity”. NVIDIA's Llama Neotron Models and Enterprise AI: NVIDIA has fine-tuned Llama models for enterprise applications, creating the Llama Neotron Suite of open models. These models are designed for various use cases, including fast response times (Super Llama Neotron Supers) and advanced teacher models (Ultra models). This highlights NVIDIA’s focus on practical, business-oriented AI solutions. This was done due to their belief that the original llama models could be better "fine tuned for enterprise use."