GPU rental prices have not converged. The same card can cost several times more depending on which cloud you rent it from, and the cheapest option changes through the day. We run FastGPU, which pulls prices from 28 GPU clouds (marketplaces, neoclouds, serverless platforms and the hyperscalers) and ranks them cheapest first.
Here is the market as it stood on September 27, 2026 at 04:11 UTC, and the few patterns that decide whether you overpay. Prices move daily, so every figure links to its live page.
The snapshot
Cheapest on-demand price per GPU-hour for each card, next to the cheapest on-demand price for the same card at a hyperscaler (AWS, Azure, Google Cloud or Oracle). Every figure in a row comes from that card's live page.
| GPU | VRAM | Cheapest on-demand | Offered by | Cheapest hyperscaler | Gap (hyperscaler / cheapest) | Providers |
|---|---|---|---|---|---|---|
| H100 SXM | 80 GB | $1.79/hr | Cudo Compute | $5.38/hr (Google Cloud) | 5.38 / 1.79 = 3.0x | 22 |
| H200 | 141 GB | $2.60/hr | GMI Cloud | $7.91/hr (AWS) | 7.91 / 2.60 = 3.0x | 18 |
| B200 | 192 GB | $3.69/hr | DeepInfra | $11.28/hr (Google Cloud) | 11.28 / 3.69 = 3.1x | 16 |
| B300 | 288 GB | $4.89/hr | DeepInfra | $17.80/hr (AWS) | 17.80 / 4.89 = 3.6x | 12 |
| GB200 | 192 GB | $8.00/hr | GMI Cloud | $16.00/hr (Oracle Cloud) | 16.00 / 8.00 = 2.0x | 3 |
| A100 80GB | 80 GB | $0.89/hr | DeepInfra | $3.22/hr (Google Cloud) | 3.22 / 0.89 = 3.6x | 18 |
| L40S | 48 GB | $0.47/hr | Vast.ai | $1.86/hr (AWS) | 1.86 / 0.47 = 4.0x | 13 |
| MI300X | 192 GB | $1.71/hr | TensorWave | none listed | n/a | 4 |
| RTX PRO 6000 | 96 GB | $0.93/hr | Vast.ai | $0.97/hr (Google Cloud) | 0.97 / 0.93 = 1.0x | 11 |
| L4 | 24 GB | $0.11/hr | Vast.ai | $0.71/hr (Google Cloud) | 0.71 / 0.11 = 6.5x | 7 |
| RTX 4090 | 24 GB | $0.14/hr | Vast.ai | none listed | n/a | 4 |
| RTX 5090 | 32 GB | $0.21/hr | Vast.ai | none listed | n/a | 3 |
1. Hyperscalers charge 3.0x to 3.6x more for the same flagship silicon
For the data-center flagships, the cheapest hyperscaler rate is 3.0x the floor on H100 SXM ($5.38/hr vs $1.79/hr), 3.1x on B200 ($11.28/hr vs $3.69/hr) and 3.6x on B300 ($17.80/hr vs $4.89/hr). The widest gap in the table is L4 at 0.71 / 0.11 = 6.5x. You pay the hyperscaler premium for the ecosystem (IAM, networking, managed services, committed-use discounts, startup credits), which can be worth it, but it is a choice rather than a default. The exception is the new workstation-class Blackwell card: an RTX PRO 6000 starts at $0.93/hr (Vast.ai) and costs $0.97/hr on Google Cloud, almost the same.
2. Spot capacity is cheaper again, with two catches
If your job checkpoints, interruptible capacity cuts the price further: H100 SXM from $0.98/hr, H200 from $1.84/hr, B200 from $1.77/hr. The first catch: spot can be reclaimed at any moment, so it suits checkpointed training and offline batch jobs, never a user-facing endpoint. The second: the cheapest listings are often whole nodes. In the open dataset, the H200 spot floor on AWS requires 8 GPUs and the B200 spot floor on Google Cloud requires 8 GPUs, so the real entry ticket is that many times the per-GPU price.
3. Blackwell is rentable on demand
B200 is listed by 16 providers from $3.69/hr (DeepInfra), B300 by 12 from $4.89/hr (DeepInfra), and GB200 by 3 from $8.00/hr (GMI Cloud). If you assumed the newest silicon was contact-sales only, much of it now has a public hourly price.
4. For small models, consumer cards are hard to beat
An RTX 4090 (24 GB) rents from $0.14/hr and an RTX 5090 (32 GB) from $0.21/hr on marketplace hosts. 24 GB is enough for a 7B model in fp16 or a 13B model quantized to 4-bit (how much VRAM an LLM needs). Marketplace hosts are individual machines, so check host reliability and keep anything you cannot lose off the box.
5. The floor moves, so check it when you rent
Marketplace prices change as hosts come and go, and a cheap listing can be rented out within the hour. Each GPU page carries a 30-day price strip showing the low, the typical price and whether now is a good moment to rent (for example H200 or L40S).
How to choose, in four steps
- Size the VRAM your model needs first. Everything else follows from it.
- Decide whether the job tolerates interruption. Checkpointed training can use spot; an endpoint cannot.
- Compare the landed cost, not the headline rate. Minimum GPU counts, egress, storage and cold-start billing can erase a cheap rate.
- Check availability. A price with no stock behind it is not a price.
If you would rather skip the spreadsheet, describe the workload and FastGPU ranks the live options that can actually run it.
Get the data yourself
The prices behind this post are an open dataset (CC BY 4.0), downloadable as JSON or CSV with no key:
curl -s https://fastgpu.co/api/v1/dataset/gpu-prices-current.json \
| jq '[.[] | select(.gpu_slug == "h100-sxm" and .offer_type == "on-demand")]
| sort_by(.price_usd_hr) | .[:5]
| map({provider: .provider_label, usd_per_gpu_hr: .price_usd_hr, min_gpus: .min_gpu_count})'
Daily history is in the same dataset, which also has a permanent DOI on Zenodo (10.5281/zenodo.22842387). If you are building an agent, the same data is exposed as an MCP server at https://fastgpu.co/api/mcp with list_gpu_prices and match_workload tools.
Method and disclosure
Prices are USD per GPU-hour, read from each provider's public API or pricing page and refreshed continuously. "Cheapest on-demand" counts on-demand and community listings, preferring ones that are not sold out; spot, reserved and serverless are shown separately. The hyperscaler reference is the cheapest on-demand offer for that card at AWS, Azure, Google Cloud or Oracle. Snapshot taken September 27, 2026 at 04:11 UTC; the linked pages show the current numbers. Some provider links on FastGPU are affiliate links; price tables and every "cheapest" figure here are ordered strictly by price.
Top comments (1)
plot twist: a green dashboard tile is not a signed usage tip.
1 cut: when the invoice fight starts, can a buyer GET a queryable meter of what ran, or only another compliance seal?
curiosity > decks.