InfiniteTalk on our own Vast GPU vs WaveSpeed
Same still, same voice track, same prompt as the WaveSpeed take podcast-eq A1-plant-steroid. Left is our RTX 5090, right is WaveSpeed. Both are 464x832 at 25 fps.
10 seconds
Our GPU: 365 s of GPU time, $0.068
WaveSpeed: first 10 s of the paid take, $0.30 at $0.03/s
20 seconds
Our GPU: 645 s of GPU time, $0.120
WaveSpeed: the full paid take, $0.61
Measured cost
| Item | Our RTX 5090 ($0.67/h) | WaveSpeed 480p |
|---|---|---|
| GPU seconds per video second | 36.5 (10 s clip), 31.8 (20 s clip), 29.7 once past the first window | not visible, billed per second |
| Cost per video second | $0.0055 to $0.0068 | $0.03 |
| A 60 s ad | about $0.33 and about 30 min on one card | $1.80 |
| Setup, fresh machine | 340 s and $0.10 (image pull plus 29 GB of weights) | none |
| Setup, warm machine from the pool | 3 s | none |
| Keeping one card warm all day | $16.05 a day | none |
At the volume we actually ran
| Day | WaveSpeed InfiniteTalk | Same seconds on our GPU, on demand | Same seconds with one card warm 24 h |
|---|---|---|---|
| Nov 2 | 8 takes, 67 s, about $2.00 | about $0.40 plus $0.43 per burst | $16.05 |
| Nov 3 | 51 takes, 449 s, about $13.50 | about $2.30 plus $0.43 per burst | $16.05 |
A burst is one fresh machine ($0.10 of setup and a 6 minute wait) plus the 30 minutes the pool keeps it after the last job (about $0.33). A card kept warm all day pays for itself above about 655 s of talking video a day; Nov 3 was 449 s.
Settings: ComfyUI core InfiniteTalk nodes, Wan2.1 I2V 14B 480p fp8, lightx2v LoRA, 6 steps, plain attention. Not yet tried: fewer steps, sage attention and fp8 matmul, each of which should cut the GPU time further. The 25 video-to-video takes on Nov 3 were priced as if they ran the same way; that mode was not tested here.