Cloud GPU Comparison: How to Compare Per-Hour Prices Across Providers and GPU Models

Compare cloud GPU prices on a per-GPU-per-hour basis, and only across the same pricing type. On GPUPerHour's live snapshot (last stock observation 2026-10-01 13:40 UTC), the cheapest secure on-demand listings ran from $0.44/hr for an RTX A6000 to $7.89/hr for a B300 SXM6, with H100 SXM5 at $2.27/hr and A100 SXM4 80GB at $0.67/hr. Those numbers are only comparable if you normalise multi-GPU instances, exclude spot and peer-to-peer offers, and check that the provider still has stock.

Normalise every price to one GPU-hour

Providers quote instances, not GPUs. An 8×H100 node at $18.16/hr is $2.27 per GPU-hour — the same as a single-GPU listing at $2.27/hr. Divide the instance price by the GPU count before you compare anything.

Two other traps:

• Bundled vs. bare pricing. Some listings include CPU, RAM, and storage; others charge for them separately. A lower GPU rate can lose once you add the rest of the node.
• Per-hour vs. per-second billing. Short jobs benefit from per-second billing even at a higher headline rate. Check the minimum billing increment before assuming the cheaper hourly rate wins.

Compare like with like: pricing types are not interchangeable

GPUPerHour's current-price column uses only secure on-demand instances and excludes spot, reserved, and peer-to-peer marketplace listings, because their prices are not comparable. Keep that separation in your own comparison:

A $0.33/hr RTX 4090 on a marketplace host and a $0.74/hr RTX 4090 on RunPod are not the same product. The gap buys you isolation, uptime guarantees, and support — decide whether you need them before chasing the lower number.

Match the GPU to the workload before comparing prices

Price only matters after you know which class of GPU the job needs. The snapshot covers 14 popular GPUs; grouping them by typical use:

• Large-model training and fine-tuning: H100 SXM5 ($2.27/hr, Vast.ai), H100 PCIe ($2.59/hr, QuantaCloud), H100 NVL ($3.11/hr, Massed Compute), H200 SXM ($3.50/hr, Ori), H200 NVL ($3.43/hr, QuantaCloud), B300 SXM6 ($7.89/hr, RunPod). These carry the highest rates and the thinnest stock.
• General training and inference: A100 SXM4 80GB ($0.67/hr, Vast.ai), A100 PCIe 80GB ($0.68/hr, LeaderGPU), L40S ($0.97/hr, Massed Compute), MI300X ($3.39/hr, Hot Aisle).
• Inference, fine-tuning small models, dev work: RTX PRO 6000 Blackwell MIG 24GB ($0.59/hr, RunPod), RTX 5090 ($0.53/hr, Vast.ai), RTX 4090 ($0.74/hr, RunPod), RTX A6000 ($0.44/hr, LeaderGPU).

If your model fits on an A100, comparing H100 rates is wasted effort — the A100 tier is roughly a quarter of the price per hour in this snapshot.

Check stock, because the cheapest listing may already be gone

The snapshot shows how few providers carry each GPU: 3 providers with H100 SXM5 stock, 2 with H200 SXM, 1 with B300 SXM6, 1 with MI300X. Availability can change before booking, and GPUPerHour only counts offers marked available by a provider with live stock tracking, observed within 15 minutes of the summary's last refresh.

Practical sequence:

1. Filter to your GPU model and pricing type.
2. Note the cheapest two or three providers, not just the first.
3. Confirm the instance is actually bookable at that rate before you build a plan around it.
4. Keep a fallback provider — with single-digit provider counts per model, one sellout removes your cheapest option.

Use 30-day history to judge whether now is a good time

The current price alone doesn't tell you if you're overpaying. GPUPerHour publishes a 30-day historical catalog average alongside each current rate, and the two often diverge:

One caveat on that column: historical averages include all catalog listings — including unavailable and spot offers — and are medians of daily catalog averages over 30 days. They do not measure historical live prices or savings, so treat them as a rough baseline, not a target you could have hit.

A worked comparison

Say you need 200 GPU-hours of fine-tuning on a model that fits in 80GB.

• A100 SXM4 80GB at $0.67/hr → about $134.
• H100 SXM5 at $2.27/hr → about $454.
• RTX 4090 at $0.74/hr → about $148, but 24GB of VRAM may not fit the model.

The A100 and H100 differ by roughly 3.4× in cost. If the A100 finishes the job in a comparable wall-clock time, the H100 premium buys nothing. If the H100 cuts training time enough to matter, the gap narrows. Run that arithmetic before defaulting to the newest silicon.

Where to pull the numbers

GPUPerHour tracks 19,669 offers across 31 providers, with 690 available at the snapshot time, and refreshes the summary at most every 5 minutes. Daily history and CSV downloads are available through its Cloud GPU Price Index. Because the page states its methodology — normalised per GPU per hour, secure on-demand only for current prices — you can reproduce the comparison yourself rather than trusting a single headline rate.