이거 보시면 엔비디아 다 알게 됩니다

Jensen Huang reveals that Nvidia's initial technology choice in 1993 was 'exactly wrong.' The company aimed to reinvent 3D graphics for PCs by turning them into game consoles, but the algorithm they developed did not work. By 1995, with 35-40 competitors in the same space, Huang realized the company would fail unless they confronted the failure and adopted the correct approach, which they did by learning from textbooks.

After discovering their original algorithm was flawed, Jensen Huang bought three textbooks on OpenGL pipeline design from Fry's Electronics for a few hundred dollars. He gave them to his engineers, and this self-taught approach led Nvidia to reinvent computer graphics and become the world leader in the field, inventing most major breakthroughs in the last 25 years.

Huang explains that Nvidia's enduring big idea was augmenting CPUs to solve problems otherwise too difficult, such as molecular dynamics, image processing, inverse physics, and deep learning. He emphasizes that the company's success came from focusing on accelerating algorithm domains rather than just building great chips, and that a unique, deeply held perspective about the world is more important than technology or market timing.

Jensen Huang explains his management philosophy by comparing the CEO to an F1 driver building a race car that fits them perfectly, rather than adapting to a standard car. He argues that the organization should be constantly tweaked to the leader's needs to remain competitive, and that the next CEO can reshape the company afterward. This 'founder mode' has scaled NVIDIA from zero to a $5 trillion company over 34 years, according to the interviewer.

Huang emphasizes that systems understanding, awareness, design, and organization will be the most critical skills as low-level tasks become automated by AI agents. He notes that NVIDIA's designers are already systems designers because chip synthesis is automated, and that software development will similarly shift to abstract problem-solving. Understanding constraints like processor, memory, and networking at a technical level will remain valuable even as agents handle implementation.

Huang describes how current AI agents already exhibit coarse recursive self-improvement by updating markdown files, long-term memory, and knowledge graphs asynchronously with each use. However, the key breakthrough needed is fine-grained controllability—being able to change a single word in a plan file or one component in a CAD file and have the agent regenerate everything else accordingly. He argues that agents don't need 100% accuracy; even 80% is sufficient if humans can guide the rest.

Jensen Huang explains that the breakthrough in generative video, specifically seeing a neural network generate articulation like a finger moving or a hand picking up a glass, convinced him that robotics articulation was imminent. This realization led Nvidia to develop 'physical AI' and 'world foundation models' that understand physics, causality, friction, and tension. He states that the 'ChatGPT moment of robots' occurred a couple of years ago, when robots could walk and be fine-tuned via reinforcement learning grounded in physics.

Huang outlines the three essential systems for advancing robotics: creating environments for robots to learn and evaluate (real-to-sim), developing simulators based on both grounded physics simulation and generative physics simulations (using Isaac Sim and Cosmos), and finally sim-to-real transfer using reinforcement learning grounded in physics and electromechanical systems. He compares this pipeline to the post-training phase of agentic AI systems.

Huang identifies self-driving cars as the first economically significant application of physical AI because they have a large enough market, relatively standardized technology for scaling, and real economic value. Nvidia's physical AI business, including autonomous vehicles, is already approximately $10 billion. He predicts it will become the next $100 billion business, taking less than 10 years but more than two or three.

Nvidia open-sourced its self-driving car stack (Alamo) because autonomous navigation is needed across many smaller markets like agriculture, mail delivery, and warehouse AMRs, none of which are individually large enough to justify a dedicated stack. By creating a comprehensive open platform, Nvidia enables diverse applications while maintaining its core business in chips and data centers for companies like Waymo, Tesla, and Mercedes.