AI models the physical world: world models, weather, and proteins

Cosmos for physical AI

NVIDIA expanded its Cosmos platform with new training tools, simulation checkpoints, and Cosmos Data Factory blueprints designed to accelerate physical AI world models, the company said. Cosmos Curator added support for multimodal multisensory inputs for curating video training datasets; Cosmos Data Factory now offers five cloud-native blueprints through GitHub, Microsoft Azure, Nebius, and CoreWeave; and updated Cosmos 3 NIM microservices deliver up to 2x faster inference with no quality loss in validated configurations. MicroAGI post-trained Cosmos 3 on videos of humans manipulating objects and distilled its 32-billion-parameter reasoner down to an 8-billion-parameter Nano model with no loss in accuracy. m:…8988

The adoption stories are about replacing physical iteration with simulation. Porsche is leading development of a shared autonomous-vehicle data flywheel across Volkswagen Group brands using Cosmos Data Factory through Voxel51 on Microsoft Azure — expected, per NVIDIA, to reduce repeated physical test drives by up to 90 percent and shorten development cycle times by up to 85 percent. ANYbotics generated over 1,500 synthetic video clips with Data Factory blueprints, validating outputs with Encord's platform. Ultralytics retrained its YOLO model with real images plus targeted synthetic defects generated with Cosmos Data Factory, improving defect localization and recall. ADNOC uses Fogsphere's MistIQ agentic platform, built with Cosmos and the Metropolis video-search blueprint, to detect safety hazards in real time across 65 sites. m:…8988

Weather as an open model

NVIDIA introduced StormScope EU, a regional version of its StormScope model for short-term forecasting trained on European weather and climate data, as part of the Earth-2 platform. The model uses EUMETSAT satellite observations to predict evolving cloud conditions across 16 Meteosat channels at 10-minute intervals, and NVIDIA showcased a live pipeline at GTC Berlin in a demo built with Earthmover. Earth-2 gives developers and researchers a family of open AI models and tools they can customize with their own data, with the open-source PhysicsNeMo framework for training and customization. m:…7306

AXA is using Earth-2's FourCastNet 3 to study regional cyclone risk beyond the historical record across more than 1,000 simulated seasons, generating insights for risk assessment and prevention. TotalEnergies is evaluating StormScope-based solar-radiation nowcasts at two photovoltaic sites. Vaisala Xweather is combining Earth-2 with its proprietary lightning data and forecasting expertise, and University of Manchester researchers adapted the CorrDiff model for UK air-quality modeling — training on the NVIDIA-powered Isambard-AI supercomputer and running locally on DGX Spark systems. m:…7306

Proteins by search

RFD4-Proteina treats protein design as a search problem: choose amino acids for a 100-residue sequence and the possibilities exceed the number of atoms in the observable universe, NVIDIA noted — an eight-residue sequence alone has more than 25 billion combinations. The method uses reward-guided inference-time search to navigate that space, aiming to find protein sequences that open paths to new therapies. Somewhere in that space, the company said, is a protein that opens a path to a new therapy. m:…4331

What it means

This is my read, not a reported fact: the common thread is simulation replacing physical iteration — test drives, historical weather records, wet-lab search. World models are becoming infrastructure you build on rather than research you read about, and the open-weights pattern (Earth-2, PhysicsNeMo) is how domain players like insurers and energy companies get in. The thing to watch is validation: synthetic data is only as good as its grounding, and the companies furthest ahead are the ones pairing simulation with real-world measurement loops. m:…8988 m:…7306