Podcast Notes | Jensen Huang: Young People Should Learn to Ask Questions to AI, NVIDIA Hopes to Leave Behind More Than Just Games and Chips
Curated & Compiled: TechFlow
TechFlow Summary: In this 63-minute deep-dive interview, renowned tech creator Cleo Abram bypassed the standard earnings figures and stock market hype, engaging NVIDIA Founder & CEO Jensen Huang in a fundamental conversation about the evolution of underlying computing and the future of human technology. From breaking through 30 years ago with gaming GPUs, to weathering Wall Street's over-a-decade of skepticism while building CUDA, to keenly identifying AlexNet in 2012 which ignited the deep learning revolution, Jensen Huang dissects in depth how NVIDIA systematically reconstructed the modern computing stack that had persisted for 65 years.
Standing at the critical threshold of transitioning from Generative AI to Physical AI, Jensen Huang elaborated in detail on the "World Model + Physics Simulation" foundation built by Cosmos and Omniverse, sketching a ten-year vision of "everything taking a robotic form" and every individual owning a lifelong companion R2-D2 robot. Simultaneously, he firmly addressed the philosophical divide on specialized versus general-purpose chip architectures, analyzed forward-looking bets on digital biology and climate prediction, and offered core advice to the younger generation: embrace the art of asking questions. When asked how he hopes to be remembered, he gave the most unadorned yet grand answer: "Very simply: they made an extraordinary impact."
Guest: Jensen Huang (Founder & CEO, NVIDIA)
Host: Cleo Abram (Renowned tech video creator, host of Huge If True)
Podcast Source: Huge If True (https://www.youtube.com/watch?v=7ARBJQn6QkM)
Original Title: NVIDIA CEO Jensen Huang's Vision for the Future
Air Date: January 27, 2025
Key Takeaways
The perfect computer must balance both serial and parallel computing: In most complex programs, roughly 10% of core code accounts for 99% of the computational load. This portion is highly suited for massive parallel acceleration by GPUs, while the remaining 90% of control logic is efficiently handled by CPUs in serial processing. NVIDIA was founded precisely to solve the ultra-high-load problems that traditional computers could not conquer.
The gaming industry is the best financial engine for foundational R&D: Real-time rendering of 3D virtual worlds naturally requires parallel computing. More importantly, the electronic gaming industry possesses an extremely large consumer market size, providing continuous self-sustaining funding for tens of billions in underlying hardware and software development without relying on government grants or academic funding.
The essence of CUDA is democratizing parallel computing: Early researchers had to disguise physics or math problems as graphics shaders to utilize GPUs. NVIDIA created CUDA, allowing developers to directly leverage GPU computing power using the widely-known C language. Although this bold bet faced over a decade of cold reception and skepticism from Wall Street due to high R&D costs, it ultimately paved the way for the subsequent explosion of deep learning.
AlexNet prompted NVIDIA to completely rebuild its foundational computing stack: Created in 2012 by Ilya Sutskever, Alex Krizhevsky, and Geoffrey Hinton, AlexNet used just two GeForce GTX 580 graphics cards to crush traditional image classification algorithms. Jensen Huang instantly recognized that software was shifting from "manually written rules" to "data-driven autonomous learning." He subsequently led NVIDIA to bet everything on AI, comprehensively restructuring chips, interconnects, algorithm libraries, and system architecture.
The core of Generative AI lies in becoming a universal cross-modal translator: What makes this wave of AI unprecedented is that architectures like Transformer can understand and mutually translate any modality of data (text, images, voice, 3D space, robotic arm trajectories), thereby shattering the barriers between human knowledge and the digital world.
Physical AI must be driven by dual wheels of "World Models + Physics Simulation": Large language models suffer from hallucinations and require retrieval and context anchoring to ground true answers; in the physical world, robots must possess common sense of physics such as gravity, inertia, friction, and spatial constancy. NVIDIA launched Cosmos as the foundational model for the physical world and leverages the Newtonian-physics-compliant Omniverse simulation engine to generate realistic synthetic data, allowing robots to undergo millions of extreme-condition simulations before entering the physical world.
In the next decade, everything will take a robotic form: Whether autonomous vehicles, smart buildings, lawn mowers, or humanoid robots, all future mobile objects will possess robotic capabilities. Everyone will have a dedicated lifelong digital companion (R2-D2), continuously evolving alongside the user across forms like glasses, phones, computers, cars, and home appliances.
Rejecting the proprietary ASIC trap and upholding flexible innovation in computing: Regarding views on hardwiring Transformers entirely into specialized chips (ASICs), Jensen Huang believes innovation will never stop at a single algorithmic architecture. Fixing a computer to a single purpose is akin to turning it into a microwave oven. NVIDIA consistently builds flexible, programmable accelerated architectures, supporting academia in exploring more efficient novel attention mechanisms like FlashAttention and hierarchical attention.
From the company gamers grew up with to the epicenter of cutting-edge scientific breakthroughs: In the RTX 50 series, out of 8 million pixels in 4K resolution, only 500,000 are computed via traditional ray tracing pipelines, while the remaining 7.5 million are accurately predicted and filled by AI. This paradigm frees scientists and engineers to focus their precious energy on core key points, amplified by AI into superhuman productivity. From digital biology and protein language parsing to microclimate forecasting and future robotics, NVIDIA's ultimate historical vision is "to make an extraordinary impact."
Highlight Quotes
On technological faith and forward-looking positioning: "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 a chance to come. Our strategy is built on hope, but it is a rationally deliberated hope."
On the division of labor between serial and parallel computing: "Perhaps only 10% of the code handles 99% of the computational processing, and this 99% task can be done in parallel; but the remaining 90% of code still must execute sequentially. A truly perfect computer isn't either-or; it simultaneously possesses exceptional serial and parallel processing capabilities."
On the driving force behind CUDA's creation: "Many great innovations are born from the same complex stew: part from lofty aspirations and inspiration, and part purely from desperate self-rescue when cornered. On one hand, we saw researchers at Massachusetts General Hospital in Boston rebuilding CT tomography images using our gaming GPUs. On the other hand, we struggled immensely rendering flowing water, explosions, and particle collisions in our own games. These internal and external factors intertwined to spark the birth of CUDA."
On the debate over architectural specialization: "Will the Transformer be the last AI algorithmic architecture human researchers discover? Absolutely not. Looking back at computer science history, no single algorithm has ever ruled forever. If you had completely solidified the algorithms of ten years ago into chip hardware, computers would have become microwave ovens, and all subsequent innovation would have been impossible."
On the essential relationship between AI and human work: "In the roughly 8 million pixels of a 4K display, we actually only physically ray-traced 500,000 of them. The remaining 7.5 million pixels were perfectly filled by AI predictions. This is a vivid metaphor for how AI will make humans seem like superheroes: humans can focus their precious energy and attention on the most critical core parts, handing off the rest of the cumbersome workflows to AI, achieving a hundredfold leap in productivity."
On survival guides for the younger generation: "If I were a student today, the first thing I’d do is learn AI. Learning to interact with AI is like becoming exceptionally good at asking questions. You can’t just randomly ask anything; having AI as your assistant requires mastering the technique and art of prompting. Whether you study law, medicine, chemistry, or biology, you must constantly ask yourself: How can I use AI to do my job better?"
On life's endpoint and historical legacy: "Very simply: They made an extraordinary impact. I hope the next generation knows not just that NVIDIA is the company gamers grew up with, but sees the entire storm center of digital biology, materials science, robotics, and autonomous driving变革, and standing at its forefront is this very company that upheld its core beliefs."
Transcript
The Birth of Heterogeneous Computing: Why Does 10% of Code Need a Dedicated GPU?
Cleo Abram: For today’s conversation, I want to skip the quarterly earnings reports and stock price figures we often see in financial news, and focus on more fundamental questions: What exactly are you building? How did computers evolve step by step over the past few decades to reach today? And how will future technologies shape the lives of ordinary people?
Jensen Huang: Perfect. I’m really looking forward to this kind of exchange. In fact, at some point, you have to choose to believe in something. What we’re doing right now is, at its core, completely reconstructing the modern computing architecture that has persisted for decades.
Cleo Abram: To understand the globally瞩目 AI era we inhabit today, we must first return to the 1990s. Game developers wanted to create more realistic, immersive 3D visuals, but the traditional hardware of the time couldn't handle real-time computation for massive pixel counts. Why couldn't computers of that era handle graphics? And how was the GPU born?
Jensen Huang: It stems from a core characteristic of computer program architecture: within a piece of software code, typically only 10% of critical instructions account for a staggering 99% of the computation. This 99% of intensive calculation is usually highly repetitive, making it ideal for splitting into thousands of tiny tasks executed in parallel; the remaining 90% of code consists mainly of system control, branching, and scheduling logic, which must be processed step-by-step in strict sequence.
Traditional computers of that era were completely dominated by Central Processing Units (CPUs). CPUs were designed to be exceptionally fast at handling serial logic, but they were powerless when facing the massive parallel demands of rendering countless pixels simultaneously. We keenly observed then: a truly ideal computer cannot rely solely on serial processing, nor solely on parallel processing. It must possess both simultaneously.
Thus, we founded NVIDIA with the mission to create dedicated accelerators capable of solving problems that traditional computers could not handle. This is the origin point of the Graphics Processing Unit (GPU).
Cleo Abram: This reminds me of a classic educational video released by NVIDIA over a decade ago: The Mythbusters team built a robot that shot paintballs one by one from a single nozzle to demonstrate CPU task processing; then they deployed a massive mechanical array with thousands of nozzles firing simultaneously in milliseconds, instantly assembling a complete Mona Lisa portrait, vividly showcasing the parallel acceleration power of GPUs.
Jensen Huang: That demonstration intuitively captured the essential difference between the two. Gaming was the first field to develop an extreme hunger for this kind of large-scale parallel computing power.
Gaming Breakthrough: Fueling Hardcore Foundational R&D Through Consumer Markets
Cleo Abram: Why did NVIDIA initially choose to enter the electronic gaming sector rather than directly targeting scientific computing?
Jensen Huang: First, because we genuinely love that application domain. 3D video games are digital simulations of real or fantasy worlds. Who wouldn't want to step into a realistic, magnificent virtual universe?
But from a commercial strategy perspective, there's a deeper, decisive consideration: We foresaw that electronic gaming would become one of humanity's largest consumer entertainment markets. Developing cutting-edge semiconductor chips requires unimaginably huge R&D capital. If you only serve niche high-performance scientific computing labs, you must rely on government research grants or expensive custom contracts, which simply cannot sustain continuous, massive technological iteration.
Electronic gaming provided an enormously vast consumer market. Hundreds of millions of players willingly paid for smoother, more stunning visuals. This allowed us to become the world's largest manufacturer of parallel processors by shipment volume. Our massive shipment scale gave us a continuous, self-sustaining pipeline of R&D funding, enabling us to push this parallel architecture to its extreme physical limits over the following decades.
Betting Everything on CUDA: From Medical Imaging Inspiration to a Universal Parallel Computing Platform
Cleo Abram: By the early 2000s, some sharp academics and developers began discovering that GPUs weren't just for rendering game textures; they could run physics collision simulations, and even external scientists used them for complex scientific computing. But from my understanding, doing so back then was extremely painful. Scientists had to disguise matrix operations as triangular pixels or texture maps, tricking the graphics card into calculating them. Then you made a hugely bold decision: launching the CUDA platform. Can you share the background behind that?
Jensen Huang: You've done your homework, and that's perfectly accurate. Many great innovations are born from the same complex stew: part comes from lofty aspirations and inspiration, and part comes purely from desperate self-rescue when backed into a corner.
One of the earliest external examples of using our GPUs for general-purpose acceleration came from Massachusetts General Hospital in Boston. Several researchers there used our gaming graphics cards to reconstruct CT tomography images, which gave us massive inspiration.
Simultaneously, NVIDIA hit an internal bottleneck: We wanted the water in our game worlds to flow realistically, explosions to scatter tens of thousands of fragments according to aerodynamics, and smoke to diffuse naturally. This required GPUs, beyond processing graphics, to simultaneously run particle physics and fluid dynamics simulations. If the graphics hardware pipeline remained rigidly locked into traditional computer graphics workflows, none of this would be possible.
So we made the firm decision to completely break free from the constraints of dedicated graphics pipelines, transform them into a universal parallel computing architecture, and created Compute Unified Device Architecture (CUDA). CUDA allowed engineers and scientists worldwide to write code in the familiar C language and freely leverage the tens of thousands of computing cores inside GPUs.
Cleo Abram: Looking back from today's vantage point, building CUDA was a visionary gamble. At the time, NVIDIA put everything on CUDA, forcibly embedding general-purpose computing logic into every shipped gaming graphics card. This increased chip area, drove up costs, compressed gross margins, and resulted in harsh criticism and skepticism from Wall Street for over a decade.
Jensen Huang: When you create a brand-new disruptive technology, critics always say, "You build it, and they might not come." But from an optimist's perspective, our belief is: "But if you don't build it, they will never have a chance to come."
We deeply understood the immense physical potential of this technology. We firmly believed that through the massive consumer GeForce market, we could deploy this advanced parallel computing architecture onto the desktops of geeks and researchers globally. It was this rationally grounded, unwavering hope that carried us through those long, lonely years.
The Epiphany Moment of AlexNet in 2012: Rebuilding the 65-Year-Old Modern Computing System
Cleo Abram: Time then jumps to 2012, witnessing a milestone battle in tech history: the ImageNet object recognition competition. Alex Krizhevsky, Ilya Sutskever, and their mentor Geoffrey Hinton submitted a deep neural network model called AlexNet, crushing all traditional computer vision algorithms and drastically reducing error rates. They trained this epoch-defining model using just two consumer-grade GeForce GTX 580 graphics cards. Inside NVIDIA at that moment, what was your immediate reaction upon seeing these results?
Jensen Huang: It was a moment that completely altered the trajectory of world history. An internal computer vision research team at NVIDIA was struggling immensely, trying to solve image recognition problems with traditional algorithmic libraries. When AlexNet emerged, we were not only shocked but felt a powerful resonance: This new algorithm utterly abandoned the old path of manually writing rules, relying entirely on massive datasets to autonomously train weights, achieving an unprecedented leap in capability.
We immediately stepped back and asked ourselves a fundamental question: "If this deep learning architecture demonstrates such terrifying scalability in computer vision, what is its absolute limit? If it can be scaled infinitely, what does that mean for the entire computer industry and the underlying hardware-software architecture?"
At that time, we made a crucial judgment: Software development paradigms were undergoing a fundamental shift. Since IBM System/360 established the instruction set architecture for modern general-purpose computers in 1964, for a full 65 years, humanity told computers what to do by precisely writing step-by-step instructions. The rise of deep learning marked computers transforming into systems that autonomously learned patterns from massive data.
Since the essence of software had changed, the underlying hardware and entire system stack supporting it had to be completely torn down and rebuilt. We immediately made the decision to go all-in on AI, comprehensively redesigning the entire computing stack from single-chip architecture, interconnect buses, networking communication, and algorithm libraries to distributed systems. The later DGX supercomputer systems originated directly from this epiphany.
Why Now? A Universal Translator Across All Data Modalities
Cleo Abram: Why did the public suddenly feel the explosion of AI only in recent years? What exactly is the underlying technological leap behind this?
Jensen Huang: Because the hardcore challenges in computer science over the past decade were tackled one after another. Initially, computer vision was thoroughly solved, followed by speech recognition reaching superhuman levels, and then natural language understanding achieved qualitative breakthroughs.
But the final kick that triggered a comprehensive phase change was the maturation of the Transformer architecture and Generative AI. Past AI models were mostly single-task classifiers, like judging whether an image contained a cat. Today's Generative AI is a universal "cross-modal translation system."
It can understand text and translate it into another text, like summarizing paragraphs or translating cross-border languages; it can also translate text into images, images into 3D assets, and sensor data into robotic arm rotation angles. It freely shuttles between various representation forms of the digital and physical worlds. That's why today's AI feels so omnipotent.
Physical AI and World Models: Cosmos Imparts Common Sense, Omniverse Anchors Physical Laws
Cleo Abram: Next, I want to dive deeply into the most critical transition happening right now: moving from text and images in the digital world to embodied intelligence and robotics in the physical world. Previously, training robots required placing them in real factories step-by-step, or having humans wear motion capture suits to collect extremely limited data. Robots wore out easily, fell, and broke, making data collection incredibly expensive and slow. Now we're beginning to train robots in virtual worlds. NVIDIA launched Cosmos and Omniverse. How are you solving the robot learning puzzle in virtual environments?
Jensen Huang: The best reference frame for understanding Physical AI is the developmental path of language models like ChatGPT.
When ChatGPT first appeared, its performance was stunning, but it had a fatal flaw: hallucination. If it rambled at length or discussed untrained blind spots, it would confidently speak nonsense because its outputs lacked anchoring to Ground Truth. Later, the industry introduced Retrieval-Augmented Generation (RAG). When you feed the model an authoritative PDF or connect it to real-time web search, that PDF and search result became the model's benchmark facts, giving its reasoning reliable anchor points.
Now we project this exact same logic into the physical world. For robots to move freely in reality, they first need a foundational model equipped with physical common sense. What is physical common sense? It must understand Earth has gravity, objects possess inertia and friction; it must have geometric and spatial perception; it must grasp object constancy, meaning when I turn away and turn back, the cup on the table remains there; it must understand causal logic, meaning if I knock over a cup, water spills 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 realm.
But Cosmos alone isn't enough. We must provide it with an "absolute physical truth benchmark," just like feeding ChatGPT authoritative PDFs. That benchmark is Omniverse. Omniverse is not a standard video game engine; it is a strictly first-principles-compliant physics simulation system whose underlying mathematical models are rigorous solvers for Newtonian mechanics, fluid dynamics, and electromagnetics.
By combining Cosmos's world model with Omniverse's physics simulation engine, we can generate endless, absolutely physics-compliant virtual environments and sudden edge cases for robots in digital space: sudden light changes, slippery roads, occluded obstacles, mechanical failures. Robots can complete extreme training equivalent to decades in the real world within hours, before their models are deployed to physical robots.
Ten-Year Vision: Everything Will Be Robotic, Everyone Will Own Their Own R2-D2
Cleo Abram: Following this logic forward, what will the physical world look like in ten years? How will ordinary people interact with these physical AIs?
Jensen Huang: Everything that moves in the future will possess robotic capabilities, and this day will arrive much faster than most imagine.
In the future, pushing a lawnmower across your yard under the scorching sun will seem absurd by today's standards. Every car will be a fully autonomous robot; humanoid robots will walk into factories and homes worldwide. All these robots will graduate in the digital twin worlds constructed by Omniverse and Cosmos before stepping into physical reality.
What excites me even more is that everyone will own a dedicated lifelong digital companion, just like R2-D2 from Star Wars. It won't necessarily be a round metallic trash can; it might live in your smart glasses, your smartphone, your office computer, sit in your car, and greet you as a physical entity when you arrive home.
This personal R2-D2 accompanies you for life, growing alongside you, deeply understanding your joys, sorrows, habits, and work methods, becoming your strongest extension for perceiving and reshaping the world.
Debate on Architecture Philosophy: Why Can't We Hardwire Transformer Directly into Chips?
Cleo Abram: While preparing for this interview, I had deep discussions with many engineer friends who insisted I ask you this: With the Transformer architecture becoming the foundation for nearly all current large models, many chip companies are attempting to manufacture specialized chips (ASICs) custom-built entirely for Transformer. Specialization naturally brings极致 efficiency on specific tasks but sacrifices general flexibility. How does NVIDIA strike the balance between specialized optimization and general architecture? How do you make multi-billion-dollar tech bets that determine life and death?
Jensen Huang: This precisely touches our most foundational core belief: Do you believe the Transformer is the last AI algorithmic architecture human researchers will ever discover?
If your answer is yes, then of course you can hardwire all the mathematical operations of Transformers directly into silicon with physical circuits. But our core belief is: The Transformer is by no means the end goal. It is merely a stepping stone on the ladder toward higher AI evolution. The frontier architectures of a few years from now will likely bear little resemblance to Transformers.
Looking back at the history of human computer algorithms and innovation, has any brilliant idea ever monopolized the landscape forever? The answer is never. The essence of a computer lies in its ability today to run wondrous programs that anyone could only dream of ten years ago. If you had completely solidified chips into single-purpose hardware ten years ago, computers would have degraded into microwave ovens, and the prosperity of the entire application ecosystem would have halted abruptly.
Taking Transformer's core Attention Mechanism as an example: Its original purpose was to calculate the correlation between each word and all others in the context. When text contains only 10 words, computation is trivial; but when the context window expands to tens of thousands of words or millions of tokens, full-to-full computation costs explode quadratically, making physical hardware implementation impossible. To solve this, novel mathematical algorithms like FlashAttention, Hierarchical Attention, and WaveAttention have emerged over the past few years.
Computer science hasn't ended, AI researchers haven't stopped exploring, and neither will NVIDIA. Our most vital mission is to continuously build highly programmable, flexible accelerated computing systems, allowing the world's smartest algorithm scientists to swim freely in this ocean of compute, discovering the next generation of humanity's great unknown algorithms.
Chip Manufacturing Limits and Physical Challenges: From Wafer Packaging to Extreme Energy Efficiency
Cleo Abram: Top-tier chips today are approaching the fundamental limits of physics. Your teams design the chips, and partners like TSMC photolithograph and manufacture them at nanometer scales. As Moore's Law slows and power bottlenecks arise, how does NVIDIA approach design thinking?
Jensen Huang: Traditional Moore's Law and Dennard Scaling have essentially reached physical termination. The past era of relying solely on shrinking transistors for free multiplications in compute power and reductions in power consumption is gone forever.
That is precisely why we must shift comprehensively from "chip thinking" to "data center thinking." The fixed optical limit of a single reticle size restricts wafer scaling. To break through single-chip physical limits, we employ Chip-on-Wafer-on-Substrate (CoWoS) packaging integration, Ultra-High-Bandwidth Inter-Chip Links (NVLink), and extremely dense High Bandwidth Memory (HBM).
More importantly, parallel accelerated computing follows entirely different physical laws. Delivering a complex task to tens of thousands of efficient cores for parallel execution improves energy utilization efficiency by dozens or even hundreds of times compared to forcing a few high-clock-speed general CPU cores to brute-force it. Through synergistic optimization across algorithms, chips, networking, and system levels, we have pioneered a new path for sustaining exponential compute growth in the post-Moore era.
The Next Grand Bet: Physical AI, Digital Biology, and Microclimate Simulation
Cleo Abram: Looking back, NVIDIA's heavy bets on GPUs, CUDA, deep learning, and autonomous driving have almost all been validated by time. Besides the robotics discussed earlier, what major bets are coming up next?
Jensen Huang: Beyond Physical AI and the hardware/software toolchains for humanoid robots, a field I am intensely excited about is Digital Biology.
Past computers understood artificial code composed of 0s and 1s; modern AI understands human natural language. Biology's essence, however, 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 use AI to deeply parse and predict biological macromolecular interactions, we can construct a digital twin of the human body. This transforms drug discovery, gene therapy, and disease mechanism research from wet-lab trial-and-error into high-precision sciences rapidly simulated and iterated in computers.
Another source of immense pride for me is climate science. Leveraging digital twins, we are constructing ultra-high-resolution regional climate and extreme weather forecasting systems capable of precisely simulating micro-meteorological evolution within a kilometer directly above your head.
Scientific tools are fundamentally "time machines" invented by humanity. Scientists strive desperately to predict the future precisely to see evolutionary trajectories ahead of time, enabling optimal design and intervention today to create the best possible future version.
Neural Rendering and Personal Supercomputers: Productivity Leaps with RTX 50 and DGX
Cleo Abram: At the recent consumer electronics show, you unveiled the latest GeForce RTX 50 series graphics cards and the new desktop-generation DGX systems. What do these new hardware pieces actually mean for ordinary creators and developers?
Jensen Huang: Behind this lies a fascinating technical cycle: Originally, GeForce gaming GPUs孕育ed modern AI, helping Alex Krizhevsky and Geoffrey Hinton train AlexNet; today, powerful, mature AI feeds back into GeForce cards, completely subverting the underlying logic of graphical rendering.
In the 4K resolution imagery presented by the latest RTX 50 series (totaling roughly 8 million pixels), our physical ray tracing pipeline actually performs intensive calculations on only 500,000 pixels. The remaining 7.5 million pixels are all inferred in real-time and perfectly filled by deep neural networks based on those 500,000 high-quality benchmarks.
This is the ultimate metaphor for AI empowering human productivity: Human energy, focus, and stamina are extremely precious and limited. If we brute-force every single step, output is inevitably constrained. But by leveraging AI's generalization and inference capabilities, humans need only pour all their energy into the core "500,000 pixels," while AI completes the rest, unleashing superhuman creative potential collectively.
Simultaneously, in 2016, we personally delivered the world's first DGX supercomputer to OpenAI, with Elon Musk accepting it himself. Back then, supercomputers were massive, cost $250,000, and were exclusive to top-tier labs. Now, we've shrunk it into a roughly $3,000 personal AI workstation that sits directly beside students or engineers. This means every ordinary student or independent developer owns a dedicated supercomputing center, capable of training their own AI models and R2-D2 robots right in their dorm rooms.
Advice for the Youth and Life Legacy: The Art of Prompting and Impactful Era Influence
Cleo Abram: For students watching this interview, the younger generation, and everyone hoping to achieve something in this era, what crucial advice do you think is vital that I haven't asked?
Jensen Huang: One of the most important pieces of advice is: If I were a student today, the first thing I'd do is全力 learn AI.
How to learn to interact with ChatGPT? With Gemini Pro? With Grok? Learning to interact with AI is essentially becoming exceptionally good at asking questions. You're exceptionally good at asking questions, and writing prompts (Prompting) for AI is highly similar. You can't just randomly ask anything. Having AI as your capable assistant requires mastering the techniques and art of prompting.
So if I were a student today, whether studying mathematics, science, chemistry, or biology, regardless of what scientific field I enter or career I pursue, I would ask myself: How can I use AI to do my job better?
If I want to be a lawyer, how do I use AI to become a better lawyer? If I want to be a doctor, how do I use AI to become a better doctor? If I want to be a chemist, how do I use AI to become a better chemist? If I want to be a biologist, how do I use AI to become a better biologist? This question should permeate everyone's thinking.
Like the era when I grew up, we were the first generation forced to face this exact question: How do we use computers to do our jobs better? The generation before us didn't even have computers. Remember, when I entered the industry before Windows 95, back in 1984, there wasn't a single computer in the office. Shortly after, computers普及了, so we had to ask ourselves how to use them well. The next generation doesn't need to ask that question anymore, but obviously, they must face the next proposition: How do I use AI to do my job better? That is the starting point and the ending point of everything.
AI has radically lowered the thresholds for comprehension, knowledge, and accessing intelligence. In the past, putting a computer in front of someone who had never used it meant they absolutely couldn't learn to operate it in a day; they needed hands-on teaching. Facing ChatGPT, if you don't know how to use it, you simply input: "I don't know how to use ChatGPT, teach me," and it immediately responds with guidance and examples. The most beautiful aspect of intelligence is that it assists you along the way, making you stronger throughout the process.
Cleo Abram: Great. If you still have time, I have one final question. This wasn't originally planned, but on the way here, I actually experienced some fear of flying. It's admittedly irrational. The flight was quite turbulent, very bumpy. Sitting in that violently shaking cabin, I couldn't help but wonder what people would say at my funeral.
Jensen Huang: "She always asked fantastic questions!" That will definitely be engraved on your tombstone.
Cleo Abram: I hope so! Yes. Besides being "deeply loved by her husband, friends, and family," I hope people discuss my "optimism," and that everyone understands the goals I tirelessly pursued. I'm incredibly curious about you. You've深耕 this industry for so many years, painting a vision of the future so grand. What theme do you hope people use to evaluate the career you've worked so hard to achieve?
Jensen Huang: Very simply: "They made an extraordinary impact."
I believe we are incredibly lucky because we established certain core beliefs long ago, continually built upon them, and steadfastly held fast, allowing us to become one of the world's most significant, perhaps historically most influential, tech companies today.
Therefore, we take this responsibility extremely seriously. We ensure全力 that the capabilities we create are not only accessible to large enterprises but also open to independent researchers and developers across all scientific fields, regardless of profitability, scale, or fame. Precisely because we deeply understand the monumental significance of our work and its profound potential impact on countless lives, we strive to distribute these capabilities to every conceivable corner.
I firmly believe that when we look back in a few years, I deeply hope the next generation will see: First, they will recognize us for the game technologies we built. But I am certain that looking back, digital biology and life sciences will have been completely reshaped, our understanding of materials science overturned, and robots assisting humans in dangerous and tedious tasks across every domain. Want to drive? Drive yourself. Don't want to? Rest briefly, or turn your car into a home theater, calmly reading during your commute. By then, you might even wish you lived further away so you could enjoy those moments longer inside your vehicle.
When you look back and discover that a company stood almost at the epicenter of all these transformations, and coincidentally, it was exactly the company gamers grew up with. I hope that is how the next generation perceives NVIDIA.
Cleo Abram: Thank you so much for your valuable time. This has been a truly enjoyable exchange.
Jensen Huang: Thank you. I'm glad too.