简体中文
EnglishHow to Compare Cloud GPU Pricing: Current Per-Hour Rates and What Drives Them
Updated
Cloud GPU pricing is quoted per GPU-hour, and the only fair comparison is between offers of the same GPU model, same pricing type, and same security tier. As of the latest snapshot on gpuperhour.com (stock observed 2026-10-01 13:40 UTC), the cheapest secure on-demand rates across 31 tracked providers range from $0.44/hr for an RTX A6000 to $7.89/hr for a B300 SXM6. Use the tables below to see where current prices sit against 30-day catalog averages, then check availability before you book — stock can change before an instance is provisioned.
Current cheapest secure on-demand prices by GPU
All figures are per GPU-hour, secure on-demand only. "Historical avg." is the median of daily catalog averages over the prior 30 days and includes all pricing types and availability states, so it is a reference point, not a savings estimate.
| GPU | 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 |
Two things stand out. First, the gap between current and historical average runs in both directions: the A100 SXM4 at $0.67/hr is well below its $1.06/hr 30-day median, while the B300 at $7.89/hr is above its $5.97/hr median. Second, "providers with stock" is often 1–3 for high-end parts, so the cheapest listing may not be the one you can actually get.
Why the same GPU has different prices
Form factor and memory variant
H100 appears as SXM5, PCIe, and NVL, priced at $2.27, $2.59, and $3.11/hr respectively. These are different physical cards with different interconnect and memory configurations, so they are not interchangeable for every workload. Compare within a variant, not across all "H100" listings.
Pricing type
The snapshot above uses only secure on-demand instances. Spot, reserved, and peer-to-peer marketplace listings are tracked separately and excluded from the current-price column because their terms are not comparable — spot can be interrupted, reserved locks you in, and marketplace hosts carry a different security tier. Mixing them into one ranking produces misleading "cheapest" claims.
Multi-GPU instances
Listings are normalised to price per GPU per hour by dividing the instance price by the GPU count. A $16/hr eight-GPU node is $2.00 per GPU-hour. Always confirm whether a quoted price is for the whole instance or a single GPU before comparing.
Supply and demand
Availability changes before booking, and the number of providers with stock is small for top-tier parts. A price observed 15 minutes ago may be gone; the summary refreshes at most every 5 minutes and only includes offers marked available by a provider with live stock tracking.
How to pick a GPU for your workload
| Workload | Typical choice | Why |
|---|---|---|
| Large-model training | H100 SXM5, H200 SXM, B300 SXM6 | Highest memory bandwidth and interconnect; expect $2.27–$7.89/hr |
| Fine-tuning / mid-size training | A100 SXM4 80GB, H100 PCIe | 80GB memory at $0.67–$2.59/hr; good balance of capacity and cost |
| Inference at scale | L40S, A100 PCIe, RTX PRO 6000 MIG | $0.59–$0.97/hr; MIG slices suit smaller models |
| Rendering / lighter GPU work | RTX 4090, RTX 5090, RTX A6000 | $0.44–$0.74/hr; consumer and workstation cards |
| AMD-based workloads | MI300X | $3.39/hr; only one provider had stock in this snapshot |
Match the GPU to the memory your model needs first, then optimise on price. A cheaper card that forces you to shard across instances often costs more per hour in total.
Steps to compare and rent by the hour
- Fix your requirements. Decide the GPU variant, minimum VRAM, and whether you need a single GPU or a multi-GPU node.
- Filter to comparable offers. Keep only secure on-demand listings of the same variant. Exclude spot, reserved, and peer-to-peer unless you specifically want those terms.
- Check the price basis. Confirm the quote is per GPU-hour, not per instance-hour, and note the GPU count if it is a multi-GPU node.
- Compare against history. Use the 30-day catalog average as a sanity check — a current price far above the median may reflect scarcity rather than value.
- Verify stock. Check the "providers with stock" count and re-check immediately before booking, since availability can change.
- Confirm billing unit and terms. Before you commit, verify the billing increment (per hour, per minute, or per second), the security tier, and whether the provider's live stock tracking is current.
Where the numbers come from
gpuperhour.com scrapes the public pricing and inventory endpoints of 31 cloud GPU providers periodically. Each listing is normalised to a price per GPU per hour and tagged with its pricing type and security tier. Current prices include only secure on-demand listings observed within 15 minutes of the summary's last refresh; the summary refreshes at most every 5 minutes. Historical figures are medians of daily catalog averages over 30 days, including all pricing types and availability states, with missing days excluded. Daily history and CSV downloads are available via the site's Cloud GPU Price Index.