# SaladCloud Pricing

**Source:** https://salad.com/pricing — **Snapshot:** 2026-08-12 (from the H3 economics research pass)
**Billing:** per-hour **batch pricing** on a distributed fleet of privately owned consumer GPUs. This is the economic outlier — the cheapest lane in the package.

| GPU | $/hr (batch) |
| --- | ---: |
| RTX 3090 | ~0.124 |
| RTX 4090 | ~0.204 |
| RTX 5090 | ~0.294 |

Why so cheap: Salad is not a conventional cloud — it schedules containers onto
a network of machines its hosts run when idle. That's the same reason for the
caveats:

- **Availability is supply-driven** — a container group may wait for hosts;
  startup behavior differs from a dedicated GPU cloud.
- **Batch pricing** means spot-like behavior; treat jobs as retryable.
- The deal only works if you can say "give me any suitable 24–32GB GPU → run
  this container → return output → terminate" — Salad's container-groups REST
  API is built exactly for that, which is why the agent treats the provider as
  interchangeable.

Economic framing (with the H3 anchors, see docs/h3-economics.md):

| GPU | Per finished 1-min H3 (4 × 15s) |
| --- | ---: |
| Salad 5090 $0.294 | **~$0.06** (~250 min per $15) |
| Salad 4090 $0.204 | ~$0.056 |
| Salad 3090 $0.124 | ~$0.08 (slow — ~39 min of GPU per minute of video) |

Treat the ~250-minutes-per-$15 number as the upper-bound model, not a
guaranteed production figure. **Verify live before committing spend.**