MiniMax H3 video gen on rented GPUs: 4090 does a clip for $0.013, L40S matches 4090 speed
Worldly_North_7213 · reddit · 2026-09-12
A developer benchmarked MiniMax H3 text-to-video (5s, 864x480, 20 steps, stock ComfyUI graph, int8convrot weights, identical seeds) across four GPU rental providers and published raw CSVs from self-funded runs totaling $1.67.
- Steady-state cost per clip: Vast.ai RTX 4090 (spot $0.40/h) is cheapest at 93s and $0.013/clip; RunPod 4090 $0.019; Nebius L40S $0.024; Hyperstack H100 PCIe fastest at 67s but $0.038.
- Two surprises: the L40S runs at 4090-class speed (3.97 s/step vs 2.99 on H100 PCIe), and the 4090 streams the 34 GB DiT through ComfyUI's dynamic VRAM path using only 23 GB VRAM — but peaks at 68 GB host RAM, so ≥64 GB RAM is a hard requirement.
- Short jobs live on session overhead: the 67 GB weight download hit 450 MB/s on Vast but only 104 MB/s on Nebius, making Nebius's first clip take 11.5 minutes (1.8–2.7 min elsewhere); a new prompt adds 15–25 s of text encoding.
The author discloses they're building a service on this, but the numbers and methodology are fully public — a directly actionable reference for anyone self-hosting video generation on rented GPUs.
More from Infra
- NVIDIA releases NVFP4 quantized DeepSeek-V4.1-Flash, 665B params on HF — TheZachMueller · 2026-10-31
- Clover-35B-A3B Quantized to 29.4 GB Runs at ~42 tok/s on a 64 GB Mac — viveky259 · 2026-10-31
- Liquid Agent brings fully local LFM3 agent to Linux with one-command install — JosephJacks_ · 2026-10-31
- A fully local office AI stack: MiniCPM, Qwen3, LFM2.5 and Whisper — HowDevelop · 2026-10-31
- a16z: GPU rental prices up 106% in 3 months while token prices fall 58% — a16z · 2026-10-31
- vLLM nightly adds PLE disk offload, with a telling [Qwen4Exp] tag on the PR — rmhubbert · 2026-10-31