Use case · video generation
Video generation GPUs: VRAM for video, rented by the clip
Video diffusion is the most VRAM-hungry workload in the catalogue — and the best case for renting: generate on an 80 GB H100 PCIE at $2.147/hr only while the queue runs, then destroy it. Nobody buys a $25k card for weekend clips.
The video cards, by headroom
| Tier | GPU | VRAM | On-demand | Interruptible | Why this card | Action |
|---|---|---|---|---|---|---|
| Good | RTX 5090 | 32 GB | $0.318 | $0.159 | 32 GB GDDR7 — Wan 14B at 720p with offloading; the affordable way in. | Deploy |
| Better | H100 PCIE | 80 GB | $2.147 | $1.073 | 80 GB HBM3 — full-quality Wan/Hunyuan without offload gymnastics, 2–3× faster clips. | Deploy |
| Best | H200 | 141 GB | $3.058 | $1.529 | 141 GB — long clips, high resolutions and batch generation without a single OOM. | Deploy |
Middle path: the 48 GB L40S at $0.466/hr runs most 720p pipelines at full precision.
A batch night, budgeted
60 five-second clips for a storyboard, overnight on interruptible:
| GPU time | ~6 min/clip × 60 on H100 PCIE interruptible | $6.44 |
|---|---|---|
| Model volume | 200 GB (Wan + Hunyuan + LoRAs), one month | $16.00 |
| Download the results | ~3 GB out | $0.03 |
| Total for the night | ≈ $22.47 |
Interruptions cost nothing here: each clip is an independent queue item that re-runs from the volume.
Make it not hurt
- Models on a volume — video checkpoints are 20–60 GB; download once, mount forever.
- Prototype at 480p on the 5090, final-render at 720p+ on 80 GB — same workflow JSON.
- Queue as items, not marathons — per-clip jobs love interruptible pricing.
- Watch VRAM, not GPU% — video pipelines OOM before they saturate compute.
AI video GPUs: FAQ
How much VRAM do video models need?
More than image models by an order of magnitude of activations: Wan 2.1 14B wants 24–32 GB for 720p clips with offloading, comfortable at 48 GB; HunyuanVideo prefers 48–80 GB. The 32 GB RTX 5090 is the realistic entry point, 80 GB cards the comfortable one.
What does a clip cost to generate?
A 5-second 720p Wan 2.1 clip takes roughly 4–8 minutes on an H100 PCIE — about $0.21 at $2.147/hr. Batch overnight on interruptible and the per-clip cost halves.
Which template do I start from?
ComfyUI — current video models (Wan, Hunyuan, LTX, image-to-video pipelines) all ship ComfyUI workflows first. Put models on a volume: video checkpoints are 20–60 GB each.