Cloud GPU Price Tracker: Cheapest Per-Hour Rates by GPU Model and Provider

Updated

As of the latest snapshot (2026-10-01 13:40 UTC), the cheapest secure on-demand cloud GPU rates tracked by GPUPerHour start at $0.44/hr for an RTX A6000 (LeaderGPU) and $0.53/hr for an RTX 5090 (Vast.ai), while datacenter-class H100 SXM5 sits at $2.27/hr (Vast.ai) and H200 NVL at $3.43/hr (QuantaCloud). These figures cover only secure on-demand instances — spot, reserved, and peer-to-peer marketplace listings are tracked but excluded from the current-price column because their pricing models aren't directly comparable. Use the table below to pick a GPU by model, then check availability before booking, since stock can change between refreshes.

Current Cheapest Rates by GPU Model

All prices are per GPU-hour, secure on-demand only, from the GPUPerHour live tracker (31 providers, 19,669 tracked offers, 690 available at snapshot time).

GPU Model Cheapest Current Price Provider 30-Day Catalog Avg. Providers with Stock
H100 SXM5 $2.27/hr Vast.ai $2.18/hr 3
H100 PCIe $2.59/hr QuantaCloud $2.77/hr 5
H100 NVL $3.11/hr Massed Compute $1.91/hr 3
H200 SXM $3.50/hr Ori $2.70/hr 2
H200 NVL $3.43/hr QuantaCloud $2.88/hr 4
B300 SXM6 $7.89/hr RunPod $5.97/hr 1
A100 SXM4 80GB $0.67/hr Vast.ai $1.06/hr 7
A100 PCIe 80GB $0.68/hr LeaderGPU $1.69/hr 4
L40S $0.97/hr Massed Compute $0.90/hr 4
RTX PRO 6000 Blackwell MIG 24GB $0.59/hr RunPod $1.79/hr 5
RTX 5090 $0.53/hr Vast.ai $0.46/hr 2
RTX 4090 $0.74/hr RunPod $0.33/hr 3
RTX A6000 $0.44/hr LeaderGPU $0.57/hr 7
MI300X $3.39/hr Hot Aisle $2.75/hr 1

How to Read the Two Price Columns

The current price is the lowest secure on-demand offer observed within 15 minutes of the snapshot refresh. The 30-day catalog average is a median of daily catalog averages over the last 30 days and includes all pricing types and availability states — spot, reserved, and unavailable listings included. That means the two columns answer different questions:

  • Current price tells you what you can book right now, on-demand, at the lowest tracked rate.
  • Catalog average tells you whether today's rate is high or low relative to the model's recent range.

A few models show a meaningful gap worth noting:

  • H100 NVL: current $3.11/hr vs. 30-day avg $1.91/hr — today's cheapest on-demand rate is well above the recent catalog median.
  • RTX 4090: current $0.74/hr vs. avg $0.33/hr — the catalog average is pulled down by spot and P2P listings that aren't in the current-price column.
  • A100 SXM4 80GB: current $0.67/hr vs. avg $1.06/hr — today's on-demand rate is below the recent median.
  • B300 SXM6: current $7.89/hr vs. avg $5.97/hr, with only 1 provider holding stock — the highest-priced and least-available model in the snapshot.

Because the catalog average mixes pricing types, don't treat it as a "normal on-demand price." Use it only as a rough signal of whether the current on-demand offer is unusually high or low.

Why Spot, Reserved, and P2P Prices Are Excluded

GPUPerHour normalizes every listing to a price per GPU per hour (multi-GPU instance price ÷ GPU count) and tags it by pricing type and security tier. The current-price column filters to secure on-demand only because:

  • Spot instances can be interrupted and their effective cost depends on how long you actually hold them.
  • Reserved instances require upfront commitment, so the per-hour figure isn't comparable to pay-as-you-go.
  • Peer-to-peer marketplace hosts carry different reliability and security profiles.

If you're comparing headline rates across providers, mixing these categories will mislead you. The tracker keeps them in the catalog for historical context but out of the live comparison.

Choosing a GPU by Workload

Workload Reasonable starting point Why
Large-model training (70B+) H100 SXM5, H200 SXM/NVL, B300 High memory bandwidth and capacity; H200/B300 for the largest models
Mid-size training / fine-tuning A100 SXM4 80GB, H100 PCIe 80GB memory at $0.67–$2.59/hr; strong price-to-capability
Inference at scale L40S, A100 PCIe, RTX PRO 6000 MIG Lower cost per hour; MIG slices for small models
Memory-bound inference (large context) H200 NVL, MI300X Higher memory bandwidth per dollar than H100 in some cases
Prototyping / small models RTX 4090, RTX 5090, RTX A6000 $0.44–$0.74/hr; cheapest entry point
AMD ecosystem testing MI300X Only AMD option in the snapshot; $3.39/hr

For training runs that will last hours or days, the gap between the cheapest and second-cheapest provider for the same model can matter more than the model choice itself — check how many providers have stock before committing.

Verifying a Quote Before You Book

  1. Check the snapshot timestamp. The summary refreshes at most every 5 minutes, and current offers must be observed within 15 minutes of the last refresh. If the timestamp is stale, re-check.
  2. Confirm the provider still has stock. Availability can change before booking; the "providers with stock" column shows how thin the market is for each model.
  3. Confirm the pricing type. Make sure the offer you're about to book is secure on-demand, not spot or reserved, if you're comparing against the table above.
  4. Check the security tier. The tracker tags listings by security tier; peer-to-peer hosts are excluded from current prices but may appear elsewhere.
  5. For daily history and CSV downloads, use the Cloud GPU Price Index linked from the tracker.

Key Limitations

  • The snapshot covers 31 providers and 14 popular GPU models — it is not the entire market.
  • Historical figures are medians of daily catalog averages, not historical live prices, so they don't measure realized savings.
  • Missing days are excluded from averages, which can skew the 30-day figure for models with intermittent listings.
  • Only 690 of 19,669 tracked offers were available at snapshot time — most tracked listings are unavailable, so the cheapest rate for a given model may come from a single provider.