So-called “agentic” AI is on everyone’s lips, including those of Nvidia. Jensen Huang, co-founder and CEO of the company, therefore spent a long time there during his opening conference of CES 2025, which was held in a packed 12,000-seat stadium and for which some people made more than two hours in line (L’Usine Digitale was _obviously_ in the VIP line).
Nemotron, a model family derived from Llama
Nvidia announced Nemotron, a family of large language models derived from Llama dedicated specifically to agentics. Nemotron is available in three variants: Nano (for embedded), Super (capable of running on a single GPU) and Ultra (for the data center). The models can be used as is or customized, for example to integrate RAG.
They have been specially trained for following instructions, chatting, calling functions, generating code, and math. A family of models for vision, called Cosmos Nemetron (not to be confused with the “Cosmos world foundation model”), was created in parallel, again with three versions. All these models are available directly for download on the Nvidia website and on Hugging Face, or in the form of NIMs microservices (requiring being an Nvidia AI Enterprise customer). SAP and ServiceNow will be among the first customers to deploy them.
Preconfigured agents to accelerate deployments
Nvidia had already launched NeMo, a framework for model development, and NIMs for model inference. This summer, he then announced Blueprints, pre-packaged models and workflows (using the Nvidia AI Enterprise suite and therefore NeMo and NIM) so that customers can deploy agents as easily as possible.
The company took advantage of CES to detail the agents now available. Five of them come from partners specializing in AI workflow orchestration: CrewAI, Daily, LangChain, LlamaIndex and Weights & Biases. Nvidia has also developed a Blueprint for converting PDF documents into podcasts (clearly inspired by Google’s NotebookLM), and another for searching and summarizing video content.
Robotics, the next big market that Nvidia wants to conquer
-Finally, four Blueprints build on Omniverse and are dedicated to “physical AI”. Translation: digital twins for robotics. Jensen Huang has insisted it’s the “next step” after agents, so it’s clear he sees a big potential market there. Nvidia has been working on the subject for years.
There is a Blueprint for streaming mixed reality on an Apple Vision Pro, another for autonomous driving simulation, a third for real-time visualization for CAD, and finally there is Mega. The latter makes it possible to train and test fleets of robots at scale in a factory or warehouse. Nvidia had already presented this type of use, but this time it has a concrete case developed by Accenture and Kion, a large German logistics company.
Kion creates the digital twin of a warehouse using CAD or BIM files and images, videos or Lidar captures. This twin is located in Omniverse where it is used to train and test the “brains” of the robots (which use Nvidia Isaac). The operation of the sensors and the system choices are fully simulated in this virtual environment. Kion’s logistics management system can assign them specific tasks, such as transferring a load from one location to another. Once the virtual robot has been trained correctly, all you have to do is load it into the real machine.
In addition to this use case, Accenture has integrated Mega into its “AI Refinery” offering intended for manufacturers and logisticians. On stage, Jensen Huang pointed out that the warehouse logistics market is worth $1 trillion. A word to the wise. At the start of the conference, Gary Shapiro, president of the CES, came to give a short introduction that was not far from genuflecting. He explained that if he had listened to Jensen a little better in the past, he could have retired by now. This is probably also true in this case.
Selected for you
