The startup that first took AI seriously in filmmaking has now declared war on Google. Meet Runway. After 7 years, they're a $5.3B AI company — and their focus isn't "better video generation" anymore.
They want to build an AI that simulates the world itself. They call it a "World Model."
What's this declaration about?
Runway was founded in 2018 by three graduates of NYU's ITP (Interactive Communications Program). Cristóbal Valenzuela (co-CEO), Anastasis Germanidis (co-CEO), and Alejandro Matamala Ortiz — from Chile, Greece, and Chile, respectively. Starting outside Silicon Valley is a core part of their identity.
They started as AI tools for filmmakers. In February 2023, they launched their first video generation model — and the Oscar-winning Everything Everywhere All at Once actually used their tools. They've since signed partnerships with Lionsgate and AMC Networks.
Then, in December 2025, they pivoted. Beyond video generation, they announced a "General World Model" that simulates the world itself. GWM-1.
Here's the thing about LLMs: they learn from text on the internet. They "understand" that things fall when thrown — from millions of sentences. But fully internalizing gravity, air resistance, and surface friction from text alone has real limits.
World models take a different approach. Instead of text, they learn from video and physical observation data. Water flowing, a ball bouncing, a robot arm grabbing an object — the model internalizes how the world works by watching it. As co-CEO Germanidis puts it: using "less biased data than language."
What's actually different from LLMs?
GWM-1 is an autoregressive model built on top of Gen-4.5. It generates frame-by-frame, runs in real time, and can be controlled interactively. Three variants have launched.
| LLM (Language Model) | World Model (GWM-1) | |
|---|---|---|
| Training data | Text (internet documents) | Video + physical observation |
| Physics | Relies on linguistic associations | Internalizes physical laws |
| Interactivity | Text generation (one-way) | Real-time control (camera, robot commands) |
| Spatial understanding | 2D text descriptions | Geometry, lighting, physics — all at once |
| Applications | Text gen, Q&A, coding | Robotics, drug discovery, gaming, education |
Three variants, three different bets:
- GWM Worlds — Infinite explorable spaces
Give it a text or image prompt and it generates a virtually infinite explorable space — geometry, lighting, and physics simulated in real time at 24fps and 720p. This changes how game dev, VR, and metaverse content get built. - GWM Robotics — Simulator for robot policy training
A diffusion backbone that gives robot policy models rich visual understanding of the physical world. Predicts video rollouts conditioned on robot actions and generates counterfactual scenarios — high-sample-efficiency training without actual hardware. Runway already runs its own robotics unit. - GWM Avatars / Runway Characters — Real-time conversational characters
Turn a single image into a real-time conversational character — natural facial expressions, eye movements, lip-sync, and gestures included. Launched in May 2026 as the "Runway Characters" API.
Competitive landscape, fast
Others in the race: Google (Genie world model + Veo video), Luma AI ($900M raised), World Labs (Fei-Fei Li, $1.29B raised). Notably, OpenAI's Sora shut down in March 2026 — even with $2.1M daily revenue against $1M daily costs, it ran at a loss.
The essentials: How to start with world models
- Start with the Runway Characters API
The fastest entry point. Create a Runway account at runwayml.com, upload an image, and get a real-time conversational character. Ready for education, customer service, or interactive content right out of the box. - Prototype virtual environments with GWM Worlds
Designing a game scene or VR experience? Spin it up with GWM Worlds first. One text prompt generates a physics-accurate explorable space. Iteration speed beats traditional 3D modeling by a wide margin. - Try GWM Robotics for policy training experiments
Building robotics or physical AI? Check the GWM Robotics API. Train policy models in simulation without hardware. Especially useful for counterfactual scenario generation in safety-critical environments. - Check Gen-4.5 for your current video workflow
Not ready for world models? Start with Gen-4.5. It's #1 on the Video Arena leaderboard (Elo 1,247), ahead of Google Veo and OpenAI. Works well for ad videos, social content, and product demos. - Enterprise API integration
For large-scale content production or robotics research, start with Runway's enterprise API docs. The Lionsgate and AMC Networks case studies are useful for prioritizing your integration approach.
Want to go deeper?
Runway Research: Introducing Runway GWM-1 Official announcement — technical architecture and demos for all three variants (Worlds, Robotics, Avatars). runwayml.com
TechCrunch: Runway raises $315M at $5.3B valuation February 2026 Series E details — NVIDIA and AMD Ventures involvement and the world model strategy. techcrunch.com
Runway + NVIDIA Rubin partnership How Gen-4.5 was ported from Hopper to Vera Rubin NVL72 in a single day — and the compute infrastructure strategy behind world models. runwayml.com
Why World Model AI Will Shake Up 2026 Korean-language breakdown comparing world model concepts with LLM limitations in physical understanding. ezyeconomy.com
Startup Fortune: Runway is turning AI video into a world model business Business analysis of the pivot from video generation to world models. startupfortune.com
Introducing Runway Characters Official GWM-1-based real-time interactive character API launch — conversational characters from a single image. runwayml.com
