# Vast.ai Pricing

**Source:** live marketplace API (`https://vast.ai/api/v0/bundles`) — **Snapshot:** 2026-08-12, cheapest on-demand offers
**Billing:** per-hour, marketplace-determined. You pay **GPU + storage + bandwidth** separately; `dph_total` is the real all-in hourly number.

Cheapest on-demand offers at snapshot time (per-machine, single GPU unless noted):

| GPU | Cheapest $/hr (dph_total) | Location |
| --- | ---: | --- |
| RTX 3090 | 0.016 | Texas, US |
| RTX 4090 | 0.136 | Texas / KR / CN |
| RTX 5090 | 0.213 | South Korea |
| L40S | 0.401 | Texas, US |
| H100 SXM | 0.990 | US |

**This is a marketplace — prices move constantly.** The snapshot is only a
baseline. The operative way to price Vast is the search at deploy time:

```bash
vastai search offers 'gpu_name=RTX 5090' --order 'dph_total asc' --limit 10 --raw
```

Gotchas that make cheap offers expensive:
- **Storage is billed hourly** (~$0.10–0.20/GB/month equivalent, prorated) and `dph_total` already includes the *offer's* storage — a 30GB disk on a 4090 can add 20–40% to the headline GPU price.
- **Interruptible offers** are cheaper but reclaimable — fine for retryable batches, dangerous for one-shot jobs.
- Watch **`dph` vs `dph_total`**: `dph` is GPU-only bait pricing; `dph_total` is what you actually pay.
- **Geolocation** matters for upload/download of weights and outputs.
- Some cheap offers are CPU-bound or have bad network; `dlperf` (their benchmark score) is a sanity filter.

**Verify live before committing spend** — always search before provisioning.