Jensen Huang 63 Minutes Deep Dive into NVIDIA's Next Decade: From CUDA to Physical AI and Robotics

Compiled by: Deep Tide TechFlow

Deep Tide Guide: In this 63-minute in-depth interview, renowned tech creator Cleo Abram avoided the usual financial report numbers and stock price hype, engaging in a fundamental dialogue about the evolution of computing and the future of human technology with Jensen Huang, co-founder and CEO of NVIDIA. From breaking through with gaming graphics cards 30 years ago, to enduring over a decade of skepticism from Wall Street while building CUDA, to keenly capturing the deep learning revolution ignited by AlexNet in 2012, Jensen Huang deeply analyzed how NVIDIA has gradually restructured the modern computing technology stack that has lasted for 65 years.

Facing the transformative shift from generative AI to Physical AI, Jensen Huang elaborated on the technological foundation of "world model + physical simulation" built by Cosmos and Omniverse, depicting a ten-year vision where "everything will have a robotic form" and everyone will have a lifelong companion in the form of an R2-D2 robot. At the same time, he firmly responded to philosophical differences regarding chip architecture specialization versus generalization, dissected cutting-edge bets such as digital biology and climate prediction, and offered core advice to the younger generation to embrace the art of questioning. When asked how he hopes to be remembered in the future, he gave the simplest and grandest answer: "Very simple: they made an extraordinary impact."

Guest: Jensen Huang (Co-founder and CEO of NVIDIA)

Host: Cleo Abram (Renowned tech video creator, host of Huge If True channel)

Podcast Source: Huge If True (https://www.youtube.com/watch?v=7ARBJQn6QkM)

Original Title: NVIDIA CEO Jensen Huang's Vision for the Future

Broadcast Date: January 27, 2025

Key Points Summary

· A perfect computer must balance serial and parallel computing: In the vast majority of complex programs, about 10% of the core code accounts for 99% of the computational load, which is extremely suitable for large-scale parallel acceleration by GPUs; while the remaining 90% of control logic is efficiently processed serially by CPUs. The original intention of NVIDIA was to solve the ultra-high load problems that traditional computers could not tackle.

· The gaming industry is the best funding engine for underlying research and development: Real-time rendering of 3D virtual worlds inherently requires parallel computing. More importantly, the video game industry has an extremely large consumer market scale, capable of continuously providing self-funding for billions of dollars in underlying software and hardware research and development, without relying on government grants or research funding.

· The essence of CUDA is to democratize parallel computing: Early researchers had to disguise physical or mathematical problems as graphics shading programs to invoke GPUs. NVIDIA created CUDA, allowing developers to directly utilize GPU computing power with the general C language. Although this gamble faced over a decade of cold reception and skepticism from Wall Street due to high research and development costs, it ultimately paved the way for the subsequent explosion of deep learning.

· AlexNet prompted NVIDIA to completely restructure the underlying computing stack: AlexNet, created in 2012 by Ilya Sutskever, Alex Krizhevsky, and Geoffrey Hinton, crushed traditional image classification algorithms using just two GeForce GTX 580 graphics cards. Jensen Huang keenly realized that software was shifting from "manually written rules" to "data-driven learning," and immediately led NVIDIA to bet everything on AI, comprehensively restructuring chips, interconnects, algorithm libraries, and system architecture.

· The core of generative AI is to become a cross-modal universal translator: This wave of AI is unprecedented because architectures like Transformers can understand and translate data from any modality (text, images, speech, 3D space, robotic arm trajectories), breaking down the barriers between human knowledge and the digital world.

· Physical AI must be driven by a dual engine of "world model + physical simulation": Large language models can produce hallucinations and need to rely on retrieval and contextual anchoring for real answers; in the physical world, robots must possess physical knowledge such as gravity, inertia, friction, and spatial constancy. NVIDIA launched Cosmos as the foundational model for the physical world and used the Omniverse simulation engine, which follows Newtonian physics, to generate realistic synthetic data, allowing robots to complete millions of extreme condition rehearsals before entering the physical world.

· In the next decade, everything will have a robotic form: Whether it's autonomous vehicles, smart buildings, robotic lawn mowers, or humanoid robots, all moving objects in the future will possess robotic capabilities. Everyone will have a personal lifelong digital companion (R2-D2) that will continuously evolve with the user's growth, taking forms across glasses, phones, computers, cars, and home entities.

· Reject the trap of specialized ASICs and uphold innovative computing flexibility: In response to the viewpoint of completely hardening Transformers into dedicated chips (ASICs), Jensen Huang believes that innovation will never stop at any specific algorithm architecture. If computers are solidified for a single purpose, it is like turning a computer into a microwave. NVIDIA has always insisted on building flexible and programmable acceleration architectures to support academia in continuously exploring more efficient new attention mechanisms like FlashAttention and hierarchical attention.

· From a company that grew up playing games to the storm center of human frontier science: In the RTX 50 series, only 500,000 pixels of the 8 million pixels in a 4K image are calculated using traditional pipeline ray tracing, while the remaining 7.5 million pixels are perfectly predicted and filled in by AI. This paradigm allows scientists and engineers to focus their valuable energy on core critical points, amplified by AI into superhuman productivity. From digital biology, protein language parsing, microclimate prediction to future robotics, NVIDIA's ultimate historical vision is to "create an extraordinary impact."

Remarkable Insights

· On technological faith and forward-looking layout: "At some point, you have to choose to believe in something. If you build it, people may not come; but if you don't build it, people will never have the chance to come. Our strategy is built on hope, but it is a rational hope that has been deeply considered."

· On the division of labor between serial and parallel computing: "Perhaps only 10% of the code in a program accounts for 99% of the computational processing, and this 99% of tasks can be done in parallel; but the remaining 90% of the code still must be executed in order serially. A truly perfect computer is not either/or, but must possess excellent serial and parallel processing capabilities simultaneously."

· On the driving force behind the birth of CUDA: "Many great innovations often arise from the same complex hodgepodge: part comes from lofty ambitions and inspiration, while another part purely comes from urgent self-rescue when in a dead end. On one hand, we see researchers at Boston's Massachusetts General Hospital using gaming graphics cards to reconstruct CT images, while on the other hand, we ourselves feel stuck when rendering flowing water, explosions, and particle collisions in games. These internal and external factors intertwined to facilitate the birth of CUDA."

· On the debate over the specialization of computing architectures: "Will Transformers be the last AI algorithm architecture discovered by human researchers? Absolutely not. Looking back at the history of computer science, no algorithm has ever dominated everything forever. If you had completely solidified the algorithms of a decade ago into chip hardware, then computers would have become microwaves, and subsequent innovations would be impossible to discuss."

· On the essential relationship between AI and human work: "In the approximately 8 million pixels of a 4K display, we actually only ray traced 500,000 of those pixels, while the remaining 7.5 million pixels were perfectly predicted and filled in by AI. This is a vivid metaphor for how AI will make humans superhuman: humans can concentrate their valuable energy and attention on the most critical core parts, while the remaining heavy and lengthy processes are handed over to AI, achieving a hundredfold leap in productivity."

· On survival guidelines for the younger generation: "If I were still a student today, the first thing I would do is learn AI. Learning how to interact with AI is very much like becoming someone who is very good at asking questions. You can't just randomly ask questions; making AI your assistant requires mastering the skills and art of prompting. Whether you study law, medicine, chemistry, or biology, you must continuously ask yourself: How can I use AI to do my job better?"

· On life's endpoint and historical legacy: "Very simple: they made an extraordinary impact. I hope the next generation not only knows that NVIDIA is the company that grew up with them playing games, but also sees that the storm center of the entire digital biology, materials science, robotics, and autonomous driving revolution is this company that adheres to its core beliefs."

Main Text

The Birth of Heterogeneous Computing: Why Does 10% of the Code Need a Dedicated GPU?

Cleo Abram: In today's discussion, I want to avoid the quarterly financial reports and stock price numbers that we often see in financial news and focus on more fundamental questions: What are you actually building? How has computing evolved over the past few decades to reach today? And how will future technologies shape the lives of ordinary people?

Jensen Huang: Great, I really look forward to this exchange. In fact, at some point, you have to choose to believe in something. What we are doing now is essentially a complete reconstruction of the modern computing architecture that has lasted for decades.

Cleo Abram: To understand the current AI era that has captured the world's attention, we must first go back to the 1990s. At that time, game developers wanted to create more realistic and immersive 3D graphics, but the traditional hardware at the time simply could not handle the real-time computation of massive pixels. Why couldn't computers at that time handle graphics? How did GPUs come into being?

Jensen Huang: This stems from a core characteristic of computer programs: in a piece of software code, typically only about 10% of the key instructions account for a full 99% of the computational load. This 99% of intensive computation usually has a very high degree of repetitiveness, making it extremely suitable to be split into thousands of tiny tasks for parallel execution; while the remaining 90% of the code mainly involves system control, branching, and scheduling logic, which must be processed step by step in order.

At that time, traditional computers were completely dominated by the central processing unit (CPU), which was designed to excel at high-speed processing of serial logic, but proved powerless in the face of the massive parallel demands of rendering vast numbers of pixels simultaneously. We keenly observed that a truly ideal computer must not only have serial processing capability but also parallel processing capability; it must possess both simultaneously.

Thus, we founded NVIDIA, determined to create dedicated acceleration processors that could solve problems beyond the reach of traditional computers. This is the origin of the birth of the Graphics Processing Unit (GPU).

Cleo Abram: This reminds me of a very classic science popularization video released by NVIDIA over a decade ago: the Mythbusters team built a robot that shot paintballs one by one with a single nozzle to demonstrate how a CPU processes tasks; then they introduced a massive mechanical array, with thousands of nozzles firing paintballs simultaneously in a thousandth of a second, instantly creating a complete image of the Mona Lisa, vividly showcasing the parallel acceleration power of the GPU.

Jensen Huang: That demonstration intuitively showcased the essential differences between the two. Gaming was the first field to have an extreme demand for such large-scale parallel computing power.

Breaking the Game: Feeding Hardcore Underlying Research with the Mass Consumer Market

Cleo Abram: Why did NVIDIA choose to first enter the field of video games instead of directly engaging in scientific computing?

Jensen Huang: Firstly, because we genuinely love this application. 3D video games are digital simulations of the real world or fantastical worlds; who wouldn’t want to enter a realistic and beautiful virtual world?

But from a business strategy perspective, there is a deeper decisive consideration: we foresaw that video games would likely become one of the largest consumer entertainment markets for all humanity. Developing top-notch semiconductor chips requires an unimaginable amount of R&D funding. If we only served niche high-performance scientific computing laboratories, we would have to rely on government research grants or expensive custom contracts, which could not sustain continuous massive technological iterations.

Video games, on the other hand, provide an incredibly vast consumer market. Tens of millions of players are willing to pay for smoother and more stunning graphics, allowing us to become the world's largest manufacturer of parallel processors. The enormous shipment scale provided us with a continuous and self-sustaining source of R&D funding, enabling us to push this parallel architecture to extreme physical limits over the following decades.

A Bold Bet on CUDA: From Medical Imaging Inspiration to a Universal Parallel Computing Platform

Cleo Abram: By the early 2000s, some perceptive scholars and developers began to realize that GPUs could not only render game textures but also run physics collision simulations and even be used by external scientists for complex scientific calculations. But my understanding is that achieving this at that time was extremely painful; scientists had to disguise matrix operations as triangle pixels or texture maps to trick the graphics card into computing. Then you made an extremely bold decision: to launch the CUDA platform. Can you talk about the background at that time?

Jensen Huang: You’ve done your homework; that’s very accurate. Many great innovations often arise from a complex mix of factors: part ambition and inspiration, and part pure desperation for self-rescue.

One of the earliest external cases using our GPU for general acceleration came from Massachusetts General Hospital in Boston. Researchers there used our gaming graphics cards for CT image reconstruction, which gave us tremendous inspiration.

At the same time, NVIDIA was also facing bottlenecks: we wanted the water flow in the gaming world to behave like real water, the thousands of fragments produced by explosions to conform to aerodynamic principles, and smoke to diffuse naturally. This required the graphics card to run particle physics and fluid dynamics simulations simultaneously alongside graphics processing. If the graphics card hardware pipeline was rigidly fixed in the traditional computer graphics process, none of this would be possible.

So we resolved to completely break the constraints of the dedicated graphics pipeline, transforming it into a universal parallel computing architecture, and created the Unified Computing Architecture (CUDA). CUDA allows engineers and scientists worldwide to write code directly in the C language they are already familiar with, freely utilizing thousands of computing cores within the GPU.

Cleo Abram: Looking back from today’s perspective, creating CUDA was a visionary gamble. At that time, NVIDIA bet everything on CUDA, forcibly integrating general computing logic into all shipped gaming graphics cards, leading to a significant increase in chip area, high costs, pressure on profit margins, and facing severe criticism and skepticism from Wall Street for over a decade.

Jensen Huang: When you create a brand new disruptive technology, critics will always say, "If you build it, people might not come." But from an optimistic perspective, our belief was, "But if you don’t build it, people will never have the chance to come."

We were well aware of the immense physical potential contained in this technology, and we firmly believed that through the vast market of consumer GeForce graphics cards, we could lay this advanced parallel computing architecture on the desks of every geek and researcher worldwide. It was this rational hope that supported us through those long and lonely years.

The Epiphany of AlexNet in 2012: Reconstructing a 65-Year-Old Modern Computing System

Cleo Abram: Then we arrived at 2012, witnessing a milestone battle in the history of technology: the ImageNet image recognition competition. Alex Krizhevsky, Ilya Sutskever, and their mentor Geoffrey Hinton submitted a deep neural network model called AlexNet, which directly crushed all traditional computer vision algorithms and significantly reduced error rates. They trained this groundbreaking model using two consumer-grade GeForce GTX 580 graphics cards. What was your first reaction when you saw this result within NVIDIA?

Jensen Huang: That was a moment that completely changed the trajectory of world history. At that time, an internal computer vision R&D team at NVIDIA was trying to solve the image recognition problem using traditional algorithm libraries and faced significant obstacles and frustrations. When AlexNet emerged, we were not only shocked but also felt a strong resonance: this new algorithm completely abandoned the old path of manually written rules, relying entirely on massive data to self-train weights, achieving an unprecedented leap in capability.

We immediately took a step back and asked ourselves a fundamental question: "If this deep learning architecture can exhibit such terrifying scalability in computer vision, what are its limits? If it can be infinitely scaled, what does that mean for the entire computer industry and the underlying software and hardware architecture?"

We made a crucial judgment at that time: the paradigm of software development was undergoing an essential shift. Since the IBM System/360 established the instruction set architecture of modern general-purpose computers in 1964, for 65 years, humanity has been telling computers what to do by precisely writing instructions step by step; the rise of deep learning marked the beginning of computers transitioning to self-learning patterns through massive data.

Since the essence of software has changed, the underlying hardware and the entire system stack supporting the software must be completely rebuilt. We then made the decision to fully bet on AI, redesigning the entire computing stack from single-chip architecture, interconnect buses, network communication, algorithm libraries to distributed systems. The subsequent DGX supercomputer system originated from this epiphany.

Why Now: A Universal Translator Across All Data Modalities

Cleo Abram: Why did the public only suddenly feel the explosion of AI in recent years? What is the essential technological leap behind this?

Jensen Huang: Because over the past decade, hardcore challenges in the field of computer science have been tackled one after another. Initially, computer vision was completely solved, followed by speech recognition reaching superhuman levels, and then natural language understanding achieving qualitative breakthroughs.

But the real catalyst for a comprehensive qualitative change was the maturity of the Transformer architecture and generative AI. Previous AI models were mostly single-task classifiers, such as determining whether a picture contains a cat; whereas today’s generative AI is a universal "cross-modal translation system."

It can understand text and translate it into another text, such as summarizing paragraphs or translating across languages; it can also translate text into images, images into 3D assets, and sensor data into the rotation angles of robotic arms. It can seamlessly traverse various representations between the digital and physical worlds, which is why today’s AI feels so omnipotent.

Physical AI and World Models: Cosmos Endows Common Sense, Omniverse Anchors Physical Laws

Cleo Abram: Next, I want to delve into one of the most important transformations happening: moving from the textual and visual elements of the digital world to embodied intelligence and robotics in the physical world. In the past, training robots required placing them in real workshops to walk step by step, or having humans wear motion capture suits to collect extremely limited data, which could easily wear out or damage the machines, making data collection very expensive and slow. Now we are starting to train robots in virtual worlds. NVIDIA has launched Cosmos and Omniverse; how are you solving the robot learning challenges in the virtual world?

Jensen Huang: The best reference frame for understanding physical AI is the development path of language models like ChatGPT.

When ChatGPT first appeared, although it performed impressively, it had a fatal weakness: hallucinations. If it rambled on for too long or discussed untrained blind spots, it would speak nonsense seriously because its output was not anchored in factual truth (Ground Truth). Later, the industry introduced Retrieval-Augmented Generation (RAG); when you feed the model an authoritative PDF or connect to the internet for real-time searches, this PDF and search results become the model's benchmark facts, providing a reliable grounding for the model's reasoning.

Now we project the same logic onto the physical world. For a robot to move freely in the real world, it first needs a foundational model with physical common sense. What is physical common sense? It must understand that the Earth has gravity, objects have inertia and friction; it must possess geometric and spatial perception abilities; it must understand object permanence, meaning that when I turn my head away and then back, the cup on the table is still there; it must understand causal logic, meaning that if I knock over the cup, water will spill out.

To inject this common sense into AI, we built Cosmos, a foundational language model for the physical world, specifically encoding the universal laws of the physical world.

But having Cosmos alone is not enough; we must provide it with an "absolute physical reality benchmark," just as we feed authoritative PDFs to ChatGPT. This benchmark is Omniverse. Omniverse is not an ordinary video game engine but a physical simulation system that strictly follows first principles, with its underlying mathematical models being rigorous solvers of Newtonian mechanics, fluid dynamics, and electromagnetism.

By combining the Cosmos world model with the Omniverse physical simulation engine, we can generate endless virtual environments and unexpected scenarios in digital space for robots that absolutely comply with the laws of physics: sudden changes in lighting, slippery surfaces, obstructed obstacles, mechanical failures. Robots can complete extreme training in hours that would take decades in the real world, and then deploy their models to real robots.

Vision for the Next Decade: Everything Will Have a Robotic Form, and Everyone Will Have Their Own R2-D2

Cleo Abram: Following this logic, what will the physical world look like in ten years? How will ordinary people interact with these physical AIs?

Jensen Huang: In the future, all mobile objects will possess robotic capabilities, and this day will come faster than most people imagine.

In the future, people will no longer have to push lawnmowers in their yards under the scorching sun, which seems somewhat absurd today. Every car will be a fully autonomous robot; humanoid robots will also enter factories and households. All these robots will complete their education in the digital twin world constructed by Omniverse and Cosmos, and then transition into physical reality.

What excites me even more is that everyone will have their own lifelong digital companion, just like R2-D2 in Star Wars. It may not necessarily be a round metal can; it could exist in your smart glasses, your smartphone, your office computer, or sit in your car, waiting to greet you in some physical robotic form when you return home.

This exclusive R2-D2 will accompany you throughout your life, growing alongside your experiences, deeply understanding your joys and sorrows, lifestyle habits, and work styles, becoming the strongest extension of your perception and transformation of the world.

Debate on Architectural Philosophy: Why Can't Transformers Be Directly Embedded in Chips?

Cleo Abram: While preparing for this interview, I discussed with many engineering friends, and they insisted I ask you this question: As the Transformer architecture becomes the foundation for most large models today, many chip companies are beginning to attempt to manufacture application-specific integrated circuits (ASICs) tailored specifically for Transformers. Specialization can indeed bring extreme efficiency for specific tasks, but it sacrifices general flexibility. How does NVIDIA balance between specialized optimization and general architecture? How do you make such a significant technological bet that concerns billions of dollars?

Jensen Huang: This touches on our core belief: Do you believe that the Transformer is the last AI algorithm architecture that human researchers can discover?

If your answer is yes, then you can certainly solder all the mathematical operations of the Transformer directly onto silicon chips; but our core belief is that the Transformer is by no means the endpoint; it is merely a stepping stone on the ladder to more advanced AI evolution. The next generation of cutting-edge architectures years from now may not even resemble the Transformer.

Looking back at the history of human computer algorithms and innovations, has there ever been a brilliant idea that could dominate forever? The answer is no. The essence of computing's charm lies in its ability today to run marvelous programs that no one could have dreamed of ten years ago. If you had completely solidified the chips of that time into single-function hardware, computers would have devolved into microwaves, and the prosperity of the entire application ecosystem would have come to an abrupt halt.

Take the core attention mechanism of the Transformer, for example; its original intention was to calculate the relevance of each word to all other words in the context. When the text has only 10 words, the computation is effortless; but when the context window expands to tens of thousands of words or millions of tokens, the computational cost of full-to-full calculations will explode quadratically, making it impossible to implement on physical hardware. To solve this problem, in recent years, new mathematical algorithms such as FlashAttention, hierarchical attention, and waveform attention have emerged.

Computer science has no end, AI researchers have not given up on exploration, and NVIDIA will not either. Our most important mission is to always create a highly programmable accelerated computing system that allows the world's smartest algorithm scientists to freely explore the next generation of unknown great algorithms in this ocean of computing power.

Limits of Chip Manufacturing and Physical Challenges: From Wafer Packaging to Extreme Energy Efficiency

Cleo Abram: Today's top chips are approaching the fundamental limits of physics. You design chips, and partners like TSMC are responsible for photolithography manufacturing them at a few nanometers in physical scale. How does NVIDIA think about design in response to the slowdown of Moore's Law and power consumption bottlenecks?

Jensen Huang: The traditional Moore's Law and Dennard Scaling have essentially reached their physical limits. The era of relying solely on shrinking transistor sizes to gain power increases and reduced power consumption for free is long gone.

This is why we must completely shift from "chip thinking" to "data center thinking." The optical limit size of a single wafer is fixed; to break through the physical limits of a single chip, we have adopted chip-on-wafer integration packaging (CoWoS), ultra-high bandwidth interconnect links (NVLink), and extremely high-density on-chip high bandwidth memory (HBM).

More importantly, parallel accelerated computing follows completely different physical laws. Assigning a complex task to thousands of efficient cores for parallel processing can improve energy utilization efficiency by dozens or even hundreds of times compared to forcing a few high-frequency general CPU cores to handle it. Through collaborative optimization at the algorithm, chip, network, and system levels, we have opened up a new path for continuously driving exponential growth in computing power in the post-Moore's Law era.

The Next Grand Bet: Physical AI, Digital Biology, and Microclimate Simulation

Cleo Abram: Looking back, NVIDIA's every significant bet on GPUs, CUDA, deep learning, and autonomous driving has been proven correct by time. Besides the robots mentioned earlier, what other major bets do you have coming up?

Jensen Huang: In addition to physical AI and the hardware and software toolchain for humanoid robots, one area I am extremely excited about is digital biology.

In the past, computers understood artificial code made up of 0s and 1s; now AI understands human natural language. The essence of biology is the complex interactive language of molecules, amino acids, proteins, and living cells. Just as we use physical models to understand the real world, if we can use AI to deeply analyze and predict the interactive language between biological macromolecules, we can construct a digital twin of the human body. This means that drug discovery, gene therapy, and disease mechanism research will shift from the past trial-and-error methods to high-precision science that can be rapidly simulated and iterated on computers.

Another area I am very proud of is climate science. With digital twin technology, we are building ultra-high-resolution regional climate and extreme weather prediction systems that can accurately simulate micro-meteorological evolution within a kilometer above your head.

Scientific tools are essentially "time machines" invented by humans. Scientists strive to predict the future to see the trajectory of future evolution in advance, allowing them to make optimal designs and interventions in the present to create the best possible future version.

Neural Rendering and Personal Supercomputers: The Productivity Leap Brought by RTX 50 and DGX

Cleo Abram: At the recently concluded Consumer Electronics Show, you showcased the latest GeForce RTX 50 series graphics cards and the new generation DGX system on the table. What do these new hardware mean for ordinary creators and developers?

Jensen Huang: This contains a fascinating technological cycle: Initially, GeForce gaming graphics cards nurtured modern AI, helping Alex Krizhevsky and Geoffrey Hinton train AlexNet; today, powerful and mature AI is feeding back into GeForce graphics cards, completely overturning the underlying logic of graphic rendering.

In the 4K resolution images presented by the latest RTX 50 series graphics cards (approximately 8 million pixels in total), our physical ray tracing pipeline actually only densely calculated 500,000 of those pixels, while the remaining 7.5 million pixels were perfectly inferred and filled in by deep neural networks based on these 500,000 high-quality references in real-time.

This is the ultimate metaphor for AI empowering human productivity: human energy, focus, and physical strength are extremely precious and limited. If we have to hard-press every detail, the output will inevitably be extremely limited; but if we leverage AI's generalization and inference capabilities, humans only need to focus all their energy on the most core and valuable "500,000 pixels," and AI can help us fill in the rest, allowing humanity to unleash superhuman creativity.

At the same time, we delivered the world's first DGX supercomputer to OpenAI in 2016, which was personally received by Elon Musk. At that time, the supercomputer was massive and priced at $250,000, only a few top laboratories could afford it; now, we have shrunk it down to about $3,000 for a personal AI super workstation that can sit right next to a student or engineer's desk. This means that every ordinary student and independent developer now has their own dedicated super computing center, capable of training their own AI models and R2-D2 robots in their dorm rooms.

Advice for Young People and Life Legacy: The Art of Prompting and Extraordinary Impact of the Era

Cleo Abram: For the students, young generation, and anyone watching this interview who hopes to make a difference in this era, what crucial advice do you think I haven't asked yet?

Jensen Huang: I believe one of the most important pieces of advice is: If I were a student now, the first thing I would do is to fully immerse myself in learning AI.

How do I learn to interact with ChatGPT? How do I learn to interact with Gemini Pro? And how do I learn to interact with Grok? Learning how to interact with AI is essentially about becoming someone who is extremely skilled at asking questions. For example, you are very good at asking questions, and writing prompts for AI is very similar. You can't just randomly ask questions; to make AI your effective assistant, you need to master the skills and art of prompting.

So if I were a student today, whether studying math, science, chemistry, or biology, regardless of which scientific field I will enter in the future or what profession I will pursue, I would ask myself: How can I use AI to do my work better?

If I want to be a lawyer, how can I use AI to become a better lawyer? If I want to be a doctor, how can I use AI to become a better doctor? If I want to be a chemist, how can I use AI to become a better chemist? If I want to be a biologist, how can I use AI to become a better biologist? This question should run through everyone's thinking.

Just like the era I grew up in, we are the first generation that had to face this question: how to use computers to do our work better. The generation before us had no computers at all; remember, when I entered the industry, it was even before Windows 95, back in 1984, and there wasn't a single computer in the office. Soon after, computers became widespread, and we had to ask ourselves how to leverage them to do our jobs well. The next generation no longer needs to ask this question, but clearly, they must confront the next proposition: how can I use AI to do my work better? This is the starting point and the end point of everything.

AI has fundamentally lowered the thresholds of understanding, knowledge, and access to intelligence. In the past, placing a computer in front of someone who had never used one would mean they could never learn to use it in a day; they needed someone to teach them hands-on. However, when faced with ChatGPT, if you don’t know how to use it, you just need to input: "I don’t know how to use ChatGPT, please teach me," and it will immediately respond and provide examples. The most wonderful aspect of intelligence is that it will assist you all the way, making you stronger in the process.

Cleo Abram: Okay, if you have time, I have one last question. This isn't what I originally planned to ask, but on my way here, I actually had a bit of a fear of flying, which is not very rational. The flight was a bit bumpy, very bumpy. Sitting in the violently shaking cabin, I couldn't help but think about what people would say at my funeral.

Jensen Huang: "She always asked great questions!" That will surely be written on your tombstone.

Cleo Abram: I hope so! Yes. Besides "deeply loved her husband, friends, and family," I hope people will talk about my "optimism," and I hope everyone can understand the goals I strive for. I am very curious about you; you have been deeply involved in this industry for so many years, and the vision you describe is so grand. What theme would you like people to use to evaluate the career you have worked hard to achieve?

Jensen Huang: Very simple: "They made an extraordinary impact."

I believe we are fortunate because we established some core beliefs long ago and have consistently adhered to them, continuously building upon that foundation, which has made us one of the most important and influential technology companies in the world, perhaps even in history.

Therefore, we take this responsibility very seriously. We are fully committed to ensuring that the capabilities we create are not only available for large enterprises but also open to independent researchers and developers in various scientific fields, regardless of profitability, size, or recognition. It is precisely because we deeply understand the significant meaning of the work we do and its potential profound impact on countless people that we hope to disseminate this capability to every corner as much as possible.

I firmly believe that when we look back in a few years, I sincerely hope the next generation will see: first, they will recognize us for the gaming technology we built; but I am confident that looking back, the fields of digital biology and life sciences will have been completely reshaped, our understanding of materials science will have been utterly transformed, and robots will be assisting humans in various fields to complete dangerous and tedious tasks. When we want to drive, we can drive ourselves, and when we don’t want to, we can take a moment to relax or turn the car into a home theater to read peacefully during our commute. By then, you might even wish you lived farther away so you could enjoy more time in the car.

When you look back and find that there is a company at the center of all these transformations, and it just happens to be the company that accompanied you while you grew up playing games. I hope this is what the next generation recognizes as NVIDIA.

Cleo Abram: Thank you very much for your valuable time; this has been a very pleasant conversation.

Jensen Huang: Thank you, I enjoyed it as well.