Live per-hour pricing across Vast.ai, Salad, Terra Compute, and Hashrate — from RTX 3060 to B200. Updated July 14, 2026.
| GPU | VRAM | Vast.ai | Salad | Terra Compute | Hashrate |
|---|---|---|---|---|---|
| RTX 5090 | 32GB GDDR7 | $0.42–$0.48 | $0.25 | — | Listed |
| RTX 4090 | 24GB GDDR6X | $0.35–$0.39 | $0.11–$0.18 | — | Listed |
| RTX 3090 Ti | 24GB GDDR6X | $0.25–$0.35 | $0.10–$0.12 | — | Listed |
| RTX 3090 | 24GB GDDR6X | $0.20–$0.30 | $0.09–$0.10 | — | Listed |
| RTX 3080 | 10GB GDDR6X | ~$0.12 | $0.06 | — | Listed |
| RTX 3070 | 8GB GDDR6 | ~$0.08 | $0.05 | — | Listed |
| RTX 3060 | 12GB GDDR6 | ~$0.08 | $0.05–$0.08 | — | Listed |
| RTX 4070 Ti Super | 16GB GDDR6X | ~$0.15 | $0.07–$0.10 | — | Listed |
| RTX 5070 Ti | 16GB GDDR7 | ~$0.18 | $0.07–$0.17 | — | Listed |
| RTX 4080 | 16GB GDDR6X | $0.20–$0.28 | $0.08–$0.12 | — | Listed |
| RTX 4060 Ti 16GB | 16GB GDDR6 | ~$0.10 | $0.06–$0.08 | — | Listed |
| H100 SXM | 80GB HBM3 | $0.90–$2.27 | — | Multi-GPU | Listed |
| H200 | 141GB HBM3e | $2.15–$6.00 | — | Multi-GPU | Listed |
| B200 | — | $4.75–$5.08 | — | — | Listed |
| B300 | — | $5.85–$6.30 | — | — | Listed |
| A100 80GB | 80GB HBM2e | $0.75–$4.00 | — | Multi-GPU | Listed |
| V100 32GB | 32GB HBM2 | ~$0.50–$1.00 | — | — | Listed |
| Tesla P40 | 24GB GDDR5 | ~$0.15–$0.25 | — | — | Listed |
Best price highlighted in green. Hashrate is a comparison index, not a provider. Terra Compute sells bare-metal server hosting at $0.075/kWh. Prices fluctuate — check each provider for current rates.
| Hardware | Upfront | Rental Equivalent | Break-Even |
|---|---|---|---|
| Used RTX 3090 24GB | $700 | $0.20–$0.30/hr on Vast.ai | 2–4 months |
| Dual RTX 3090 48GB | $1,700 | $0.40–$0.60/hr | 3–5 months |
| Mac Studio M3 Ultra 96GB | $4,199 | $0.50–$0.90/hr (H100) | 8–14 months |
| Mac Studio M3 Ultra 192GB | $7,499 | $1.50–$2.27/hr (H100) | 5–8 months |
| 4× Tesla P40 96GB | ~$1,200 | $0.60–$1.00/hr | 2–3 months |
Assumes 24/7 operation. Break-even = months until owning costs less than renting the equivalent GPU. After break-even, you pay only electricity ($12–$50/mo depending on hardware).
| VRAM | Example GPUs | Max Model Size (Q4) | Example Models |
|---|---|---|---|
| 8–12 GB | RTX 3060, 3070, 4060 Ti | 8B | Llama 3 8B, Mistral 7B, Gemma 7B |
| 16 GB | RTX 4070 Ti, 4080, 5070 Ti | 13B | CodeLlama 13B, Mistral 13B, Qwen 14B |
| 24 GB | RTX 3090, 4090, 5090, P40 | 27B Q4 / 70B Q3 | Qwen 3 27B, GPT-OSS 20B, Llama 3.3 70B Q3 |
| 48 GB | Dual 3090, A6000 | 70B Q4 | Llama 3.3 70B, DeepSeek V3 Q4, Qwen 3 72B |
| 80–96 GB | H100, A100, Mac Ultra 96GB | 120B Q4 | DeepSeek V3, Llama 4, Nemotron |
| 192 GB | Mac Ultra 192GB, 4× P40 | 200B+ Q3 | Any model that fits in memory |
For 24/7 inference on 27B-class models, buy a used RTX 3090 ($700) — it pays for itself in 2–4 months vs. renting. For variable workloads, rent on Salad (cheapest consumer GPUs). For 70B+ models, rent H100 on Vast.ai or buy dual 3090s ($1,700). For silent 24/7 operation, Mac Studio M3 Ultra 96GB ($4,199).