Under 100 milliseconds. That's the time-to-first-frame Runway demonstrated for its new real-time video generation model at Nvidia's GPU Technology Conference in San Jose on March 18. For reference, a human eye blink takes about 150ms.
The demo ran on Nvidia's Vera Rubin supercomputer - a machine spec'd at 36 Vero CPUs, 72 Rubin GPUs, and over 54TB of combined memory. Type a prompt, get HD video output faster than you can process what you're seeing. Previous AI video tools like Runway's own Gen-3 Alpha take seconds to minutes per clip. Sub-100ms puts this in the territory of interactive, real-time generation rather than batch processing.
Impressive Hardware, No Product Yet
The gap between "works on a supercomputer at a tech conference" and "available to creators" is wide. Runway hasn't announced pricing, availability, or even a name for this model. The Vera Rubin hardware it ran on isn't something you can rent on AWS.
That said, the direction is clear. Runway has been building toward what it calls playable world generation - interactive virtual environments that respond in real time. Real-time video generation is a prerequisite for that. If the latency holds up when the compute requirements shrink (and historically, they always shrink), this changes what tools like Runway can do.
Right now, AI video generation is a "submit and wait" workflow. You write a prompt, hit generate, wait 30 seconds to two minutes, review the result, iterate. Real-time generation turns that into something closer to live directing - adjust a prompt and immediately see the output shift. For video editors and content creators, that's the difference between a render queue and a creative instrument.
The practical timeline for any of this reaching consumer tools is unclear. But Runway choosing to demo at GTC rather than its own event signals this is as much about Nvidia selling hardware as Runway shipping features.