# Generative Video at the Speed of Light — Keegan McCallum, uRun

Source: [AI Engineer](https://www.youtube.com/watch?v=Xln-On3syJk)  
Feed7 permalink: https://feed7.dev/p/generative-video-at-the-speed-of-light-keegan-mccallum-urun-1x7i0w3  
Published: 2026-08-18T16:30:29.000Z  
Trust: Source Linked (source_linked)

## Why Included

Real-time video models are becoming cheap and responsive enough for agent interfaces, but builders still need global GPU routing, streaming infrastructure, and multi-model orchestration.

## Source Summary

McCallum says at least **40 real-time or long-horizon models** appeared this year. At current cited pricing, **$10 buys about 3 hours** of continuous generated video and $50 buys about 15 hours, with steering possible in under a second.

## Practical Implication

Builders can start treating video as a live agent surface rather than a render job: interactive avatars, webcam transformations, and observable agent output. The underlying harness needs GPU placement, WebRTC infrastructure, frame-level controls, and asynchronous model pipelines.

## Agent-Ready Context

McCallum says at least **40 real-time or long-horizon models** appeared this year. At current cited pricing, **$10 buys about 3 hours** of continuous generated video and $50 buys about 15 hours, with steering possible in under a second.

Builders can start treating video as a live agent surface rather than a render job: interactive avatars, webcam transformations, and observable agent output. The underlying harness needs GPU placement, WebRTC infrastructure, frame-level controls, and asynchronous model pipelines.

The cost and quality claims come from the provider’s talk and are not a cross-provider benchmark. Real-time output is described as comparable to older frontier quality, so higher fidelity may still require a separate render step.

## Context Map

- Layer: infra
- Domains: video
- Topics: generative-media, harness-engineering, tool-use

## Uncertainty

- The cost and quality claims come from the provider’s talk and are not a cross-provider benchmark. Real-time output is described as comparable to older frontier quality, so higher fidelity may still require a separate render step.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.