The useful question is what changes for users, developers or buyers, and whether the announcement stays industry context or becomes something people can actually use.

This year’s ICML conference underscores the growing reliance on open frontier models and infrastructure in AI research, with NVIDIA’s contributions prominently featured. Of the approximately 2,000 accepted papers citing NVIDIA GPUs, 145 specifically reference Nemotron open models and datasets as foundational tools. Research spans physical AI, robotics, autonomous vehicles, and biomedical fields, reflecting sustained investment in areas like vision, video generation, and reinforcement learning for large language models.

Robot world models emerged as a key theme, exemplified by projects like DreamDojo, which leverages NVIDIA Cosmos open models to simulate robot behavior in unseen environments. The approach enables researchers to evaluate policies, plan actions, and teleoperate virtual robots, reducing the costs and risks associated with physical testing. In life sciences, NVIDIA BioNeMo open models facilitated research into protein function and drug discovery, with models like KERMT and benchmarks such as FLIP2 advancing predictive capabilities in molecular behavior.

Synthetic data generation gained traction at ICML 2026, with Nemotron and physical AI datasets enabling large-scale training without heavy reliance on human-labeled data. NVIDIA’s open infrastructure, including Nemotron’s research stack and NeMo Curator, provides reproducible tools for data curation, reasoning, safety, and efficient inference. The Cosmos 3 family of omnimodels further expands capabilities in robotics, autonomous vehicles, and vision AI, offering tools to perceive, reason, plan, and act in physical environments.

Collaborations highlighted the broader impact of NVIDIA’s open models. Companies like Merck & Co. used KERMT for drug molecule prediction, while Sakana AI built models on Nemotron 3 Ultra to advance AI research automation. KiloCode reported token cost reductions of up to 90% by integrating Nemotron, and organizations such as NAVER extended the architecture for Korean-language AI research. The trend extends to industrial deployments, with companies adopting Isaac GR00T and Cosmos models to accelerate humanoid robot development.