Local H3 video gen: 15s clip in ~5 min at zero cost vs ~$0.60 per generation on platforms
AIandDesign · x · 2026-09-22
A post breaks down the economics of running the H3 video model locally: one user generates a 15-second video in about 5 minutes on a 5090 at zero marginal cost, while the same generation on a hosted platform would run roughly $0.60 — enabling endless test-fail-iterate loops without worrying about credits.
AIandDesign jokes that anyone who bought a 5090 before GPU prices exploded will have "earned it back" in just over 3000 such generations, framing a local-vs-cloud iteration cost debate.
More from Infra
- Gas-for-AI-datacenters forum co-located with CoRL shows energy demand spillover — carlosdponx · 2026-11-12
- a16z podcast: Mike Volpi on Cisco's 70+ acquisition playbook and the AI buildout — a16z · 2026-11-12
- AMD to pass Intel in server revenue by mid-2027 as share gains 6 points a year — BenBajarin · 2026-11-12
- Study: Reflective Data Center Roofs Actually Cool Daytime Ground by 1.8-2.2°C — SumitGup · 2026-11-12
- Tsinghua's BaGuaLu-2 tames MoE expert-load spikes at runtime, accepted to SC26 — jiqizhixin · 2026-11-12
- antirez backs low-bit DeepSeek 4.1 Flash over Qwen 27B dense on Apple M5 Macs — antirez · 2026-11-12