Last updated: May 2026
The cheapest VPS with GPU in May 2026 starts at $0.29/hr for RTX 4090 (Vast.ai marketplace), $0.34/hr for RTX 4090 on RunPod Community Cloud, $1.39/hr for A100 80GB, and $1.49/hr for H100 (spot). Hyperscalers (AWS, GCP, Azure) charge 2-3x the same hardware. This guide ranks the real options — Vast.ai, RunPod, Lambda Labs, Paperspace, SynpixCloud, TensorDock — with per-second-vs-hourly billing notes, on-demand vs spot pricing, and how to actually pick.
For free GPU options first, see Free GPU Compute. For buying hardware instead, see Best GPU for AI 2026.
VPS GPU pricing (May 2026, per hour, lowest published)
| GPU | Vast.ai | RunPod (Community) | RunPod (Secure) | Lambda | SynpixCloud |
|---|---|---|---|---|---|
| RTX 3090 24GB | $0.16 | n/a | n/a | n/a | n/a |
| RTX 4090 24GB | $0.29 | $0.34 | $0.59 | n/a | $0.39 |
| RTX 5090 32GB | $0.50-$1 | $0.69 | $0.99 | n/a | $0.79 |
| L40 48GB | $0.31 | $0.69 | $0.99 | n/a | n/a |
| A100 PCIe 40GB | $0.52 | $0.60 | $1.49 | $1.10 | $1.39 |
| A100 SXM 80GB | $0.67 | $0.79 | $1.99 | $1.29 | $1.39 |
| H100 80GB | $1.49 (spot) | $1.49 (spot) | $2.49 | $2.99 | $2.49 |
| H200 141GB | $2.50 | $2.99 | $3.79 | $3.79 | $3.49 |
| B200 192GB | $5-$8 | $5.50 | $6.50 | $5.99 | $6.49 |
Hyperscaler equivalents (for reference): AWS / Azure / GCP H100 on-demand is $5-$8/hr; B200 is $9-$12/hr. The specialist clouds are 2-3x cheaper for the same hardware.
Vast.ai — cheapest, marketplace model
Vast.ai is a marketplace where individual GPU hosts (homelabs, mining farms, small datacenters) list their hardware. Buyers bid; prices float based on supply. Result: the absolute cheapest GPU prices anywhere.
- Pricing: Spot from $0.024/hr (older GTX 1070), $0.16/hr (RTX 3090), $0.29/hr (RTX 4090), $1.49/hr (H100 spot).
- Billing: Per-second.
- Trade-offs: Reliability varies by host. Some are excellent, some are sketchy. Spot instances can be interrupted. Network throughput varies.
- Best for: Prototyping, batch training with checkpointing, hyperparameter sweeps, anything where cost matters more than uptime.
RunPod — best UX, per-second billing
RunPod is the convenience pick. Two tiers:
- Community Cloud: $0.34/hr RTX 4090, $0.60/hr A100 PCIe — peer-provisioned hosts (similar to Vast.ai but managed by RunPod).
- Secure Cloud: $0.59/hr RTX 4090, $1.99/hr A100 SXM — RunPod-owned hardware in proper datacenters.
Why teams choose RunPod:
- Per-second billing — short experiments don't get billed for full hour.
- Templates — one-click deploy of PyTorch, ComfyUI, Stable Diffusion WebUI, Ollama, etc.
- Persistent volumes — keep your data between sessions.
- API + CLI — automate spin-up / shutdown from CI.
- Reliable — Secure Cloud has datacenter-grade uptime.
Best for: Solo developers, small teams, anyone who wants GPUs to just work without per-instance reliability variance.
Lambda Labs — ML team default
Lambda Cloud targets ML engineers. Strengths:
- Pricing: H100 8x from $2.99/GPU-hr on-demand. Competitive but rarely the absolute cheapest.
- UX: Clean Jupyter / SSH workflow, strong docs for ML practitioners.
- Reserved instances for sustained workloads — reservation discounts make Lambda competitive with specialist clouds for long-running training.
Best for: Funded teams doing serious training where reliability + reserved pricing matters more than rock-bottom hourly rate.
SynpixCloud — newer competitor, sharp pricing
Newer entrant with aggressive on-demand pricing: A100 80GB at $1.39/hr, H100 at $2.49/hr. Per-hour billing.
Best for: Cost-conscious teams that want stable pricing without bidding (Vast.ai) or marketplace variance.
TensorDock — competitive bidding marketplace
Similar to Vast.ai with a competitive-bid model. Sometimes lists GPUs Vast.ai doesn't have in stock.
Best for: Backup option when Vast.ai inventory is tight.
Paperspace — high price, persistent storage
Paperspace's H100 VM is $5.95/GPU-hr with a 1-hour minimum — meaningfully more expensive than RunPod / Lambda. The strength is persistent storage and team workflows.
Best for: Teams that prioritize persistent storage + Notebook UX over raw hourly cost.
Hyperscaler GPU (AWS, GCP, Azure)
Same hardware, 2-3x the price. Reasons to still use them:
- Enterprise procurement. AWS / Azure contracts you already have.
- Deep IAM / VPC integration. GPUs that talk to your existing AWS resources without bridging.
- Spot capacity for specific instance types. Some workloads can't run on RunPod / Vast.ai due to hardware specifics.
For most teams that just need GPU rental as a workload, hyperscalers are needless overpay.
On-demand vs spot — the savings math
Spot / preemptible instances cost 50-95% less than on-demand but can be interrupted with minutes of notice. Vast.ai's spot model is the most aggressive (instances can be outbid and reclaimed). RunPod's spot tier offers 30-70% discount on Secure Cloud.
Use spot for:
- Training with frequent checkpointing.
- Hyperparameter sweeps (many short jobs).
- Batch inference where each request is independent.
- Anything you can resume after interruption.
Don't use spot for:
- Customer-facing inference.
- Long fine-tuning runs without checkpointing.
- Demos or live workloads with strict uptime needs.
Picking the right GPU for your workload
RTX 4090 24GB ($0.29-$0.69/hr): 7B-13B model fine-tuning, Stable Diffusion / video gen, embedding generation, most inference. The sweet spot for solo developers.
RTX 5090 32GB ($0.50-$1/hr): 30B at 4-bit dense, 70B MoE with low active params, big batch sizes. Worth the premium over 4090 when 24GB hits the wall.
A100 80GB ($1.39/hr+): 70B at 4-bit, 30B-40B at FP16, larger Stable Diffusion XL training, multi-LoRA serving.
H100 80GB ($1.49-$2.99/hr): 70B FP16, modern frontier-class fine-tuning, FP8 precision training, low-latency production inference.
H200 141GB ($2.50-$3.79/hr): When 80GB H100 isn't enough — 100B+ inference, longer-context training.
B200 192GB ($5-$8/hr): Frontier training, very large MoE serving. Overkill for 95% of workloads.
Billing models — read the fine print
| Provider | Billing |
|---|---|
| Vast.ai | Per-second |
| RunPod | Per-second |
| Lambda Labs | Per-second |
| Paperspace | 1-hour minimum |
| SynpixCloud | Per-hour |
| TensorDock | Per-hour |
| AWS / GCP / Azure | Per-second (most modern instance types) |
Per-second billing matters more than headline rate for short experimental runs. A 5-minute job on $5/hr per-second beats $1.50/hr with 1-hour minimum.
Common mistakes when renting GPU VPS
- Renting H100 when RTX 4090 would do. 7x cost premium for nothing. Right-size your GPU.
- Skipping checkpointing on spot. Spot interrupted = work lost. Always checkpoint.
- Forgetting egress. Some providers charge for outbound data; check before pulling 100GB of trained model weights.
- Hyperscaler-by-default. Unless you need AWS / Azure integration, you're overpaying 2-3x.
- Per-hour minimums. Avoid Paperspace's 1-hour minimum for short experiments.
- No storage budget. Persistent volumes cost money; clean up unused volumes.
Quick setup recipe (RunPod, 5 minutes)
- Sign up at runpod.io. Add $10 credits (or use referral link for $0).
- Pick a GPU and a pod template (PyTorch, Ollama, ComfyUI, etc).
- Launch. Pod boots in 30-60 seconds.
- Connect via SSH or web terminal.
- Run your workload. Stop the pod to stop billing.
Repeat the same flow on Vast.ai if cost beats UX.
