Pool topic: Real-time, continuous, and interactive AI video generation. Question revision: 1. Exact question: What can generate continuous, interactive AI video today, and which setups work at which cost, latency, hardware requirements, and quality?
Checked 3 October 2026 · Independent project report
Eight H100s finish 14.375 seconds of H3 video in 13.506 seconds. Is that live?
Project says: a warm FastH3 Preview v1 job saved a 1344×768, 24 fps MP4 with native audio before its own playback duration ended. That is a useful Hopper capacity lead, not a tested show.
Keep the two tests separate
| Test | Result | Limit |
|---|---|---|
| Eight H100 80 GB GPUs, 345 frames, 14.375 seconds of output | 13.503, 13.506 and 13.554 seconds to saved MP4, warm | Four-forward sparse checkpoint; load and compilation excluded. Quality suite not run. |
| Matched eight-H200 stock versus patched pipeline | 23.810 to 14.085 seconds | Different hardware from the H100 headline. Do not use this speedup as an H100 A/B. |
What to test before buying capacity
The project pins its patches to one FastVideo revision and reports no accepted-quality score. A complete clip faster than playback can fill a queue, but it says nothing about cold starts, retries, audio continuity, input-to-screen delay or what a viewer sees when the queue runs dry. Ask for a timestamped second and third clip at a receiver, one delayed job and the billed GPU time. Without those, live service cost is Unknown.
A builder asking whether a single RTX 5090 can cut a 15-second H3 clip to one minute has a different decision. This eight-H100 result is not a 5090 recipe or a cost-per-finished-clip comparison.