Summary
Keegan McCallum, founder of uRun — a new inference provider focused on interactive media — argues that the generative video space is experiencing an efficiency revolution as significant as its quality improvements. While most attention goes to flagship models like Sora or SeeDance, McCallum highlights a parallel wave of distilled, real-time-capable models: at least 40 such models with real-time or long-horizon capabilities have shipped in the past year alone.
The centerpiece example is Helios, a distillation of the Wan 2.1 14B model that uRun serves in production. McCallum gives concrete pricing: $10 buys roughly 3 hours of continuously generated real-time video, and $50 covers a full day of interactive AI visual experience. He contrasts this with the “slot machine” paradigm of current generative video — submit a prompt, receive a file, hope for the best — and outlines the use cases that only become viable when video is steerable in under a second: magic-mirror webcam transforms, real-time content creation, agent monitoring, and world-model-based interactive experiences.
On the infrastructure side, McCallum walks through what building on top of these models actually requires: globally distributed GPUs, WebRTC/ICE/TURN setup, multi-model streaming pipelines, and frame-synchronized controls. uRun’s answer is a drop-in React component that abstracts this complexity, letting developers integrate real-time generative video without building the underlying harness — positioning uRun as the AWS of interactive video inference.
📺 Source: AI Engineer · Published August 18, 2026
🏷️ Format: Keynote Launch
