NVIDIA opens its Cosmos world models to robot and car makers at CES 2025, with Figure, Agility, XPENG and Uber among the first users

NVIDIA used its CES 2025 keynote on 6 January 2025 to launch Cosmos, a platform of world foundation models, video tokenizers and data-processing tools built to generate physics-aware video for training robots and autonomous vehicles. The company released the first Cosmos models under its NVIDIA Open Model License, which permits commercial use, and named 14 robotics and self-driving developers plus Uber as early adopters. On the same day it unveiled the Isaac GR00T Blueprint, a set of simulation workflows that turn a small number of human demonstrations into large synthetic motion datasets for humanoid robots.

The scale claims were large. In its launch blog, NVIDIA said the Cosmos models were trained on 9,000 trillion tokens drawn from 20 million hours of real-world human interactions, environment, industrial, robotics and driving data. The accompanying preprint, posted to arXiv on 7 January 2025 as paper 2501.03575, describes curating about 20 million hours of raw video at resolutions from 720p to 4K into roughly 100 million clips for pre-training and about 10 million for fine-tuning. NVIDIA said its NeMo Curator pipeline can process those 20 million hours in 14 days on Blackwell GPUs or 40 days on Hopper GPUs, against more than three years for a CPU-only pipeline, and that the Cosmos tokenizer delivers 8 times more total compression and 12 times faster processing than leading tokenizers.

NVIDIA grouped the models into three tiers. Nano is meant for real-time, low-latency inference at the edge, Super serves as a performant baseline, and Ultra targets maximum fidelity and the distillation of custom models. The first open checkpoints, published on Hugging Face as Cosmos 1.0, came in two architectures. Diffusion models with 7 billion and 14 billion parameters generate video from text (Text2World) or extend an input image or video (Video2World). Autoregressive models with 4 billion and 12 billion parameters predict future frames, with 5 billion and 13 billion parameter Video2World variants that also accept text. According to NVIDIA's model cards, the 7B diffusion model outputs 121 frames, about five seconds at 24 frames per second, at 1280x704 pixels and needs roughly 380 seconds on a single H100 GPU, while the 4B autoregressive model turns 9 input frames into 24 new frames at 1024x640 in about 62 seconds.

NVIDIA made the models available through Hugging Face, its NGC catalog and GitHub, with access through the NVIDIA API catalog, NIM microservices to follow and deployment on DGX Cloud. The company said Cosmos Guardrails screen prompts before generation and check generated videos afterwards, and that videos produced with the autoregressive and diffusion models on its API catalog carry an invisible watermark so they can be identified as AI-generated.

The adopter list in NVIDIA's release reads 1X, Agile Robots, Agility, Figure AI, Foretellix, Fourier, Galbot, Hillbot, IntBot, Neura Robotics, Skild AI, Virtual Incision, Waabi and XPENG, along with Uber. NVIDIA said Waabi would use Cosmos for data curation and simulation evaluation in autonomous driving, and that Hillbot would use it to generate terabytes of 3D environments for robot training. Pras Velagapudi, chief technology officer of Agility, said in the release that Cosmos lets the company generate and augment photorealistic scenarios for training without as much expensive real-world data capture. Uber chief executive Dara Khosrowshahi said working with NVIDIA could help speed up the timeline for safe and scalable autonomous driving.

NVIDIA founder and chief executive Jensen Huang framed the launch as a turning point. “The ChatGPT moment for robotics is coming,” he said in the release, arguing that world foundation models matter to robots and autonomous vehicles the way large language models matter to chatbots, but that few developers have the expertise and resources to train their own. “We created Cosmos to democratize physical AI and put general robotics in reach of every developer,” he said.

The Isaac GR00T Blueprint, announced the same day, has three workflows. GR00T-Teleop captures human movements through an Apple Vision Pro headset in a digital twin, with a simulated robot mirroring them to produce ground-truth data. GR00T-Mimic multiplies those demonstrations into a larger synthetic motion dataset. GR00T-Gen, built on NVIDIA Omniverse and Cosmos, expands the data further through domain randomization and 3D upscaling. Policies are trained in Isaac Lab, NVIDIA's open-source robot-learning framework. NVIDIA said Boston Dynamics and Figure had started adopting Isaac GR00T, invited humanoid hardware and software makers to apply for early access to its humanoid developer program, and cited an expectation that the humanoid robot market will reach $38 billion over the next decade.

NVIDIA later put numbers on the approach. A developer blog, revised during 2025, said GR00T-Mimic generated 780,000 synthetic trajectories, equivalent to 6,500 hours of human demonstration, in 11 hours, and that combining synthetic with real data improved performance by 40% compared with real data alone. At Computex on 18 May 2025, NVIDIA said a follow-on GR00T-Dreams blueprint built on Cosmos Predict models let its researchers develop the GR00T N1.5 humanoid model in 36 hours, against about three months of manual data collection.

Cosmos extended a stack NVIDIA had been assembling for years, with Omniverse for digital twins, Isaac Sim and Isaac Lab for simulation and training, and Jetson modules for on-robot compute. Each layer feeds GPU demand in data centres, which is where NVIDIA earns most of its money. The robotics adopters were mostly venture-backed companies still short of volume revenue. According to MEKAWIRE's verified company data, Figure's last disclosed round before CES was a $675 million Series B in February 2024, and Skild AI had raised $300 million at a $1.5 billion valuation in July 2024.

Cosmos arrived weeks after two prominent generative world and video models. Google DeepMind unveiled Genie 2 on 4 December 2024 as a foundation world model that generates action-controllable 3D environments for training embodied agents, consistent for up to a minute, without a broad public release. OpenAI opened Sora Turbo to ChatGPT Plus and Pro subscribers on 9 December 2024, generating up to 20 seconds of 1080p video. Cosmos differed on two counts. Its weights could be downloaded under a licence allowing commercial use, and it was pitched at robot and vehicle training data rather than consumer video.

In plain terms, a world model takes the recent past, as video frames plus an optional text prompt, and predicts what happens next. Cosmos tokenizers compress video into compact tokens. Diffusion models refine noise into a whole clip at once, while autoregressive models predict tokens step by step, as a language model predicts words. The preprint positions the pre-trained models as general-purpose bases that developers post-train on their own data, such as a specific robot's cameras or a car's sensor rig, to obtain a customised simulator or a way to evaluate policies before testing them on hardware.

The open questions were substantial. NVIDIA did not publish field results showing that robots trained on Cosmos video outperform those trained on real data, and generated video can look plausible while breaking physics. NVIDIA's own pages described the training corpus differently, with the GR00T blog citing 18 quadrillion tokens including 2 million hours of driving, robotics, drone and synthetic footage, against 9,000 trillion tokens from 20 million hours in the Cosmos blog. Generation is expensive, at minutes of H100 time per five-second clip, and the NVIDIA Open Model License is the company's own licence rather than a standard open-source one.