TL;DR
- For SDXL the RTX 4090 wins images per dollar at $0.163 interruptible, roughly 2.9 seconds per image; for Flux the RTX 5090 leads.
- Rank cards by images per dollar, not hourly rate: 3,600 divided by seconds per image divided by hourly price.
- Flux dev FP8 needs about 17 GB, so 24 GB cards are the practical floor and 32 GB removes the ceiling.
- On PowerGPU cheap hourly cards like the RTX 3060 lose on throughput; keep checkpoints on a volume and stop idle instances.
The answer table
Eight rentable cards, live prices, images per dollar (interruptible rate — image queues pause gracefully):
| GPU | $/hr int. | SDXL s/img | SDXL img/$ | Flux s/img | Flux img/$ |
|---|---|---|---|---|---|
| RTX 3060 12 GB | $0.021 | 11.0 s | 15,584 | 38 s | 4,511 |
| RTX 3090 24 GB | $0.054 | 5.6 s | 11,905 | 16 s | 4,167 |
| RTX 4070 12 GB | $0.037 | 6.5 s | 14,969 | 21 s | 4,633 |
| RTX 4090 24 GB | $0.163 | 2.9 s | 7,616 | 7.5 s | 2,945 |
| RTX 5090 32 GB | $0.219 | 2.1 s | 7,828 | 5.2 s | 3,161 |
| RTX A4000 16 GB | $0.035 | 8.9 s | 11,557 | 29 s | 3,547 |
| L40S 48 GB | $0.257 | 3.4 s | 4,120 | 9.0 s | 1,556 |
| RTX 5070 Ti 16 GB | $0.065 | 4.8 s | 11,538 | 15 s | 3,692 |
Timings: community-typical SDXL 1024² @30 steps and Flux dev @20 steps; treat as ±20% and re-run on your workflow. Prices re-render live with every weekly sheet update.
Method: $/image, not $/hour
images_per_dollar = 3600 / seconds_per_image / price_per_hourThis single division reorders the whole market. The "cheap" hourly cards at the top of the table lose to the RTX 4090 the moment throughput enters the equation — and the gap widens on Flux, where small cards spill to system RAM.
SDXL economics
- Interactive sessions — an evening of prompting (~3 h) on a 4090: $0.98 on-demand. Use on-demand for sessions: an interruption mid-flow is worth more than the 50%.
- Batches — queue overnight on interruptible; per the table, roughly 7,616 images per dollar.
- Budget floor — the 3090 remains the best "always cheap" card: 24 GB means no workflow ever refuses to load.
Flux economics
- 24 GB is the entry ticket (FP8 + offloaded text encoders). Below that, generation works but throughput collapses — the img/$ column shows the 3060's honest number.
- 32 GB removes the ceiling — full-precision weights, LoRA stacks and upscalers resident: the 5090 leads img/$ despite the highest hourly rate on the table.
- Serving users? Concurrency changes the criteria — see serverless endpoints with the ComfyUI template.
Three traps that triple bills
- Re-downloading models every session. 30 GB of checkpoints at every boot is slow and silly. A 120 GB volume costs $9.60/mo and mounts in seconds.
- Idle instances "kept for later". Stop them — a stopped instance bills only disk. The pause button is the biggest discount on this page.
- Benchmarking with someone else's workflow. Your LoRA count, resolution and sampler move s/image by 3×. Rent two candidates for one hour ($0.766 total) and measure your own pipeline.
Put the numbers to work
Every price in this guide is our live rate — fixed, ≥30% under the market median, billed per second. Deploy the exact setup above from the console in about 30 seconds, paid in crypto, no card and no KYC.
