Generate images and video without overpaying for VRAM
Diffusion and video models are memory-hungry in different ways than LLMs. Match the card to resolution, frame count and latency target — from RTX 5090 to H100.
Recommended start
RTX 5090 $1.29/hr
48 GB
GDDR6 on L40S
32 GB
GDDR7 on RTX 5090
1 second
billing granularity
What makes this workload expensive
The bill is rarely about the GPU price alone. These are the problems we design around.
VRAM grows with pixels and frames
Higher resolution and longer clips multiply activation memory. A card that renders stills may not fit video.
Latency vs. batch
Interactive apps need fast single images; batch pipelines want maximum throughput per dollar.
Bursty creative workloads
Campaigns and launches create sudden peaks, then quiet weeks.
Built for the shape of the job
Consumer to datacenter
RTX 4090 and RTX 5090 for cost-efficient image work, L40S for 48 GB, H100 for large video models.
Pay for the burst
Spin up dozens of GPUs for a render or campaign and release them per second when the queue drains.
Object storage for outputs
S3-compatible buckets for generated assets with free ingress.
Serverless option
Put a model behind an endpoint that scales with requests.
Where to start, where to scale
RTX 5090
Start here32 GB VRAM · 1× NVIDIA RTX 5090
$1.29/hr
L40S
48 GB VRAM · 1× NVIDIA L40S
$1.19/hr
RTX 4090
24 GB VRAM · 1× NVIDIA RTX 4090
$0.89/hr
H100 SXM
80 GB VRAM · 1× NVIDIA H100
$2.69/hr
From first run to production
-
1
Measure peak VRAM at your target resolution and frame count.
-
2
Choose the smallest card that fits with headroom.
-
3
Scale out horizontally for batch jobs; use serverless for interactive traffic.
-
4
Store outputs in object storage and release GPUs when the queue is empty.
Guides for this workload
October 13, 2026
Cutting inference cold starts: container images and weights
October 3, 2026
Sizing GPUs for image and video generation models
June 9, 2026
Why VRAM, not FLOPS, limits your batch size
For most image models at common resolutions, yes. Larger video models typically need 48–80 GB or more.
Yes — LoRA training for diffusion models fits comfortably on 24–48 GB cards.
Write them to S3-compatible object storage; ingress is free.
Other workloads
Start training in minutes
Create an account, add a card, and launch your first GPU instance. Per-second billing means you only pay for what you use.
No minimum commitment · Cancel anytime · $10 free credit for new accounts