“From $X per GPU-hour” is the number everyone quotes for the big clouds, usually without saying which machine you have to rent to get it. I added two clouds whose price lists can be read without an account — Azure and Oracle Cloud — to the rental prices we already collect hourly, and kept the machine attached to every number.
One reading, 2 Oct 2026, 13:00 UTC, USD per GPU-hour:
| GPU | Vast.ai median (offers) | RunPod from | Azure from | Oracle Cloud from |
|---|---|---|---|---|
| A100 | 0.74 (58) | 1.39 | 3.40 — ND96asr v4, 8 × 40 GB, VM 27.20/h | 3.05 — BM.GPU4.8, 8 × 40 GB, machine 24.40/h |
| H100 | 2.27 (8) | 2.69 | 6.98 — NC40ads H100 v5, 1 × 94 GB | 10.00 — BM.GPU.H100.8, 8 × 80 GB, machine 80.00/h |
| H200 | 5.12 (7) | 3.59 | 10.60 — ND96isr H200 v5, 8 × 141 GB, VM 84.80/h | 10.00 — BM.GPU.H200.8, 8 × 141 GB, machine 80.00/h |
| B200 | 7.81 (6) | 5.98 | not listed as its own size | 14.00 — BM.GPU.B200.8, 8 × 180 GB, machine 112.00/h |
Three things the table shows at this reading:
- The per-GPU price hides the unit of sale. Azure’s H100 at 6.98 is a one-card VM you can take alone. Oracle’s 10.00 exists only as an eight-card bare-metal machine at 80.00 an hour. For the A100 both clouds’ lowest per-GPU price sits in an eight-card machine with the 40 GB card; a single 80 GB A100 on Azure lists at 3.67.
- The gap to the marketplace is larger on the two older cards. List price against the Vast.ai median of the same hour: A100 4.6× (Azure) and 4.1× (Oracle), H100 3.1× and 4.4×; H200 about 2× for both, B200 1.8× (Oracle). One reading, and the marketplace side is thin for the newer classes (six to eight offers) — a description of today, not a trend.
- The two clouds do not rank the same way per class. Oracle is below Azure on the A100 and the H200, above it on the H100.
What it does not say: availability (a list price is not free capacity), what you pay with a commitment or a negotiated rate, or that a cloud VM and a marketplace host are the same product — they are not; the table compares prices, nothing else.
How the numbers are made: Azure publishes whole-VM prices per region, so the per-GPU price is the Linux pay-as-you-go price divided by the GPU count from Microsoft’s size documentation, lowest across documented sizes and public regions; sizes the price list carries but the documentation does not describe are left out rather than guessed. Oracle prices a GPU-hour directly and has one list without regions. Both are read four times a day and shown beside the marketplace medians, never blended into them.
The live table: GPU rental prices per hour: Vast.ai, RunPod, Clore.ai, updated hourly — AI Compute Radar — and the exact rules: Methodology — how every number is made — AI Compute Radar
If you rent on one of these clouds and the instance named here is not the one people actually use, tell me which one is — that is the kind of correction I can check against the price list.