GPU cloud pricing is chaotic. Prices differ by 10x across providers for the same GPU, spot markets can collapse overnight, and hyperscalers charge multiples of what specialized providers do. To cut through the noise, we pulled live data from 4,969 GPU instances across 18 providers — updated every 6 hours — and compiled every stat that actually matters for someone renting compute in 2026.
All numbers below reflect current market rates as of March 2026. Use the live comparison tool to see exact prices in real time.
Market Overview: Key Numbers
Spot vs On-Demand: The 62% Gap
Of the 4,969 instances tracked, 2,095 are spot instances (42%) and 2,874 are on-demand (58%). The average spot price is $2.83/hr vs $7.39/hr for on-demand — a 62% saving. This gap is significantly wider than AWS or GCP spot discounts (typically 60-90% below on-demand for EC2, but the base price is already higher).
Spot instances on GPU clouds like Vast.ai and RunPod are interruptible — the host can reclaim the GPU with limited notice. For fault-tolerant training jobs with checkpointing, this is the most cost-effective option in the market. For production inference, stick to on-demand.
| Type | Instances | Avg Price | Best Use Case |
|---|---|---|---|
| Spot / Interruptible | 2,095 | $2.83/hr | Training, batch jobs, experimentation |
| On-Demand / Reserved | 2,874 | $7.39/hr | Production inference, SLA-required workloads |
Price by GPU Model: Complete Market Data
The table below shows single-GPU pricing only (multi-GPU instances excluded to avoid distorting medians). Prices are current as of March 2026.
| GPU | Listings | Spot From | On-Demand From | Median |
|---|---|---|---|---|
| NVIDIA B200180GB | 23 | $1.67/hr | $3.60/hr | $4.99/hr |
| NVIDIA H200141GB | 46 | $0.33/hr | $1.99/hr | $2.29/hr |
| NVIDIA H10080GB | 70 | $0.80/hr | $1.58/hr | $2.59/hr |
| NVIDIA A10040–80GB | 103 | $0.08/hr | $0.09/hr | $1.29/hr |
| NVIDIA L40S48GB | 81 | $0.26/hr | $0.86/hr | $1.86/hr |
| NVIDIA RTX 509032GB | 34 | $0.13/hr | $0.33/hr | $0.65/hr |
| NVIDIA RTX 409024GB | 34 | $0.17/hr | $0.34/hr | $0.34/hr |
| NVIDIA RTX 309024GB | 22 | $0.05/hr | $0.22/hr | $0.22/hr |
| NVIDIA A1024GB | 372 | $0.08/hr | $0.09/hr | $1.20/hr |
| NVIDIA T416GB | 543 | $0.07/hr | $0.34/hr | $0.54/hr |
Provider Landscape: 18 Providers, Wildly Different Prices
The 18 providers we track range from hyperscalers (AWS, GCP, Azure) to specialized GPU clouds (Vast.ai, RunPod, Lambda Labs) to newer entrants (Verda, CloudRift). The price gap between the cheapest and most expensive provider for the same GPU routinely exceeds 10x.
| Provider | Instances | Cheapest GPU | Category |
|---|---|---|---|
| Vast.ai | 128 | $0.01/hr | Marketplace |
| Verda | 86 | $0.05/hr | Specialized |
| RunPod | 881 | $0.07/hr | Specialized |
| Azure | 448 | $0.07/hr | Hyperscaler |
| Vultr | 115 | $0.06/hr | Cloud |
| GCP | 2026 | $0.12/hr | Hyperscaler |
| Hyperstack | 9 | $0.15/hr | Specialized |
| Cudo Compute | 6 | $0.24/hr | Specialized |
| Crusoe | 11 | $0.40/hr | Specialized |
| CloudRift | 43 | $0.39/hr | Specialized |
| Lambda Labs | 207 | $0.50/hr | Specialized |
| OCI | 320 | $0.64/hr | Hyperscaler |
| AWS | 619 | $0.10/hr | Hyperscaler |
| Nebius | 32 | $1.25/hr | Cloud |
| CoreWeave | 8 | $6.50/hr | Specialized |
The Hyperscaler Tax: How Much More AWS/GCP/Azure Charge
The H100 is the clearest example of hyperscaler pricing vs. the market. AWS's cheapest single H100 is $1.16/hr — but that's because they run competitive spot pricing. Verda offers H100s from $0.80/hr. For on-demand H100, AWS on-demand pricing sits above $4/hr for the p4d/p5 instances. Lambda Labs charges $2.49/hr. Scaleway charges $3.22/hr. The spread is massive.
This pattern holds across GPU models. The hyperscaler premium is primarily a function of:
- Enterprise SLAs: Uptime guarantees, compliance certifications (SOC 2, HIPAA, PCI), and dedicated support cost money.
- Ecosystem lock-in pricing: AWS/GCP/Azure price GPU instances knowing you're already paying for their storage, networking, and managed services.
- Market inertia: Enterprise procurement teams default to known vendors. The alternative providers compete on price to win new customers.
Price Ranges by Workload Type
Different workloads have different GPU requirements, which translates to very different price ranges. Here's the practical breakdown:
| Workload | Recommended GPU | Price Range |
|---|---|---|
| 7B model inference (quantized) | RTX 3090 / T4 | $0.05–$0.34/hr |
| 7B–13B model inference (fp16) | RTX 4090 / RTX 5090 | $0.13–$0.59/hr |
| 13B–34B model inference / fine-tuning | L40S / A100 40GB | $0.26–$2.38/hr |
| 70B model inference / fine-tuning | A100 80GB / H100 | $0.08–$2.99/hr |
| Large-scale pre-training | H100 / H200 / B200 | $0.80–$5.29/hr |
| Stable Diffusion / image generation | RTX 3090 / RTX 4090 | $0.05–$0.59/hr |
Key Takeaways
- Spot instances save 62% on average vs on-demand across the full market. For training jobs with checkpointing, spot is almost always the right choice.
- The RTX 4090 is one of the most cost-effective GPUs in the market at $0.17–$0.34/hr per GPU, handling 7B–13B inference at a fraction of A100 pricing.
- The A100 remains the value king for training. At $0.08/hr spot, it offers more VRAM-per-dollar than almost anything on the market for training workloads that fit in 80GB.
- H200 spot ($0.33/hr) is dramatically underpriced given its 141GB VRAM and 4.8 TB/s bandwidth. Check availability before defaulting to H100.
- Hyperscalers charge 3–10x more than specialized providers for the same GPU. The premium is justified only if you need enterprise SLAs or deep integration with their managed services.
All data is sourced from GPU Tracker's live price feed, updated every 6 hours. See current prices across all providers and GPU models at gputracker.dev.
