Most AI companies chase the same prize: making language models bigger, faster, and more capable. But there’s a growing contingent betting that language is only part of the picture — and that the next breakthrough in artificial intelligence won’t come from words at all, but from how AI learns to see and understand the physical world.
Runway, the AI video generation startup valued at $5.3 billion and founded by three NYU arts graduates, is at the center of that bet. The company built its name on tools that let filmmakers and creatives generate cinematic video from text prompts — think of it as Photoshop for generative video. Its technology has been used in Oscar-winning films like “Everything Everywhere All At Once,” and it counts Lionsgate and AMC Networks among its partners. Earlier this year, Runway added $40 million in annual recurring revenue in a single quarter.
But the company has quietly been shifting its ambitions far beyond Hollywood. In December 2025, Runway released its first world model — an AI system trained not on text and language, but on raw observational data from the real world. World models learn how physics, motion, and environments actually behave, which means they can simulate outcomes, predict reactions, and eventually help scientists run experiments that would take months or years in the physical world.
This is the core insight that sets Runway apart from the industry’s giants. While Google, OpenAI, and Anthropic have poured billions into scaling language models — models that essentially compress the text of the internet into a statistical representation — Runway’s founders believe that route hits a ceiling. Language, they argue, is filtered through human bias and existing knowledge. It can describe the world, but it cannot observe it directly.
Co-founder and co-CEO Anastasis Germanidis put it bluntly: “Language models are trained on the entire internet, on message boards and social media, on textbooks — distilling the existing human knowledge. But to get beyond that, we need to leverage less biased data.”
That “less biased data” means video — billions of hours of real-world movement, interaction, and physical behavior. The theory is that a model trained on enough sensory data will develop an implicit understanding of cause and effect, gravity, object permanence, and even social dynamics. Once you have that, you’re no longer limited to generating entertaining clips. You can simulate drug interactions in a digital petri dish, train robots in virtual environments, and model climate scenarios faster than any supercomputer could calculate.
Why This Matters for Startups
Runway’s trajectory holds a powerful lesson for founders building in AI’s shadow. The company didn’t follow the typical Silicon Valley playbook. Its founders weren’t Stanford PhDs with Google on their resumes. They were artists and filmmakers who started with a narrow, practical question: “Can we use AI to help people make better videos?” That question led to a product, which led to revenue, which led to a research breakthrough they hadn’t planned for.
The world model capability emerged from building video generation tools, not from a grand research lab strategy. In other words, Runway’s product-led approach accidentally unlocked a scientific moat that competitors with billions more in funding are still racing to match.
That’s a pattern worth watching. The next generation of transformative AI companies may not come from pure research outfits. They may come from startups that built something useful, learned from the data their users generated, and followed the technology where it led — even if that meant outgrowing their original mission.
The Competitive Landscape
Runway isn’t alone in pursuing this vision. Luma and World Labs are on similar trajectories, and Google has its own Genie world model in development. But Runway has an edge the others don’t: an actual paying customer base and a product that generates real revenue today, not just promises for tomorrow. Its 155 employees across six offices have shipped multiple model generations — from Gen-2 (which was laughably primitive at launch) to Gen-4.5, which is genuinely impressive — in just over three years.
Whether that head start will matter against Google’s computing resources and talent pool remains to be seen. But for a company that started with three artists in New York, the bet is no longer speculative. It’s already in motion.
Based on reporting from TechCrunch. Read the original article here.
The Takeaway for Startup Founders
The lesson of Runway’s journey is this: don’t let where you start define where you can go. A tool for filmmakers became a platform for simulating physics. A tiny team with no Big Tech pedigree is now competing with Google on one of AI’s hardest problems. If you’re building something today that feels small, give it room to grow into something you couldn’t have imagined at the beginning. The most important breakthroughs often emerge from the least expected places.
