This GPU cloud pricing comparison is updated quarterly to reflect current market rates for renting GPU compute. Prices change frequently as new providers enter the market and hardware availability shifts. The figures below were last verified in April 2026.
GPU Cloud Pricing Comparison: H100 Instances
| Provider | GPU | VRAM | Price/hr (on-demand) | Notes |
|---|---|---|---|---|
| RunPod Secure Cloud | H100 SXM5 | 80 GB | ~$2.49 | Datacenter, high uptime |
| Lambda Labs | H100 SXM5 | 80 GB | ~$2.49–3.50 | Cluster available |
| CoreWeave | H100 SXM5 | 80 GB | ~$2.65–3.00 | Enterprise, NVLink |
| Vast.ai | H100 SXM5 | 80 GB | ~$1.89–2.20 | Host-dependent reliability |
| Google Cloud (on-demand) | H100 A3 | 80 GB | ~$5.50+ | Spot much cheaper |
| AWS P4d instances | A100 (via p4d) | 40 GB×8 | ~$32/hr (8-GPU) | $4/GPU/hr equiv. |
GPU Cloud Pricing Comparison: A100 Instances
| Provider | GPU | VRAM | Price/hr | Notes |
|---|---|---|---|---|
| RunPod | A100 SXM4 | 80 GB | ~$1.64 | Managed, reliable |
| Lambda Labs | A100 SXM4 | 80 GB | ~$1.99 | Premium datacenter |
| Vultr | A100 PCIe | 80 GB | ~$2.40 | US/EU regions available |
| Vast.ai | A100 (mixed) | 40/80 GB | ~$1.20–1.50 | Variable quality |
| DigitalOcean | H100 (new) | 80 GB | ~$3.99+ | Simplified management |
GPU Cloud Pricing Comparison: RTX 4090 Instances
| Provider | GPU | VRAM | Price/hr | Notes |
|---|---|---|---|---|
| RunPod Community Cloud | RTX 4090 | 24 GB | ~$0.74 | Consumer GPU, good uptime |
| Vast.ai | RTX 4090 | 24 GB | ~$0.30–0.45 | Cheapest available, variable |
| Hetzner | RTX 4000 Ada | 20 GB | ~$0.80 | EU datacenter, stable |
Best Price per VRAM: Which GPU Gives the Most Memory per Dollar?
For workloads that need raw VRAM (large model inference, multi-LoRA serving), the A100 80 GB at ~$1.64/hr on RunPod gives 80 GB for $1.64 — or $0.0205/GB/hr. The RTX 4090 at $0.74/hr gives 24 GB for $0.0308/GB/hr. The A100 wins on cost-per-GB for VRAM-heavy workloads.
H200 and B200 Pricing: The New Tier for 2026
The GPU cloud pricing landscape expanded significantly in 2026 with wider availability of H200 and Blackwell B200 instances. The H200 (141 GB HBM3e) is available on RunPod Secure Cloud at $3.59/hr and Lambda Labs at $3.29/hr. The B200 (192 GB HBM3e, FP4 support) appears on RunPod at $5.98/hr and Nebius at $5.50/hr. These higher-tier GPUs cost more per hour but deliver significantly more tokens per second for 70B+ inference workloads, which can result in lower effective cost per million tokens — particularly for teams running sustained high-volume endpoints. For a full cost-per-token analysis across H100, H200, and B200, see our H200 vs B200 vs H100 cost per token comparison.
Spot and Reserved Pricing: Reducing the Bill by 50–90%
Spot GPU instances represent the largest available discount in cloud compute. Spheron lists H100 SXM spot at $1.03/hr (versus $2.49 on-demand — a 59% discount) and B200 spot at $2.12/hr (versus $4.99 on-demand). GCP A3 spot instances list at $2.25/hr. Spot instances are preemptible with short notice, making them unsuitable for serving live inference endpoints but highly cost-effective for training runs with checkpoint support. Budget roughly 10% of your spot spend for jobs that require restart due to preemption — the math still strongly favors spot for training at any meaningful scale.
Reserved GPU contracts (1-year or 3-year) offer a middle path. H100 reserved pricing has dropped to approximately $1.70/hr across specialist providers. B200 reserved is available at $2.25/hr. Reserved contracts work best for teams with predictable baseline compute demand and at least 6 months of data to back the utilization forecast.
For a complete breakdown of which provider to choose based on your workload, see our best GPU VPS for AI guide. For head-to-head provider comparisons, see RunPod vs Vast.ai vs Lambda Labs.
How We Collected These Prices
All prices are sourced directly from provider pricing pages and verified via test account creation. Spot/preemptible prices are not included in the main tables — on-demand only. Prices are in USD and exclude storage costs. Last verified: June 2026.
Sources
- RunPod GPU Pricing
- Lambda Labs GPU Cloud
- Vast.ai Pricing
- Vultr Cloud GPU Pricing
- GPU Cloud Pricing Index 2026
- H100 vs H200 vs B200 GPU Rental Comparison — GPUaaS.com
Related reading: owned clusters vs cloud
When sustained utilization exceeds ~20%, on-premise economics often beat reserved cloud. See GPU cluster TCO 2026 and power infrastructure for US clusters. For an alternative to renting GPUs entirely, see custom AI silicon vs GPU rental 2026, covering TPU, Trainium, and Maia economics.
Author: Iovanny Olguín Ávila
Computer Systems Engineer with an MSc in Computer Science. I apply quantitative analysis and data-driven methodologies to evaluate financial instruments, investment vehicles, and emerging technologies. My technical background allows me to cut through marketing language and analyze the actual mechanics of financial products — from HELOC structures to Medicare Advantage plan design to business credit card reward algorithms.