NVIDIA (NVIDIA) founder and CEO Jensen Huang described artificial intelligence (AI) as the next industrial revolution in his keynote speech at the GTC conference in Washington, D.C., and likened it to the greatest test of America's capabilities since the dawn of the space age. He emphasized that we are at the forefront of technology and will move forward at the fastest speed. NVIDIA has announced a series of revolutionary new products and partnerships aimed at creating “AI factories” for scientists, engineers and dreamers around the world.
The computing model revolution begins to accelerate computing and the end of Moore's Law
Jensen Huang reviewed the era of innovation in American history, from the birth of the transistor at Bell Labs, to the IBM System/360, to Apple's personal computer. He noted that Dinard scaling, the foundation of the semiconductor industry, stopped about a decade ago. Transistor performance and power consumption growth has slowed significantly due to the laws of physics, meaning Moore's law cannot continue.
In order to solve problems that cannot be solved by general-purpose computers, NVIDIA invented a new computing model "accelerated computing." NVIDIA invented the GPU and CUDA programming model to extend computing capabilities beyond traditional CPUs by leveraging parallel computing capabilities. CUDA is already the core treasure of the company. After 30 years of development, CUDA 13 has been launched, CUDA 14 will be launched soon, and hundreds of millions of GPUs are running fully compatible CUDA. Even accelerated computing requires redesigning algorithms, creating new libraries, and rewriting applications. NVIDIA has developed more than 350 CUDA X libraries for different fields such as computational exposure (cuLitho), numerical optimization (cuOpt), medical imaging (Monai), and genomics (Ariel).
An AI factory that transforms from tools to workersHuang Renxun pointed out that the impact of AI goes far beyond chatbots. AI has revolutionized the entire computing stack, moving from traditional hand-written code to machine learning, a training data-intensive programming style that runs on GPUs. He explained that the computing unit of AI is Token. Token can represent text, images, videos, 3D structures, and even chemicals, proteins, genes, etc. Once information is tokenized, AI can learn its language and meaning, translate and generate it.
In the past, the software industry was about creating tools like Excel or browsers. However, the essence of AI is work, and AI systems are essentially workers who can use these tools. This new computing demand requires a new type of system—AI factory. Unlike the general data centers of the past that were used to store files and run multiple applications, AI factories are specifically designed to produce tokens, which must be as valuable as possible and must be produced at extremely high rates while being cost-effective.
AI development is facing two exponential growth pressuresHuang Renxun emphasized that the development of AI is currently facing two exponential growth pressures. First, as AI models move from pre-training, post-training to inference, the computing requirements increase exponentially. Second, as AI models become smarter, people are willing to pay for them, creating a virtuous cycle. The smarter the model, the more people use it. The more you use it, the more computing resources you need.
In order to maintain this virtuous cycle and significantly reduce costs at the end of Moore's Law, NVIDIA adopted Extreme Co-Design. This design starts from scratch, taking into account new computer architectures, chips, systems, software, model architectures and applications. Through this approach, NVIDIA launched the Blackwell architecture and Grace Blackwell NVLink 72 system. The GB200 was the first rack-scale AI supercomputer, and the next generation is the Vera Rubin.
Jen-Hsun Huang showed the design of the Vera Rubin system, which treats the entire rack as a computer, connecting 72 GPUs (or 144 GPU cores) into a huge virtual GPU through NVLink 72. The overall Grace Blackwell NVLink 72 performance gains are staggering. Grace Blackwell delivers 10x more performance per GPU than the H200. Although GB200 is one of the most expensive computers, its Token generation capability is extremely strong and it can produce Tokens at the lowest cost in the world. This demonstrates the benefits of extreme co-design: 10x performance improvement and 10x cost reduction.
NVIDIA expects Blackwell and Rubin architectures to bring huge market growth. Blackwell and Rubin's early-stage cumulative business visibility has reached $500 billion by 2026, which is five times Hopper's lifetime growth rate. Additionally, NVIDIA emphasized its commitment to bringing manufacturing back to the United States, with Blackwell and future generations of AI factory systems set to be manufactured in locations such as Arizona, Indiana, and Texas.
Innovative breakthroughs and partners in six major fieldsIn his speech, Huang Renxun pointed out that NVIDIA announced major progress in six key areas in order to develop the entire AI ecosystem:
1. 6G and telecommunications infrastructure: In order to put the United States at the core again in the 6G revolution, NVIDIA and Nokia have reached a major cooperation. NVIDIA has launched a new product line "NVIDIA Arc" (Aerial Radio Network Computer). "NVIDIA Arc" combines Grace CPU, Blackwell GPU and ConnectX network to run the Aerial library. Nokia will use "NVIDIA Arc" as its future base station and can upgrade millions of Airscale base stations around the world to implement 6G and AI functions.
2. Quantum computing: The future of quantum computing is the collaborative work of GPU supercomputing and quantum processors (QPU). To this end, NVIDIA has released NVQLink, a new interconnect architecture that can directly connect quantum processors and NVIDIA GPUs. NVQLink is capable of transmitting terabytes of data thousands of times per second for quantum error correction and AI calibration. Currently, the platform is supported by 17 quantum computing companies and nearly all U.S. Department of Energy (DOE) laboratories.
3. Enterprise computing and network security: AI will greatly enhance network security challenges, requiring more powerful defenders. NVIDIA partners with CrowdStrike to create AI agents for cybersecurity in the cloud and edge. Additionally, NVIDIA is working with Palantir to accelerate its Ontology platform to process structured and unstructured data at the speed of light for national security and enterprise insights.
4. Robots and physical AI (Physical AI): Physical AI requires three types of computers, namely Grace Blackwell for training models, Omniverse computers for simulation (digital clones), and Jetson Thor robot computers for operation. NVIDIA launched Omniverse DSX as a blueprint for building and operating Gigascale AI factories. DSX enables co-design of computing, power and cooling facilities. NVIDIA is collaborating with Hon Hai Figure, Agility and Disney Research to launch Disney Blue, a cute robot whose movements are fully simulated and learned in Omniverse.
5. Autonomous driving: Robots are being commercialized on wheels. NVIDIA has released the Drive Hyperion platform, a standardized sensor suite and computing platform designed to enable global automotive companies such as Lucid, Mercedes-Benz, Stalantis and others to develop Robo Taxi. NVIDIA also announced a partnership with Uber to connect Drive Hyperion cars to the global network.
6. Open-Source Models: Huang Renxun emphasized that the United States must lead the development of open-source models (Open-Source Models) because this is the lifeblood of scientists, researchers, and new startups. NVIDIA has always been a leader in open source contributions, with 23 models on major rankings, ranging from language models, entity AI models to biological models.
Huang Renxun concluded that we are undergoing two platform transformations, from general computing to accelerated computing, and from traditional handwriting software to AI. NVIDIA is building new platforms for 6G (ARC), robotic cars (Hyperion) and AI factories (DSX), and bringing manufacturing back to the United States.