GPU rental vs paying per generation
Rent the hardware and the per-generation markup disappears — but only if you keep the GPU busy and manage the pipeline yourself. GPU rental itself comes in two very different flavours, priced differently, so we show them separately. Background: when does renting a GPU make sense?
Managed cloud — RunPod (official fixed prices)
Fixed prices set by RunPod. Secure Cloud = vetted datacenters; Community Cloud = individual hosts, cheaper but variable.
| GPU | VRAM | Secure Cloud ($/hr) | Community ($/hr) |
|---|---|---|---|
| RTX 2000 Ada | 16 GB | $0.24 | $0.50 |
| RTX A4000 | 16 GB | $0.25 | $0.17 |
| RTX A4500 | 20 GB | $0.25 | $0.19 |
| RTX A5000 | 24 GB | $0.27 | $0.16 |
| RTX 4000 Ada | 20 GB | $0.28 | $0.20 |
| A40 | 48 GB | $0.44 | $0.35 |
| RTX 4000 Ada SFF | 20 GB | $0.44 | $0.18 |
| RTX 3090 Ti | 24 GB | $0.46 | $0.27 |
| L4 | 24 GB | $0.49 | $0.44 |
| RTX 3090 | 24 GB | $0.50 | $0.22 |
| RTX 4070 Ti | 12 GB | $0.50 | $0.19 |
| RTX 4080 | 16 GB | $0.50 | $0.27 |
| RTX 4080 SUPER | 16 GB | $0.50 | $0.28 |
| RTX A2000 | 6 GB | $0.50 | $0.12 |
| RTX PRO 6000 MaxQ | 96 GB | $0.50 | $1.64 |
| RTX A6000 | 48 GB | $0.53 | $0.33 |
Marketplace — Vast.ai (typical market rate)
Live host-set prices, filtered to VERIFIED DATACENTER hosts and on-demand tier only. Interruptible instances run ~50% cheaper but can be paused anytime. Storage and bandwidth billed separately by host.
| GPU | Typical market rate ($/hr) |
|---|---|
| Tesla P100 | $0.11 |
| RTX A4000 | $0.16 |
| RTX 5060 Ti | $0.19 |
| RTX PRO 4000 | $0.30 |
| RTX 5070 Ti | $0.40 |
| RTX 4090 | $0.64 |
| RTX 5090 | $0.76 |
| A100 SXM4 | $0.80 |
| RTX PRO 5000 | $0.89 |
| RTX PRO 6000 S | $1.74 |
| RTX PRO 6000 WS | $1.87 |
| H100 SXM | $2.93 |
| H200 | $4.34 |
| H200 NVL | $4.91 |
| B200 | $6.00 |
How to read this: a marketplace has no price list — individual listings appear and disappear by the minute, and hosts price by their hardware specs. The number above is the typical going rate for verified datacenter hosts (on-demand tier), a slow-moving statistic we refresh several times a day — use it to budget, not to shop. When actually renting, filter listings on Vast.ai by your required specs and by DLPerf (their performance-per-dollar score). Why two tables? Marketplace host quality varies wildly — unverified listings can look 3–4× cheaper than they effectively are; a managed cloud publishes one fixed price list and manages the hardware for you.
Break-even calculator
Compare your own GPU cost per generation against the cheapest tracked platform price.
This compares marginal compute cost only — your setup time, storage and bandwidth fees, and failure handling aren't priced in. Rented idle time is pure loss, which the overhead multiplier approximates.