Physical AI News (Japan)

Physical AI, robotics and compute — reported from Japan.

日本語
· Product launches· 日本語で読む

NVIDIA CEO Huang Announces Physical AI Platform ‘Cosmos’ at CES 2025

NVIDIA CEO Jensen Huang said in his CES 2025 keynote that AI is advancing rapidly and that the world has entered the era of physical AI. He announced “Cosmos,” featuring world foundation models and a video-processing pipeline for robotics and autonomous driving. The keynote also introduced the GeForce RTX 50 series and DRIVE Hyperion for autonomous driving, among other products.

Key facts

  • NVIDIA CEO Jensen Huang delivered a 90-minute keynote at CES 2025 to an audience of more than 6,000.
  • NVIDIA Cosmos was announced as a physical AI platform with models and a video-processing pipeline for robotics, autonomous vehicles and vision AI.
  • The GeForce RTX 5090 has 92 billion transistors and AI processing performance of 3,352 trillion operations per second (TOPS), NVIDIA said.
  • NVIDIA DRIVE Hyperion was announced as an autonomous driving platform based on AGX Thor.
  • NVIDIA announced that it will use DRIVE AGX in-vehicle computers and DriveOS in next-generation vehicle development with Toyota.

What happened

In a Jan. 7, 2025, post, NVIDIA’s official blog published details of CEO Jensen Huang’s 90-minute CES 2025 keynote. Huang said AI is advancing rapidly and described “physical AI,” following perception AI and generative AI, as AI that reasons, plans and acts. He announced “NVIDIA Cosmos” for robots and autonomous vehicles, among other products.

Why it matters

NVIDIA said real-world autonomous driving data is limited and positioned synthetic data generated with Omniverse and Cosmos as a way to expand training data. NVIDIA said it will use DRIVE AGX and DriveOS in next-generation vehicle development with Toyota. The announcement concerns how in-vehicle computing platforms and safety features are combined in vehicle development.

Analysis

The announcement centers not only on individual GPU performance but also on how to build the computing and data infrastructure used by robots and autonomous vehicles. Cosmos generates virtual environments from video and other inputs and uses simulation to handle multiple scenarios. For autonomous driving, Huang cited “three computers”: DGX for training AI models, Omniverse for testing and synthetic data generation, and DRIVE AGX for in-vehicle computing. This setup separates training, validation and onboard execution, making how Cosmos and synthetic data connect these stages a key technical question. Meanwhile, the 3,352 trillion TOPS cited for the RTX 5090 and DLSS 4’s up to 8x performance improvement are selling points for gaming and PC products. These figures cannot be taken directly as measures of robot performance. In physical AI, separate questions include the difference between simulated environments and real-world sensor inputs, and how much generated data contributes to performance in real environments. NVIDIA said real-world autonomous driving data is limited and synthetic data is important, but the keynote materials did not provide data-quality metrics or figures for improvements in real vehicles or robots. In a related article, we covered a case in which NVIDIA Photonics CPO switches connect GB300 NVL72 racks equipped with more than 10,000 GPUs: Lambda Connects GB300 NVL72 Racks With NVIDIA CPO Switches, With More Than 10,000 GPUs. That concerns the network layer of large-scale computing infrastructure and is not the same deployment case as the robotics models or simulation platform announced at CES. In considering physical AI development, it is necessary to look separately at computing scale, video data processing, simulation and testing on real hardware. A report on rules covering remote access to advanced U.S. AI semiconductors, U.S. BIS Regulates Remote Access to Advanced AI Semiconductors, shows that conditions for using computing infrastructure are a separate issue from technical architecture. The CES announcement did not discuss regulatory compliance, so it is not possible to infer how the rules apply to Cosmos or related computing resources. Questions to follow include the effectiveness of data generated with Cosmos on real hardware, the specific scope of use by adopting companies, and safety assessments for in-vehicle systems.

Details

Cosmos is a physical AI platform that integrates generative models, tokenizers and a video-processing pipeline. It creates virtual environments for simulating robots and autonomous vehicles from text, image and video inputs, and can predict and evaluate multiple future scenarios. NVIDIA said it would make the platform available on GitHub under an open license. Companies NVIDIA named as early adopters include 1X, Agile Robots, Agility, Figure AI, Fourier, Galbot, NEURA Robotics, Skild AI, Uber and XPENG. The keynote also introduced the Blackwell-based GeForce RTX 50 series. The GeForce RTX 5090 has 92 billion transistors and AI processing performance of 3,352 trillion operations per second (TOPS), NVIDIA said. For autonomous driving, NVIDIA announced DRIVE Hyperion, based on AGX Thor. NVIDIA said it will use DRIVE AGX in-vehicle computers and DriveOS in next-generation vehicle development with Toyota.

Implications for Japan

NVIDIA announced that it will use DRIVE AGX in-vehicle computers and DriveOS in next-generation vehicle development with Toyota. For Japan’s automotive industry, this is a concrete example of combining an in-vehicle computing platform and software in vehicle development, but the announcement materials do not specify the vehicle models or scope of implementation.

Sources

Translated from the Japanese original: /news/nvidia-no-huang-ceo-ces-2025-de-fijikaru-ai-kiban-cosmos-wo-4f15a2. The Japanese version is authoritative if they differ.