简体中文
EnglishHow to Compare Cloud GPU Prices Per Hour Across Providers and Models
Updated
Compare cloud GPU prices by normalizing every listing to a price per GPU per hour, then filtering to secure on-demand offers only. As of the latest snapshot (2026-10-01 13:40 UTC) on GPUPerHour, the cheapest tracked secure on-demand rates include H100 SXM5 at $2.27/hr (Vast.ai), H200 NVL at $3.43/hr (QuantaCloud), B300 SXM6 at $7.89/hr (RunPod), A100 SXM4 80GB at $0.67/hr (Vast.ai), RTX 5090 at $0.53/hr (Vast.ai), and RTX 4090 at $0.74/hr (RunPod). These figures only hold for the moment they were observed — availability and price can change before you book.
Normalize every quote to $/GPU-hour
Providers list prices in different units: per instance, per multi-GPU node, or per hour of wall time. To compare fairly, divide the instance price by the number of GPUs it contains.
- An 8×H100 node at $18.16/hr = $2.27 per GPU-hour.
- A single-GPU RTX 4090 at $0.74/hr = $0.74 per GPU-hour.
Without this step, a cheap-looking multi-GPU node can appear more expensive than it is, and a single-GPU listing can look artificially cheap next to a node price.
Compare only like-for-like pricing types
GPUPerHour tracks several pricing types but its current-price column uses only secure on-demand listings. Spot, reserved, and peer-to-peer marketplace listings are tracked but excluded from the headline comparison because their prices are not directly comparable.
| Pricing type | Included in current-price comparison? | Why |
|---|---|---|
| Secure on-demand | Yes | Same commitment, same security tier |
| Spot | No | Interruptible; price not comparable |
| Reserved | No | Committed term changes the effective rate |
| Peer-to-peer marketplace | No | Different security and reliability profile |
If you mix a spot price with an on-demand price, you are comparing different products. Decide first whether you need guaranteed availability (on-demand) or can tolerate interruption (spot), then compare within that category.
Current cheapest rates by GPU model
From the GPUPerHour market snapshot (secure on-demand, per GPU-hour):
| GPU | Cheapest current price | Provider | Historical catalog avg. (30 days) | 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 |
Use historical averages to judge whether a price is actually good
The historical catalog average is a median of daily catalog averages over the last 30 days, including all pricing types and availability states. It is not a measure of historical live prices or savings — it is a reference band.
- If the current price is below the historical average, the offer is cheap relative to the last 30 days of catalog listings (e.g., A100 SXM4 at $0.67/hr vs. $1.06/hr average).
- If the current price is above the historical average, you are paying a premium for current availability (e.g., H100 NVL at $3.11/hr vs. $1.91/hr average, B300 at $7.89/hr vs. $5.97/hr average).
A low current price with only one provider in stock is also a risk: if that provider sells out, the next cheapest option may be materially higher.
Check freshness and stock before you commit
GPUPerHour's methodology states that current offers must be:
- Marked available by a provider with live stock tracking.
- Observed within 15 minutes of the summary's last refresh.
- The summary refreshes at most every 5 minutes.
The snapshot above carries a last stock observation of 2026-10-01 13:40 UTC. Before booking, confirm the price and availability are still current — the site notes availability can change before booking. Daily history and CSV downloads are available via the Cloud GPU Price Index if you want to check trends rather than a single moment.
Match the GPU to your workload before optimizing price
The cheapest GPU is not always the right one. Use the model tier as a first filter:
- H100 / H200 / B300 / MI300X — large-model training and high-throughput inference; highest cost per hour.
- A100 80GB — mid-tier training and inference; strong price-to-memory ratio at current rates.
- L40S / RTX A6000 / RTX PRO 6000 — inference, rendering, and workstation-class workloads.
- RTX 4090 / RTX 5090 — fine-tuning, smaller inference, and dev/test; lowest cost per hour.
Once the model tier is fixed, compare providers within that tier using the normalized $/GPU-hour and the historical average as your reference.
Quick comparison checklist
- Convert every quote to $/GPU-hour (instance price ÷ GPU count).
- Filter to secure on-demand if you need guaranteed availability.
- Compare current price against the 30-day historical catalog average.
- Check how many providers have stock — a single-provider price is fragile.
- Verify the observation timestamp and re-check before booking.