Jensen Huang on AI, NVIDIA, and the Future of Computing
Designing computers at the scale of AI
Lex Fridman asks NVIDIA CEO Jensen Huang how the company’s work has expanded from GPU design to extreme co-design across chips, networking, power, cooling, software, and data centers. Huang says distributing workloads across many computers creates problems that require coordinating the full system. He describes NVIDIA’s broad, technically focused leadership team and its approach of solving design questions together.
From accelerators to CUDA
Huang traces NVIDIA’s move from specialized accelerators toward general-purpose computing, including programmable shaders, FP32, and CUDA. He explains why CUDA was put on GeForce despite its cost: the installed base could help attract developers. He says the investment strained the company’s gross profit, but helped establish the platform that later supported deep learning.
“The install base is in fact the single most important part of an architecture.” — Jensen Huang
Leadership and anticipating change
Huang describes how he reasons through future possibilities and gradually builds support among employees, the board, customers, and industry partners before announcing major decisions. He says NVIDIA uses research, conversations across the AI industry, and flexible architecture to anticipate changing models and computing needs.
Scaling AI and building AI infrastructure
Huang discusses four scaling approaches: pre-training, post-training, test-time reasoning, and agentic scaling. He argues that synthetic data can extend training, that inference requires substantial computation because it involves reasoning, and that agents can use tools and spawn sub-agents. He connects these developments to changing hardware needs, including NVIDIA’s designs for systems running large language models and agentic workloads.
Agents, tools, and security
Huang argues that useful AI agents will need access to files, research, and tools, comparing their use of software to a person using tools in a home. The conversation turns to OpenClaw and security. Huang says NVIDIA introduced OpenShell and NemoClaw, describing controls that limit agents’ access to sensitive information, code execution, and external communication.
Power, supply chains, and data centers
Huang says NVIDIA is working to improve tokens generated per second per watt and reduce token costs. He discusses the challenges of expanding the AI supply chain and the coordination required to manufacture complete rack-scale systems. He also proposes using spare grid capacity: data centers could reduce or move workloads when the grid is strained, while utilities could offer different levels of power availability.
Fridman asks about Elon Musk’s approach to building data centers. Huang praises Musk’s systems thinking, questioning of existing processes, attention to work on the ground, and urgency. Huang then explains NVIDIA’s “speed of light” method: compare designs with physical limits, question why current processes take as long as they do, and make systems as complex as necessary but as simple as possible.
China, open-source AI, and TSMC
Huang attributes China’s technology development to factors including education, competition among cities and companies, open-source collaboration, and a culture that values engineering. He says NVIDIA develops and releases open models both to understand changing AI architectures and to broaden access to AI across industries and countries, including fields beyond language.
He describes TSMC’s strengths as its technology, its ability to coordinate complex customer demands while maintaining manufacturing performance, and a culture that combines technology development with customer service. Huang confirms that Morris Chang once offered him the chance to lead TSMC, but says he declined because he believed NVIDIA’s work was important.
NVIDIA’s platform, AI factories, and space
Huang identifies CUDA’s installed base, developer trust, execution speed, and broad ecosystem as NVIDIA’s advantages. He describes the changing unit of computing—from a GPU, to a computer, to a cluster, and now an AI factory—and explains his view that AI is shifting computers from retrieving stored information to generating context-aware output.
The conversation also covers computing in space. Huang says NVIDIA GPUs are already used there for satellite imaging, where AI can process data at the edge. He discusses engineering challenges such as cooling, radiation, redundancy, and performance degradation, while saying that reducing waste on Earth is an immediate opportunity.
Work, AI, and human abilities
Huang explains how he handles pressure: break problems into manageable parts, share concerns with people who can act, and focus on what can be done. He discusses resilience, learning from others, and being willing to reason in public and have others challenge the steps in his thinking.
Games and DLSS 5
Asked about criticism that DLSS 5 could make games look like “AI slop,” Huang says he understands the concern. He describes DLSS 5 as guided by 3D geometry, structured data, and artists’ textures and intent, and says it is a tool developers can choose to use. He names Doom as especially influential and also discusses Virtua Fighter, Cyberpunk 2077, and Skyrim.
AGI, jobs, and using AI
Huang says AI systems may already meet one definition of AGI: creating a company that earns more than a billion dollars, even temporarily. He distinguishes the purpose of a job from its tasks, using radiology and software engineering to argue that automating tasks can increase demand for people doing the wider work. He acknowledges that AI-driven changes will cause pain and advises people across professions to learn how to use AI.
Intelligence, humanity, and mortality
Huang distinguishes intelligence as a functional capability from humanity, which he associates with character, compassion, generosity, and lived experience. He says he does not think chips will feel anxiety or other emotions, while remaining open to being surprised by what scaling can achieve. On mortality and succession, he says he wants to keep working and describes continuously sharing knowledge with his team.
Hope for the future
Huang says he has confidence in people’s kindness and in humanity’s capacity to solve problems. He describes possible advances in health, pollution reduction, space travel, and scientific understanding, and closes by thanking Fridman for the interview.