Researchers and Startups Shift Toward World Models

Runway raised a $315 million Series E at a $5.3 billion valuation in February 2026, TechCrunch reported, with the round led by General Atlantic and joined by Nvidia, Fidelity Management & Research, AllianceBernstein, Adobe Ventures, Mirae Asset, Felicis, Premji Invest, and AMD Ventures. The company said the funds will go toward pretraining its next generation of world models, following the December release of its first world model, which reporting from AI Business and TechBuzz says shipped in specialized variants for environment simulation, robotics, and digital avatars. Separately, Yann LeCun announced his departure from Meta in November 2025 and, per TechCrunch and MIT Technology Review, launched AMI Labs in March 2026 with a $1.03 billion seed round to build JEPA-based world models for robotics and autonomous machines. A Towards AI piece that first flagged this trend, along with an Orange County Register report and a Threads excerpt of The Information's reporting, describe additional startup formation by ex-Nvidia and ex-DeepMind researchers pursuing similar world-model research.

Runway's pivot puts it in direct competition with Fei-Fei Li's World Labs, which has raised roughly $1.2 billion (including a $1 billion round backed by Autodesk, Nvidia, and AMD) and shipped its Marble 3D-world-generation product, and with Google DeepMind's Genie world-generation research - both identified by TechCrunch as Runway's closest rivals. Nvidia is pursuing its own Cosmos world foundation models while also financially backing rival labs, underscoring how compute vendors are hedging across the field rather than betting on one winner.

Unlike LLMs trained as next-token predictors, world models are trained to simulate physical dynamics, spatial relationships, and cause-and-effect, which shifts evaluation away from perplexity and human-preference scoring toward metrics for physical realism, collision fidelity, and long-horizon stability. Teams building for robotics or autonomous systems should expect heavier reliance on simulated or instrumented real-world datasets, domain-randomized training scenarios, and closed-loop rollout testing rather than text-scaling alone.

Track which datasets, simulators, and evaluation suites get open-sourced or standardized as more labs enter the field, and watch whether funding broadens beyond a handful of well-capitalized startups (Runway, AMI Labs, World Labs) or consolidates around a few dominant platforms. The claim that ex-Nvidia and ex-DeepMind researchers are separately raising nine-figure rounds comes from a single Threads excerpt of paywalled Information reporting and is worth confirming as more detail becomes public.

This is not a simple model-class swap. Integrating world models into a production stack requires new data pipelines, simulation engineering, physics-aware evaluation, and safety regimes that differ materially from LLM-centric deployment - teams evaluating this space should budget for that infrastructure shift separately from any single vendor's model quality.