1. Birth of Physical AI
2. Building a Digital Twin Environment for AI Learning
3. NVIDIA Anticipates Cooperation with Robot Companies Worldwide
4. Dawn of the AI PC Era: Collaboration with Windows
5. AI Agents Spreading Across All Industries
NVIDIA CEO Jensen Huang revealed the autonomous driving and robot development platform 'Cosmos' in his CES 2025 keynote address, heralding innovation. CEO Huang emphasized that "the ChatGPT moment for robotics is coming" and that Cosmos will accelerate the development of robots and autonomous vehicles.
'Cosmos' is a platform that helps autonomous vehicles and robots learn and understand data from the real world. NVIDIA is cooperating with Japan's Toyota to develop autonomous driving technology. Plans are to supply AI chips and operating systems to Toyota.
Leveraging NVIDIA's semiconductor technology and software platform, it was designed so developers can efficiently train and test robots in virtual environments.
Cosmos is an AI development platform composed of World Foundation Models, advanced tokenizers, video processing pipelines, and more. It supports the development of autonomous vehicles and robots by generating Physical AI-based synthetic data. Through this, various environments such as warehouses, factories, and roads can be simulated with high quality. Combined with NVIDIA's 'Omniverse,' simulation of various situations is also possible.
In the autonomous driving field, Toyota and Uber are cooperating with NVIDIA through Cosmos. For robots, cooperation is underway with leading robot startups including 1X, Agile, Agility, and Figure AI. Cosmos is scheduled to be released as open source.
[NVIDIA CEO Jensen Huang]
Birth of Physical AI
NVIDIA presented Physical AI as a new domain of AI. Current large language models (LLMs) operate by receiving context as input and generating tokens one by one. However, Physical AI expands by generating real-world actions like autonomous vehicles or robots instead of text.
Physical AI models must understand the surrounding environment, receive commands instead of questions, and perform actual actions. This is an essential element for robots and artificial intelligence to naturally interact with humans. To realize this, AI must understand the dynamics of the physical world such as gravity, friction, and inertia, and learn spatial relationships and causal relationships.
To realize this vision, NVIDIA announced the world's first Physical AI model, NVIDIA Cosmos, at CES 2025. Cosmos is a World Foundation Model focused on understanding the physical world. This model was trained on 20 million hours of video data.
This data focuses on dynamic elements related to physical dynamics such as human movement, hand manipulation, and rapid camera movements.
The goal is not to generate creative content but to train AI to understand the physical world.
Cosmos features auto-regressive models, diffusion-based models, advanced tokenizers, and AI-accelerated data pipelines.
Auto-regressive model: An AI system capable of modeling and predicting the world in real time.
Diffusion-based model: Generates high-quality images and videos to provide training data.
Advanced tokenizer: A function for learning physical concepts of the real world.
AI-accelerated data pipeline: End-to-end AI acceleration for large-scale data processing.
Cosmos can receive text, image, or video prompts to generate virtual worlds as video. This is expected to become an essential technology in the autonomous vehicle (AV) and robotics fields.
The Cosmos model receives text, image, and video prompts as input and generates virtual world states as video. Cosmos's generation prioritizes elements required in actual use cases such as autonomous driving and robotics, including realistic environments, lighting, and object persistence.
Building a Digital Twin Environment for AI Learning
NVIDIA is building a digital twin environment for AI learning and validation by combining Cosmos with the Omniverse platform.
Omniverse provides a physics simulation-based AI training environment, enabling robots and AI systems to predict and learn all scenarios that may occur in real environments.
For example, in industrial automation, factory digital twins can be used to pre-simulate and optimize robot operations. For autonomous vehicles, safety can be enhanced by experimenting with various driving conditions in virtual environments and applying optimal models.
Developers use NVIDIA Omniverse to generate physically accurate geographic simulations. They can then output Omniverse renderings to Cosmos to generate realistic, physically-based synthetic data, reflecting conditions such as weather and time-of-day scenarios.
Developers use Cosmos to create environments for reinforcement learning. Policy models can be improved through AI feedback, or model performance can be tested and validated from multiple sensor perspectives.
NVIDIA supports developers in creating physically accurate simulations that recreate real environments using Omniverse. The data thus generated is used for training and testing Cosmos models. Through this, various environments such as warehouses, factories, and roads can be simulated with high quality.
For autonomous vehicles, for example, realistic simulation data considering road environments, pedestrians, weather changes, and lighting conditions is needed. Omniverse generates synthetic data based on real physics laws so that autonomous driving AI models can learn more accurately and safely.
NVIDIA presented the goal of equipping AI with not just information processing but the ability to understand and adapt to the world.
NVIDIA's AI technology is accelerating autonomous vehicle development in cooperation with major automakers such as Toyota. With Waymo and Tesla's success, it was expected that autonomous driving would finally be realized in reality.
NVIDIA's solutions include △DGX systems for AI training, △Omniverse and Cosmos systems for simulation and synthetic data generation, and △AI computers for real-time in-vehicle processing.
NVIDIA cooperates with nearly all major automobile companies: not only Waymo, Zoox, and Tesla, but also BYD, JLR, Mercedes-Benz, Toyota, Lucid, Rivian, Xiaomi, and Volvo.
Aurora, Kodiak, and Waabi are developing autonomous trucks using NVIDIA technology. The autonomous vehicle industry will grow to a scale of trillions of dollars going forward.
NVIDIA's vision is to reduce transportation costs, strengthen safety, and maximize road efficiency through autonomous driving technology. Cosmos will collect thousands of drives in the future and convert them into billions of miles of driving data.
He indicated that "we will secure a vast amount of training data for autonomous vehicles" and that "we will continue to collect data as long as we live," suggesting continued technology advancement.
The data thus generated supports securing physically accurate and valid large amounts of data.
CEO Huang stated: "The autonomous driving industry has already arrived. There will be tremendous advances in the coming years," adding "just as computer graphics innovated at an amazing pace, autonomous driving technology will also advance rapidly in the coming years."
An operating system (OS) for electric vehicles was also revealed.
Jensen Huang explained that NVIDIA's 'Drive OS' has achieved top safety ratings, is built with 7 million lines of code, and was developed by 15,000 engineers over one year. He added that it has already been verified through 2 million tests.
NVIDIA's automotive chips and OS are utilized for Toyota's autonomous vehicles.
Toyota stated that Drive OS integrated NVIDIA's autonomous driving platform with safe real-time AI processing and advanced driving capabilities as an operating system.
The automotive AI processor 'Thor' was developed in collaboration with Japan's Toyota, Aurora, and Continental.
Thor provides 20x improved performance over the previous generation and is a general-purpose robotics computer capable of serving as the brain of not only autonomous vehicles but also industrial and humanoid robots. Thor will enter production in 2025.
Thor functions to integrate various sensor information needed for autonomous driving such as cameras, LiDAR, and LADAR to accurately predict driving paths. It is the brain of autonomous vehicles.
CEO Huang expressed anticipation that "the autonomous vehicle market will become the first trillion-dollar robotics market." He predicted that NVIDIA's automotive segment revenue would reach approximately $5 billion in fiscal year 2026.
NVIDIA Anticipates Cooperation with Robot Companies Worldwide
Another field NVIDIA is focusing on with AI is robotics. Going forward, AI robots will play essential roles in all industries.
The three important elements of robot development are as follows.
Computers for training AI, edge AI computers for running actual robots, and simulation and digital twin technology.
NVIDIA is building these elements into one integrated ecosystem.
Robot companies worldwide are using NVIDIA technology to develop smarter and more efficient robots. We are now entering an era when AI robots are becoming widespread across manufacturing, logistics, and service industries.
Robots fundamentally change industries. AI robots will become systems that understand environments, adapt, and learn autonomously, going beyond simple automation machines.
CEO Jensen Huang brought 14 humanoid robots onto the stage during this keynote. The stage featured representative robots developed in the United States including Boston Dynamics's 'E-Atlas,' Agility Robotics' 'Digit,' Figure's 'Figure 02,' and Apptronik's 'Apollo.'
Chinese company products included Unitree's 'H1,' Xiaopeng's 'Iron,' Galbot's 'G1,' RobotEra's 'Star1,' Agibot's 'A2,' and Fourier's 'GR-2.' From other countries, Norway's 1X 'Neo,' Israel's Mentee 'Menteebot,' Germany's NeuraRobotics '4NE-1,' and Canada's Sanctuary AI 'Phoenix' also appeared.
Dawn of the AI PC Era: Collaboration with Windows
NVIDIA plans to continue AI technology innovation across various industries including AI PCs, physical AI, robotics, and autonomous driving.
In particular, NVIDIA Cosmos is expected to establish itself as the core platform that enables AI to understand and interact with the physical world in real environments. Through this, AI's potential is expected to further expand in various fields such as industrial automation, smart factories, autonomous driving, and robotics.
NVIDIA's AI technology presents a next-generation paradigm that helps AI act directly in the real world, going beyond simply processing data. A new chapter of AI innovation is opening, and the impact AI will have on our lives and industries as a whole will grow even more.
CEO Huang revealed the ambitious plan to transform hundreds of millions of Windows PCs worldwide into AI PCs. NVIDIA is building AI-based PC environments in cooperation with all major PC manufacturers. Just as Microsoft and Apple led the PC revolution supplying computers to ordinary homes in the 1980s, AI PCs are expected to become a new revolution in our lives.
CEO Huang said that video AI can run on hundreds of millions of Windows PCs worldwide. He emphasized "NVIDIA is prepared in cooperation with AI PC manufacturers (OEMs) worldwide" and "AI PCs will soon come near your home."
NVIDIA's goal is for AI to run smoothly not only in the cloud but also on personal PCs. CEO Huang declared "just as Windows 95 brought a multimedia revolution, we will now innovate Windows PCs into 'AI PCs.'
NVIDIA cooperates with Microsoft to provide an AI-optimized environment utilizing 'Windows Subsystem for Linux 2 (WSL2).' WSL2 is technology enabling a Linux environment to run within the existing Windows OS. It is compatible with NVIDIA's CUDA and AI acceleration technology. Through this, users can run and utilize AI models even on personal PCs.
AI Agents Spreading Across All Industries
NVIDIA emphasized that AI agents will play core roles across all industries, going beyond simple chatbots.
AI is changing our work and lives: △AI research assistants that analyze papers and financial reports to summarize key information, △AI for software development, △AI code assistants helping developers with coding, △code security inspection and optimization support, △AI for pharmaceutical and life science research, △acceleration of drug candidate substance discovery, △smart city and industrial AI, △traffic flow optimization through camera data analysis, and △safety monitoring and automation of industrial facilities.
CEO Huang emphasized that AI is not simply a technology but will become a 'digital workforce' of the future. He stated "the IT department of enterprises will now become the HR department of AI agents," announcing the arrival of a new era where AI collaborates with people to perform tasks.
At CES 2025, he presented a strong vision for the advancement of AI. The ambition is to build an AI ecosystem encompassing cloud, enterprises, and personal PCs. NVIDIA's goal is to enable AI, which started in the cloud, to run on all devices.
CEO Huang emphasized that the future of AI lies in 'Agentic AI.' Breaking away from the existing AI model method of answering questions, the forecast is that it will evolve in a way that thinks on its own, responds internally, and solves complex problems.
Agentic AI is a system where multiple AI models collaborate and interact with users. For example, when AI receives a user's question, it derives optimal answers through multi-stage processes including △information retrieval, △PDF document analysis, △calculator use, and △chart generation.
For this, NVIDIA developed AI microservices called 'NIM.' NIM provides sophisticated AI software in container form, enabling developers to easily build AI applications. Through this, AI models can be optimized for various industry groups. It supports NVIDIA's AI technology to be utilized anywhere including cloud, enterprise data centers, and personal PCs.
NVIDIA announced the AI model family 'Llama Nemotron' based on Meta's Llama model. CEO Huang explained "we optimized the Llama model for enterprise environments" and that companies would be able to utilize AI more efficiently through this.
Llama Nemotron is provided in models of various sizes including Super and Ultra models.
- Super model: Suitable for applications requiring fast response times
- Ultra model: A powerful AI model for large-scale data learning and knowledge transfer
In particular, the Ultra model contributes to enhancing AI quality by playing the role of evaluating and complementing other AI models. This model is scheduled to be utilized in various industry groups in cooperation with ServiceNow, SAP, Siemens, and others.
At CES 2025, NVIDIA once again demonstrated that it is creating a new AI paradigm as a leader in AI computing.
CEO Huang's keynote drew more than 10,000 attendees, proving him to be the top star of CES 2025 leading the AI era. At the venue, Mandalay Bay in Las Vegas, USA, attendees lined up for hundreds of meters starting 2-3 hours before the keynote began.
