GPU rental & training
Per-GPU-hour + per-second serverless. Use when self-hosting an open-weights model beats a managed endpoint.
23 options
provider
| provider | gpu | kind | $/hr ↑ | spot | bill | region | notes |
|---|---|---|---|---|---|---|---|
| Vast.ai | H100 80GB | marketplace | $1.65 | $0.34 | per-second | global | Market floor. Reliability varies by host. |
| Spheron / GMI | H100 80GB | marketplace | $1.03 | $0.47 | per-second | global | Decentralized GPU marketplace. |
| Spheron / GMI | A100 80GB | marketplace | $1.07 | $0.60 | per-second | global | |
| Vast.ai | A100 80GB | marketplace | $2.00 | $0.67 | per-second | global | |
| RunPod | H100 SXM 80GB | cloud | $2.69 | $1.19 | per-second | global | |
| Lambda | A100 80GB | cloud | $1.79 | — | per-minute | us+eu | |
| AWS | A100 80GB (p4de) | cloud | $4.10 | $1.85 | per-second | global | |
| RunPod | A100 80GB | cloud | $1.89 | — | per-second | global | |
| SF Compute | H100 80GB | cluster | $1.96 | — | per-hour | us-sfo | Vetted clusters, sell-back unused. B300s announced for fall 2026. |
| RunPod | H100 PCIe 80GB | cloud | $1.99 | — | per-second | global | |
| Modal | A100 40GB | serverless | $2.10 | — | per-second | global | |
| Spheron / GMI | B200 | marketplace | $2.12 | — | per-second | global | |
| SF Compute | H200 80GB | cluster | $2.49 | — | per-hour | us-sfo | |
| Modal | A100 80GB | serverless | $2.50 | — | per-second | global | |
| AWS | H100 80GB (p5) | cloud | $4.10 | $2.50 | per-second | global | Was $6.88; repriced mid-2025. $1.90/hr with 3-yr SP. |
| CoreWeave | H100 80GB (8x HGX) | cloud | $6.16 | $2.83 | per-hour | us | Per-GPU price in 8x node; $49.24/hr for full node |
| Lambda | H100 SXM 80GB | cloud | $2.99 | — | per-minute | us+eu | Reserved 1-yr: $1.89/hr |
| Modal | H100 80GB | serverless | $3.95 | $3.95 | per-second | global | Preemptible base; non-preempt ~3.75×. Includes Modal Sandbox runtime. |
| CoreWeave | B200 (8x) | cloud | $8.60 | $3.96 | per-hour | us | |
| Vast.ai | B200 | marketplace | $4.50 | — | per-second | global | |
| RunPod | B200 | cloud | $4.99 | — | per-second | us | |
| Lambda | B200 | cloud | $4.99 | — | per-minute | us | Reserved 1-yr: $3.79/hr |
| Replicate (raw GPU) | H100 80GB (8x) | serverless | $5.49 | — | per-second | us | Per-GPU in 8x cluster; $43.92/hr full node. |
Serverless GPU (per-second)
Pay only while your function runs.
| Provider | GPU | $/sec ↑ | $/hr equiv | Notes |
|---|---|---|---|---|
| RunPod Serverless | H100 80GB | $0.00078 | $2.81 | |
| fal | H100 80GB | $0.00079 | $2.84 | fal serverless functions; pay only while function executes. |
| Modal Function | H100 80GB | $0.00110 | $3.96 | Preemptible base rate. |
| Modal Sandbox | H100 80GB | $0.00110 | $3.96 | Managed code-exec sandbox with optional GPU attached. |
| Replicate | H100 80GB | $0.00153 | $5.51 |
Training cost calculator
Estimate end-to-end LoRA training cost. Single H100, batch 1, standard config.
Estimated wall-clock
0.28 hrs
| Rank | Provider | GPU | Rate | Total |
|---|---|---|---|---|
| 1 | Vast.ai | H100 80GB | $0.34/hr | $0.09 |
| 2 | Spheron / GMI | H100 80GB | $0.47/hr | $0.13 |
| 3 | RunPod | H100 SXM 80GB | $1.19/hr | $0.33 |
| 4 | SF Compute | H100 80GB | $1.96/hr | $0.54 |
| 5 | RunPod | H100 PCIe 80GB | $1.99/hr | $0.55 |
| 6 | AWS | H100 80GB (p5) | $2.50/hr | $0.69 |
| 7 | CoreWeave | H100 80GB (8x HGX) | $2.83/hr | $0.79 |
| 8 | Lambda | H100 SXM 80GB | $2.99/hr | $0.83 |