Las Vegas, January 7 — Today, following the highly anticipated opening keynote at CES, Nvidia founder and CEO Jensen Huang participated in an in-depth interview with global media, including Zhidx, for nearly an hour. During the interview, he praised Chinese car companies, responded to the high pricing of the RTX 5090, admitted that AI PC talent is insufficient and demand has yet to explode, proposed that the three major Scaling Laws will coexist, and provided a more detailed analysis of Nvidia’s new releases and strategic moves from yesterday.
Huang also changed out of the “ray-tracing” leather jacket he wore yesterday, stating that the jacket might have been a bit too much for today’s occasion.
Huang stated that the influence of companies like Ideal, Xpeng, Xiaomi, NIO, and BYD on the smart driving trend may be even greater than Tesla’s.
When asked about the trade war, he expressed confidence that the government would make a clear decision, and regardless of the outcome, Nvidia would do its best to help customers and the market.
The flagship RTX 5090 released by Nvidia this time, along with its sub-flagship model, has a significant difference in price and performance. Huang explained that this choice was to meet the needs of extreme enthusiasts, while the lower memory configuration of the entry-level product was meant to balance with its computing power. A reporter tried to extract details about the 60 series, but Huang refused to comment.
The discussion about the end and changes in the Scaling Law continues, with Huang believing that in the near future, the three major Scaling Laws will coexist. These include the improvement of computing power, feedback of data during the testing and inference process, and his belief that there are no physical bottlenecks in AI development.
Although Nvidia’s market value has surpassed the $3 trillion mark, Huang demonstrated an extremely restrained, streamlined, and focused development strategy. Nvidia will not delve into infrastructure such as power generation, and there are no new acquisition plans at the moment. It also does not intend to get heavily involved in the computing layer or the library layer, nor does it plan to serve enterprises directly. Instead, it will focus on products that only Nvidia can deliver, with the latest examples being the Digits personal supercomputer and NVLink 72.
Despite constant releases in the AI PC field, sales have not yet fully taken off. Huang believes that the AI PC demand from designers, developers, and other groups is real. They will use WSL2 as a virtual platform to bring cloud-based AI technology to PCs.
Physical AI is one of the key trends in Huang’s view, as it will help AI understand the world. However, there is still a lack of a foundational model for the physical world. Nvidia hopes to generate a GPT-3-level physical model through Cosmos, which will be of great significance to robotics and autonomous driving.
Agent technology is a focus for companies like AWS and Microsoft, and Nvidia has also launched an enterprise-focused agent blueprint. However, Huang stated that Nvidia would not directly compete with these companies but would primarily provide the foundation and support for enterprise agent development.
Below is a complete translation of Huang’s interview (for readability, some modifications were made to improve clarity without altering the original meaning).
1. Jensen Huang: Nvidia is still a small company and has no new acquisition plans
- Reporter: First of all, congratulations. It has been a fantastic year. You are still leading the industry by at least two years. Last year, you defined a new era for computing, specifically data centers. I completely agree. You had to start by building, and now you have completed the entire system, including GB200 and NVL72. Is it time for Nvidia to start considering infrastructure, power, and other parts of the system?
Jensen Huang: Nvidia’s rule is to only do things that others haven’t done, or things we can do much better. So the bar for what we start doing is actually very high. That’s why we don’t consider ourselves an old, established company. The reason we do what we do is because if we don’t build NVL72, who will? Who has the capability to build it? If we don’t build switches like Spectrum X, who can do it?
Despite having 32,000 employees, we are still a relatively small company. We are still a small company, and we need to ensure our resources are highly focused on areas where we can make a unique contribution.
- Reporter: Last year at the GTC conference, you shared about increasing investments and activities in Israel, solidifying your position as one of the country’s largest employers. You continued to expand in 2024 and acquired two Israeli companies. How do you plan to further increase your investment in Israel? Specifically, will we hear about a new ongoing deal soon?
Jensen Huang: We are attracting top talent globally. We’ve received over one million job applications, and our team in Israel has grown to nearly 5,000 people. We’ve become one of the fastest-growing corporate employers in Israel, and I’m very proud of that.
Even under the current tense situation, our team in Israel has delivered exceptional results. They have remained focused and produced outstanding outcomes. In the last seven weeks alone, the team successfully developed key products like Spectrum X and Bluefield 3. I am incredibly proud of their professionalism and dedication.
As for new acquisition plans, there is no news to announce today.
2. AI PC demand does exist, and it will bring cloud-level AI capabilities
- Reporter: You released a lot of information about AI PCs last night. Sales of these products haven’t really taken off this year. Do you think Nvidia can help change this situation and drive adoption of these products? What is hindering their development?
Jensen Huang: That’s a great question. AI technology initially developed in cloud environments. If you look at Nvidia’s growth trajectory over the past few years, you’ll see it’s been primarily focused on cloud computing. This is because training AI models requires powerful supercomputers. These models are massive, and deploying them in the cloud and calling them via APIs is relatively easy.
We believe there are still designers, software engineers, creative professionals, and hobbyists who prefer to use PCs for developing these things. Of course, one challenge is that AI has primarily been in the cloud, and a lot of effort and engineering resources have been invested in the cloud, so there are fewer developers working on AI PCs.
However, we have noticed a large number of designers, software engineers, creative professionals, and tech enthusiasts who prefer to do AI development on PCs. One challenge is that because the AI ecosystem is mainly concentrated in the cloud, and cloud service providers have invested so many resources into its development, there are relatively fewer talents focusing on edge-side AI development.
But the good news is that Windows PCs are fully capable of supporting AI development. Especially with WSL2 (Windows Subsystem for Linux 2), this powerful virtualization platform, we can bring cloud-based AI technologies to PCs. We are working hard to ensure that PC hardware and the WSL2 platform can perfectly support these technologies.
I am confident that this is the right direction for the future. I’m personally very passionate about it, and all the PC manufacturers are showing great interest. We’re working with partners to make sure that every PC has the capability to run AI, including optimizing the Windows system, so we can bring the entire cloud-based AI ecosystem to personal computers.
3. A foundational model for the physical world is needed, and Cosmos will help AI understand the world
- Reporter: Some parts of your keynote at the CES opening last night felt like they belonged in a SIGGRAPH conference, very technical. Your audience is much broader now, so could you explain the significance of these concepts?
Jensen Huang: First of all, I admit that my explanation wasn’t as clear as it could have been, but that’s not the main issue. You have to understand that Nvidia is fundamentally a technology company, not a consumer products company. Our technology will influence and shape the future of consumer electronics. Although we’ve been warmly invited, we remain a company focused on technology. Of course, that’s not an excuse for not explaining the technology better.
One of the most important breakthroughs we announced yesterday is the development of a foundational model that can understand the physical world. Just like GPT is a foundational model for understanding language, and Stable Diffusion is a foundational model for understanding images, we’ve now created a foundational model that understands the physical world.
It can understand concepts like friction, inertia, gravity, the existence and persistence of objects, and geometric and spatial relationships. These are all things that children intuitively grasp; children’s understanding of the physical world is something current language models cannot match.
We firmly believe that a foundational model capable of understanding the physical world is essential. Now that we have this model, anything you can do with GPT or Stable Diffusion can be done with Cosmos. For example, you can have a conversation with it.
You can talk to this world model and ask it about the current environment. Based on information it gathers from cameras, it can understand and describe scenes like “there are many people sitting around a table, and they are in a room.”
Cosmos is essentially a model that understands the real world. So why do we need such a model? Because if you want AI to naturally operate and interact in the real world, it must possess this understanding.
Where will it be applied? Autonomous vehicles need to understand the physical world, robots need to understand the physical world. This foundational model is the key starting point for making these applications possible. Cosmos will make robots a reality.
4. The three major Scaling Laws will coexist, and test-time computation will form a positive feedback loop
- Reporter: Last night, you mentioned that we are witnessing the emergence of some new Scaling Laws, especially test-time computation. I believe OpenAI’s O3 model suggests that from a computational perspective, the cost of scaling inference is very high. What has Nvidia done to provide more cost-effective AI inference chips? From a broader perspective, how do you benefit from the scaling of test-time computation?
Jensen Huang: First of all, for test-time computation, the most direct solution in terms of both performance and cost-effectiveness is to improve our computational capabilities. The inference performance of the new architecture may be 30 to 40 times higher than the previous generation, Hopper. By increasing performance by 30 to 40 times, we can reduce costs by 30 to 40 times because the other costs in data centers remain essentially unchanged.
So, the best improvement method—and the reason why Moore’s Law has been so important in computing history—is that it lowers the cost of computation. The reason I mention that our GPU performance has increased 1000 to 10,000 times in the past 10 years is because that, in turn, means costs have decreased by 1000 to 10,000 times. Over the past 20 years, we have reduced the marginal cost of computation by a million times. This is what has made machine learning so practical. You only need to get the computer from here to there.
So, the best way to improve is by reducing the cost of computation, which is why Moore’s Law has been so important in the history of computing. I mentioned that our GPU performance has increased 1000 to 10,000 times in the past 10 years, and that also means costs have decreased by the same magnitude.
In the past 20 years, we’ve reduced the marginal cost of computation by a million times. This drastic reduction in cost has made machine learning feasible. In the future, we will continue to drive the improvement of computational power, which will continue to lower computing costs.
Looking at this from another angle, now, by expanding test-time computation to get answers, multiple attempts are required. These answers are then used as data for the next round of training. This data will become pre-training data. All the data we collect will be used in both pre-training and post-training data pools. By continually improving the performance of supercomputers and reducing training costs, we can ultimately lower the cost of AI inference for everyone.
This process takes time. These three Scaling Laws will continue to play a role in the future and will interact with each other. From a technical perspective, with every iteration, we are improving the intelligence of the model. At the same time, users’ demands for AI continue to rise, and they are posing more challenging questions, which requires the system to possess greater intelligence. This growing demand will lead to the continued development of test-time scaling, forming a positive and ever-expanding feedback loop.
5. DLSS 4 is about “predicting the future,” but rendering remains an important foundation
- Reporter: Regarding DLSS 4, you gave a demonstration last night. Could you elaborate on it, such as the multi-frame generation technology? Is it still about rendering two frames and generating/interpolating in between? Also, there was mention of text re-compression in your materials and videos. Is this a technology that game developers need to adopt specifically, or is it a driver-level function that benefits all games, or at least most PC games?
Jensen Huang: In Blackwell, we’ve enabled the shader processors to run neural networks, so we can mix code and neural networks in the shader pipeline. This is critical because textures and materials are processed in shaders.
If shaders can’t run AI, they can’t take advantage of the algorithmic advancements brought by neural networks, like compression technology. Texture compression is now far better than the algorithms used in the last 30 years. The compression ratio has significantly improved, with many textures being compressible by an additional factor of 5. Given the large size of modern games, this is a big leap forward.
The second point is about materials. Materials determine how light interacts with surfaces, and their anisotropic properties cause light to reflect in specific ways, such as creating effects like gold leaf or pure gold. These properties occur at the atomic level. It’s difficult to describe this process mathematically, but we can use AI to learn it. This neural material technology will be a major breakthrough, making computer graphics more vivid and realistic.
Regarding DLSS, frame generation is not interpolation; it is literally predicting the future. Yes, you are predicting the future. You’re not interpolating the past; you’re predicting what comes next. The reason for this is that we’re trying to increase the frame rate.
- Reporter: Is AI now playing a more decisive role in PC gaming? Have traditional geometric features that needed rendering been replaced by generation?
Jensen Huang: No, let me explain why. When ChatGPT was first released, many people thought it could generate content directly, but those of us on the inside knew this wasn’t realistic. The system needs real foundational data as a condition. Just like we use context to guide a conversation or prompt in ChatGPT, it’s important to understand the background before answering questions. This background could be a PDF, search results, or clear prompts.
The same applies to games. You need to provide context, considering not only the storyline but also spatial relationships. By providing initial geometric shapes or textures as conditions, the system can generate new content or enhance resolution. This method is the same as how ChatGPT uses context. In enterprise applications, we call it gradient retrieval-enhanced generation. Future 3D graphics will be based on this 3D-conditioned generation.
Take DLSS 4 as an example. Four frames of images have a total of 33 million pixels, but we only rendered 2 million of them. The magic is that we generated the remaining 31 million pixels from those 2 million. The key is that these 2 million pixels must be accurately located. With this foundation, we can generate the remaining 31 million pixels. What’s even more important is that because we saved on computational resources, these 2 million pixels can be rendered in great detail, providing high-quality reference for generating the other pixels.
This transformation will impact all aspects of gaming, from pixels to geometry and animation. This process took six years. When DLSS was first announced, many didn’t believe it, partly because I didn’t explain it well. But now everyone recognizes that generative AI is the future. The key is to have conditional settings and creative guidance from artists.
The relationship between Omniverse and Cosmos is the same. Omniverse, as the 3D engine of Cosmos, is essentially a generative engine. We can precisely control the level of rendering and reduce direct rendering to generate more content. When we reduce control and simulation, we can simulate larger worlds. Behind the scenes, we have a powerful generative engine creating the real world.
6. RTX 5090 Targets Extreme Enthusiasts, 5070 12GB Balances Memory and Computational Power
- Reporter: I have a question regarding gamers and consumers. We noticed a significant gap between the RTX 5090 and 5080. The number of CUDA cores in the 5090 is more than twice that of the 5080, and the price has doubled as well. Why did you create such a large difference between the flagship and the near-flagship models?
Jensen Huang: The reason is simple. Once someone wants the best product, they will always choose the best. You know, the market isn’t so segmented. And our enthusiasts, if they want the best, they won’t settle for something slightly better or save $100 by choosing something a bit worse. They just want the best.
Of course, $2,000 is not a small amount, and it’s certainly a high value. But keep in mind that this technology is being used in your home theater-level PC setup. And that PC—where you’ve probably already invested around $10,000 in your monitor and sound system—will definitely need the best GPU. So many of our customers are simply after the absolute best.
- Reporter: I’m a reporter from South Korea. As a gamer, I’m very excited. When you talked about memory, you specifically mentioned HBM. Why didn’t you choose Samsung’s?
Jensen Huang: I believe Samsung doesn’t really produce products related to graphics cards, right? Do they produce them? Please don’t tell them I said that. Sorry, my mistake. I’m not very clear on that.
There’s no special reason. As you know, SK and Samsung are two of the biggest manufacturers. I’m not exactly sure about the specifics, and it might not be something that needs to be recorded. They’re working hard. Yes, they are working hard and will definitely succeed. I’m confident that Samsung will successfully develop HBM.
I’m confident in Samsung’s success. After all, Samsung was the first to create HBM. The first batches of HBM memory we used came from Samsung. They will rise again. They are a great company. Koreans are very ambitious, which is a good thing. Although they need to redesign, I’m sure they will succeed. They’re moving forward quickly and are deeply invested in it. SK and Samsung are excellent companies, especially in memory, so I believe they will continue to succeed. As you saw yesterday, we use a lot of HBM memory in our products, and HBM is very important to us.
- Reporter: My question is, why does the 5070 still use 12GB of memory? Many games now require larger memory, especially at target resolutions.
Jensen Huang: We’ve always been looking for the optimal balance between the compute engine, computational power, bandwidth, and memory capacity. While it’s difficult to achieve perfection, that’s our goal. Memory and computational power need to match. Too much memory would waste computational resources, and too much computational power would be limited by memory capacity. Finding this balance is a significant challenge for us.
- Reporter: After seeing the performance of the 5070 last night, which is comparable to the 4090, and with the price dropping so much, it’s really exciting. What are your expectations for the 60 series?
Jensen Huang: Last night, we released four RTX Blackwell GPUs, and even the lowest-performing one surpassed the strongest GPUs on the market today. That’s truly incredible. This really demonstrates the power of AI technology. Without innovations like AI, Tensor Cores, and DLSS 4, we couldn’t achieve this level of performance. As for the 60 series, I don’t have anything particularly special to say.
- Reporter: What advantages does Nvidia’s Blackwell architecture have in terms of unification? How does this unified architecture benefit developers and end users?
Jensen Huang: Yes, this is indeed an important advantage. Our Blackwell architecture GPUs are highly versatile and can support all kinds of applications, from agentic AI to complete robotic systems. Whether it’s in cloud servers, autonomous vehicles, robotics, or gaming systems, Nvidia’s architecture remains consistent across all platforms. This was a strategic decision we made after careful consideration.
The core reason for choosing a unified architecture is to provide software developers with a seamless development platform. Developers only need to develop once, and their programs will run across any Nvidia-powered platform. As I mentioned yesterday, we can develop AI models in the cloud and then easily deploy them on personal computers. This capability is unique.
Specifically, AI containers from the cloud can be directly downloaded to run on PCs. For example, models like SD-XL, Flux, and Llama can simply be downloaded from the cloud and deployed on a PC for immediate use. This convenience will be more widely applied in the future.
7. Nvidia Doesn’t Directly Serve Enterprises, Focuses on Computational and Library Layers
- Reporter: While the demand from hyperscale customers for your products is quite evident, I’d like to understand your sense of urgency in expanding revenue streams, particularly in reaching new enterprise clients and the government data center market, as well as in hyperscale computing, especially with companies like Amazon developing their own AI chips. How much pressure are you feeling? Could you provide more details on your progress with enterprise and government customers?
Jensen Huang: Our urgency comes from the real needs of our customers. I’m very pleased to see customers using our technology in the cloud. Our technological progress is happening very quickly. Doubling performance every year means halving the cost. This progress is even faster than the best days of Moore’s Law. We will continue to actively meet the needs of our customers, no matter where they are.
As for the enterprise market, there are mainly two service industries. Our strategy is to collaborate with these two sectors to help them build solutions. NeMo, NIMs, and Blueprints are toolkits for developing Agentic AI. For example, our collaboration with ServiceNow has been very successful, and they are about to launch a series of agent-based services using these technologies. This is our basic strategy. We also collaborate with solution providers like Accenture. Accenture does a great job helping clients adopt these systems. So, the first step is to provide assistance.
- Reporter: Many companies are developing Agentic AI. How do you plan to compete or cooperate with companies like AWS, Microsoft, and ServiceNow?
Jensen Huang: Nvidia is not a company that directly serves enterprises. We are a technology platform company. We are focused on developing toolkits, libraries, and AI models for companies like ServiceNow. This is our focus. We mainly collaborate with companies such as ServiceNow, SAP, Oracle, Synopsys, Cadence, Siemens, and others.
While we have strong advantages in professional fields, we don’t want to delve too deeply into the computational and library layers of AI. Our job is to develop tools for these companies. This is a challenging task because what we are actually doing is integrating ChatGPT into containers. Optimizing these endpoints and microservices is very complex. However, once completed, customers can use our products on any cloud platform.
That’s why we developed NIMs and NeMo. Our goal is not to compete with customers, but to help them. If cloud service providers want to use these technologies, we will fully support them. In fact, many cloud service providers are already using NeMo to train their logical models, and their app stores also feature NIMs.
These technologies are things we developed later on. NIMs and NeMo are as important to our platform as CUDA and CUDAx libraries. The CUDAx libraries were crucial in promoting the NVIDIA platform. We develop these libraries for the industry so that enterprises don’t have to start from scratch.
8. Digits is Primarily Targeted at Data Scientists, Small Models Likely to Appear on AI Glasses
- Reporter: You released Digits yesterday. What do you think is the biggest unmet demand in the current non-gaming PC market?
Jensen Huang: Thank you for your question. Let me first explain Digits. It is a deep learning GPU-based intelligent training system, primarily targeted at data scientists. Most data scientists now work on PCs, Macs, or workstations. For most PC users, running machine learning and data science software like pandas or PyTorch is not an issue.
We recently launched a compact device. It’s small enough to fit on a desk and supports wireless connectivity. It operates much like a cloud service, essentially like a private AI cloud. Why did we develop such a device? Because developers need to frequently access computational resources.
Relying entirely on cloud services can be expensive. Now with this device, it’s like having a private development cloud. This is extremely valuable for data scientists, students, and engineers who need to engage in continuous development.
The future of Digits is very promising. While AI originated in the cloud and may remain cloud-centric in the future, current computational devices are struggling to keep up with the pace of AI development. That’s why we are developing new solutions.
As for superintelligence, it’s not really a new concept. We have experts in every field in our company. I am lucky to work alongside such talented people. I myself am just an ordinary person, but my management team, leaders, and scientists are the best in their respective fields.
This is the future that is coming. You’ll have various AI assistants to help you with tasks like writing, problem analysis, supply chain planning, programming, chip design, and even marketing activities and podcast production. These AI assistants will be on call, ready to help you. Of course, the application of this technology is vast. But remember, it’s humans using tools, and machines are ultimately just tools.
- Reporter: I’d like to ask about the new model products released yesterday, especially the small models. Can these models run on smart glasses? Was this something you considered during development? Based on your direction, it seems like smart glasses could become an important platform for people to experience AI assistants.
Jensen Huang: Yes, I am very interested in smart glasses that can answer questions like “What am I looking at?” or “How do I get there?” They could help with reading, and the possibilities brought by combining them with AI are exciting. I would use Cosmos like this: Cosmos in the cloud handles visual understanding, and if it needs to run on the device itself, a smaller model of Cosmos would be refined for local operation. This way, Cosmos becomes a knowledge transfer tool, transferring knowledge into smaller AI models.
This is feasible because small AI models, while less general, are highly focused in specific areas. That’s why we can perform targeted knowledge transfer. We always start by building the base model and then progressively refine it into smaller models.
9. Chinese Automakers Have a Bigger Impact on the Trend of Autonomous Driving, Which Was Unclear Five Years Ago
- Reporter: I have a question about autonomous vehicles. In 2017, NVIDIA showcased a demo car at CES and collaborated with Toyota at the GTC in May. From 2017 to 2025, what changes have occurred with this technology? What issues existed back then, and what technological breakthroughs have happened now?
Jensen Huang: I believe that all movable devices in the future will have automation capabilities. There will no longer be any regular tasks that require human labor. In 20 years, if we still see someone driving a car home, it would be interesting, but clearly unnecessary. Future cars will still have manual driving options, but they must all be equipped with autonomous driving functionality. Right now, the billion cars on the road cannot drive autonomously, but in 20 years, all of those billion cars will have autonomous driving capabilities. We can still choose to drive, but the trend is already very clear.
Five years ago, we weren’t certain about the stability of this technology, but now we are very confident. The sensors, computing, and software technologies are already quite mature. There is ample evidence to show that the next generation of cars, especially electric vehicles, will almost all come equipped with autonomous driving features.
I believe two companies have driven this change, causing traditional automakers to rethink their stance. One is Tesla, which has had a significant impact, but perhaps an even greater influence has come from China’s remarkable technological advancements. The technologies from emerging Chinese electric vehicle companies, such as BYD, Li Auto, Xpeng, Xiaomi, and NIO, are very impressive. Their autonomous driving technology has already developed very well and is now expanding globally. This sets the standard for future cars—they must have powerful autonomous driving capabilities. I believe this industry has undergone a fundamental transformation.
Both the maturity of the technology and our understanding of it have taken time, but the situation is now very promising. Waymo is an important partner of ours, and their performance in San Francisco has been excellent.
10. Physical AI is a Missing Key Factor; A GPT-3-Level Physics Model Needs to Be Built
- Reporter: The products you launched yesterday are a good start, but one of the major challenges in neural rendering is the various windows in DirectX. What do you need to do to reduce the resistance in engine performance? How do you collaborate with Microsoft?
Jensen Huang: Microsoft has always been very cooperative with our development work. If the API needs to change, they are very accommodating. However, most of our work in DLSS doesn’t require API changes, but mainly changes to the engine. This involves semantic understanding of scenes, which is not just about simple draw calls. The scenes exist mainly in game engines like Unreal Engine and the LithFire engine. That’s why DLSS is now integrated into over a hundred engines.
Especially from DLSS 2/3/4 onwards, once integration is complete, even if a game is developed for 3, it can still benefit from some improvements brought by 4. So, we must build pipelines for AI processing based on scene semantics. This is definitely something that needs to be done.
- Reporter: Major technological changes are never accomplished by a single company. For example, the internet, personal computers, and the recent green technologies. But these technologies eventually converge at some point, bringing about a huge transformation. When it comes to AI, do you think there are still any missing parts that will hinder our development, or is everything ready? I know AI is complex and has many applications, but I want to know what you think is still lacking.
Jensen Huang: Let me explain from two aspects. The first is in language and cognitive AI. We are enhancing AI’s cognitive abilities, enabling it to have multimodal and strong reasoning capabilities. The second is how to apply this technology to AI systems. AI is not a single model; it’s a knowledge system. Agentic AI is about integrating various models—retrieval models, search models, image generation models, reasoning models, planning models, etc. This is a complete knowledge system.
In recent years, the industry has not only been improving foundational AI but also exploring AI applications. However, we are still missing one key element, and that is physical AI. Physical AI needs foundational models, just like cognitive AI. Just like GPT-3 was the first language model to reach a practical level, enabling us to develop various functionalities on top of it.
Therefore, we must bring physical AI to that level as well. This is the reason we are developing Cosmos. Once we reach this level and push the model to the market, we can activate a large number of application scenarios. With foundational models, downstream tasks can be carried out smoothly. This foundational model can also serve as a teacher model, just as we mentioned earlier. This is crucial for Cosmos.
The second missing part is the problem we are solving with Omniverse—combining Omniverse and Cosmos into a physical system. Based on the principles of physics, we can use this foundation to control the generation process, making the content generated by Cosmos more reliable, not just visually realistic. The combination of Cosmos and Omniverse is likely to become a significant breakthrough point.
11. AI Development Has No Physical Limits; Will Strive to Meet Demand Amid Trade War
- Reporter: Regarding the trade war, it’s still quite active. Are you concerned that it might affect everyone’s profit outlook?
Jensen Huang: I’m not worried. I believe the government will make wise decisions in the trade negotiations. What we need to do is help our customers and the market to the best of our ability, regardless of the outcome. We can wait and see how things develop.
- Reporter: How does NVIDIA approach its market strategy? What’s the next step for AI development? Are there any physical limitations?
Jensen Huang: We only act when the market truly needs us, when there’s a gap and we should fill it. We tend to do things that are different from what the existing market is doing, or things that no one else would do if we don’t. That’s our philosophy—don’t repeat what others are already doing. We’re not here to capture market share; we’re here to create new markets.
We won’t enter an existing market just to capture share. That’s not our style. We prefer to pioneer entirely new markets. For example, there is no product like Digits on the market. If we don’t develop it, no one else may, because the software system is too complex. Similarly, if we don’t develop advanced neural graphics, no one else will. These are missions we have to take on.
As for demand, we certainly need better sensors and cameras to address issues. In the coming years, with the proliferation of more advanced cameras, especially smart cameras and smart glasses, we will collect vast amounts of video data, and that’s something I’m really looking forward to.
- Reporter: I’m from Israel. In the past few years, AI, particularly generative AI, has had significant economic growth. I’d like to ask you, for the companies that are part of this wave, do you think their growth speed is sustainable? Can they maintain this momentum in the short term?
Jensen Huang: When it comes to AI’s future development, as far as I know, there are no physical limits. The reason we’ve been able to accelerate AI computing so quickly is that we’ve unified the development of CPUs, GPUs, NVLink, and all the software and systems. If these efforts were scattered across 20 different companies, it would take a lot of time to integrate them. It’s because we control all the integrated technologies and software support that we can accelerate the development of systems so rapidly. From Hopper and H100 to H200 and the next generation, we’ll continue to improve the performance of each unit.
The second point is that by optimizing the entire system, the actual performance gains far exceed the simple transistor performance improvements. While Moore’s Law has slowed, and there’s not much generational improvement in transistor performance, the overall system performance is still increasing significantly year after year. So, I don’t see any obvious physical bottlenecks at the moment.
As computing power improves, several important expansion laws will continue to play a role: first, researchers can train larger models with more data; second, reinforcement learning and synthetic data generation capabilities will continue to improve.
Third, as costs continue to decrease, the range of applications will expand further. As long as there are no fundamental physical limitations, AI will continue to develop rapidly.
12. Satisfied with Cooperation with MediaTek; Possible Expansion of Taiwanese Operations in the Future
- Reporter: I’m from Taiwan. The keynote mentioned that the Digits AI supercomputing CPU is collaborating with MediaTek. Can you talk about the cooperation with MediaTek and other Taiwanese companies such as TSMC? Will NVIDIA establish a headquarters in Taiwan in the future?
Jensen Huang: We have many employees in Taiwan, but our current building is too small. I need to find a solution, and we may announce something at Computex. We are looking for real estate, so if you know of a good place, please make sure to tell me first.
Regarding MediaTek, we have collaborations in multiple areas. For example, in the autonomous driving field, we are working together to provide fully software-defined intelligent automotive solutions for the industry.
Our collaboration in the automotive sector has been very smooth. In addition, the new Grace GB10 CPU was also co-developed with MediaTek. We jointly designed the architecture, achieving interconnectivity between chips and memory consistency between the CPU and GPU.
MediaTek has done an excellent job in both design and manufacturing. They produced a perfect chip on the first attempt, with outstanding performance. Everyone knows MediaTek’s advantages in low power consumption, and they truly live up to their reputation. We’re very happy to collaborate with such an outstanding company.
13. Computer Graphics Entering Exponential Growth Phase; All Industries Will Experience Such Transformations in the Future
- Reporter: I’m from Indonesia, and I’d like to ask for your advice on learning directions for students.
Jensen Huang: This is a great question to wrap things up. Let me first share my personal experience, and then talk about my advice to the new generation of students. My generation was the first to have to learn how to use computers for scientific research. The previous generation still relied on calculators, slide rules, and pen and paper. My generation had to learn to program computers, design chips, and simulate physics. Computers became the tools of our work. The next generation, however, will need to learn how to work with AI because AI is the computer of the new era.
In the future, fields such as biology, forestry, agriculture, chemistry, and quantum physics will all require thinking about how to leverage AI for work. In computer science, the focus will be on how to use AI to advance the development of AI itself. The same applies to fields like supply chain management and operations research.
If someone wants to be a journalist, they need to think about how AI can improve the quality of reporting. If they want to be a writer, they need to think about how AI can enhance writing skills. In the future, every student will need to learn to use AI, just as current students must know how to use computers.
This is the fundamental difference and reflects the profound impact of the AI revolution. It’s not just about large language models—AI will permeate every aspect of the future. This is the most transformative technology, and its development is happening at an astonishing pace.
Thank you all for your attention. It’s great to see the industry evolve from using GPUs to advance AI, to now using AI to advance graphics technology.
Our work in RTX Blackwell, DLSS 4, neural rendering, and neural shading all benefits from AI progress, which is now in turn driving the development of graphics technology. It’s worth noting that computer graphics were originally experiencing slowed growth, but the introduction of AI has brought about supercharged acceleration. Now, we can achieve fully ray-traced images at 200-400 frames per second.
We’ve entered the exponential growth phase of computer graphics, and this applies to almost every field. That’s why I believe it’s not just our industry—very soon, all industries will undergo rapid transformations.
Thank you all, and Happy New Year!
Source: Internet
Disclaimer:
- This channel does not make any representations or warranties regarding the availability, accuracy, timeliness, effectiveness, or completeness of any information posted. It hereby disclaims any liability or consequences arising from the use of the information.
- This channel is non-commercial and non-profit. The re-posted content does not signify endorsement of its views or responsibility for its authenticity. It does not intend to constitute any other guidance. This channel is not liable for any inaccuracies or errors in the re-posted or published information, directly or indirectly.
- Some data, materials, text, images, etc., used in this channel are sourced from the internet, and all reposts are duly credited to their sources. If you discover any work that infringes on your intellectual property rights or personal legal interests, please contact us, and we will promptly modify or remove it.