What a GPU-hour costs
Publicly listed rental prices at 21 providers, normalised to a single GPU-hour, plus what the same cards cost to buy. Every figure links to the page it was read from.
In the last closed month (August 2026) the H100 SXM listed-rate index rose 2.6%. Over 12 months, +9%.
Size your workload
Describe the workload — we show who is cheapest today and what a month costs.
H100 SXM — every listed rate
| Provider | Type | Region | Vs. median | $ / GPU-hour |
|---|---|---|---|---|
| Cudo Compute | specialist | no-kristiansand-1, GPU onlycaptured · unchanged since first read, Sep 7, 2026 | -55% | $1.79 |
| GMI Cloud | specialist | fromcaptured · unchanged since first read, Sep 7, 2026 | -50% | $2.00 |
| Jarvislabs | specialist | globalcaptured · unchanged since first read, Sep 7, 2026 | -32% | $2.69 |
| RunPod | specialist | Community Cloudcaptured · unchanged since first read, Sep 6, 2026 | -32% | $2.69 |
| Hyperstack | specialist | globalcaptured · unchanged since first read, Sep 7, 2026 | -19% | $3.20 |
| Verda | specialist | globalcaptured · unchanged since first read, Sep 7, 2026 | -18% | $3.25 |
| RunPod | specialist | Secure Cloudcaptured · unchanged since first read, Sep 6, 2026 | -12% | $3.49 |
| Scaleway | specialist | fr-par-2captured · unchanged since first read, Sep 8, 2026€3.17 listed · ECB 1.1592 on September 11, 2026 | -8% | $3.67 |
| Vast.ai | marketplace | floorcaptured · unchanged since Sep 12, 2026 | -5% | $3.75 |
| Nebius | specialist | globalcaptured · unchanged since first read, Sep 7, 2026 | -3% | $3.85 |
| Modal | specialist | GPU onlycaptured · unchanged since first read, Sep 7, 2026 | -1% | $3.95 |
| Lambda | specialist | globalcaptured · unchanged since first read, Sep 7, 2026 | +1% | $3.99 |
| Together AI | specialist | globalcaptured · unchanged since first read, Sep 7, 2026 | +1% | $3.99 |
| Paperspace | specialist | GPU Dropletscaptured · unchanged since first read, Sep 8, 2026 | +11% | $4.41 |
| Paperspace | specialist | Paperspacecaptured · unchanged since first read, Sep 8, 2026 | +50% | $5.95 |
| CoreWeave | specialist | globalcaptured · unchanged since first read, Sep 7, 2026 | +55% | $6.16 |
| Amazon Web Services | hyper | us-east-1captured · unchanged since first read, Sep 7, 2026 | +73% | $6.88 |
| Google Cloud | hyper | us-west8, GPU onlycaptured · unchanged since first read, Sep 7, 2026 | +147% | $9.80 |
| Oracle Cloud (OCI) | hyper | globalcaptured · unchanged since first read, Sep 7, 2026 | +152% | $10.00 |
| Microsoft Azure | hyper | eastuscaptured · unchanged since first read, Sep 7, 2026 | +179% | $11.06 |
Showing 20 of 20 on-demand rates. Prices are per single GPU-hour; node prices divided by GPU count. Click a provider for its source page.
History, H100 SXM
Weekly median across providers (solid) and the cheapest listed rate (dashed) since January 2025. Grey is the reconstructed SkyPilot-catalog series, green our own scrape.
Buy vs rent
Retail price of the card ($29,950) divided by the median rate at the selected utilisation. Rental fee only — before power, network and hosting.
At this utilisation the card pays for itself in under a year — renting only wins for short experiments.
Hyperscalers vs specialists
Median within each group, for the same card.
For the same card Amazon Web Services, Google Cloud, Oracle Cloud (OCI) and Microsoft Azure charge 170% more. Only networking, compliance and credits justify the gap.
Spread across every card
Cards × providers, colour is the deviation from that card's median: who is dear everywhere, and who is cheap on one model only. An empty cell means the card is not listed there.
| RunPod | Verda | Amazon Web Services | CoreWeave | Modal | Oracle Cloud (OCI) | Paperspace | Cudo Compute | Google Cloud | |
|---|---|---|---|---|---|---|---|---|---|
| H100 SXM | $2.69 | $3.25 | $6.88 | $6.16 | $3.95 | $10.00 | $4.41 | $1.79 | $9.80 |
| H200 | $3.59 | $4.00 | $7.91 | $6.30 | $4.54 | $10.00 | $4.47 | — | $9.31 |
| A100 80GB | $1.19 | $1.79 | $3.43 | $2.70 | $2.50 | $4.00 | $3.18 | $1.50 | $3.93 |
| L40S | $0.79 | $1.37 | $1.86 | $2.25 | $1.95 | $3.50 | — | $0.87 | — |
| RTX 4090 | $0.34 | — | — | — | — | — | — | — | — |
| RTX 6000 Ada | $0.74 | $1.04 | — | — | — | — | $1.57 | — | — |
The cost of not looking
The same spread in money: what a team overpays renting 8 GPUs at the selected workload from the dearest provider instead of the cheapest.
| H100 SXM | $54,143 | |
|---|---|---|
| H200 | $47,421 | |
| A100 80GB | $16,410 | |
| L40S | $16,002 | |
| RTX 6000 Ada | $4,847 |
How it is computed, and where it comes from
What we measure
Publicly listed rental prices for 20 NVIDIA models at 21 providers — 220 current quotes — and the retail price of the cards.
Where it comes from
Each provider's own pricing page or API; retail from reseller pages. Every figure links back to its source.
How it is normalised
Node prices are divided by GPU count, per-second rates multiplied out, EUR shown at the dated ECB rate.
What we never do
Never estimate a price, never convert a currency silently, never publish a number a reader cannot verify on the linked page.
Seven files under CC BY 4.0. History is stored append-only and never rewritten.
As of , the median on-demand price of an NVIDIA H100 SXM across 16 providers is $3.97 per GPU-hour — the Cloud GPU Price Index. Cheapest listed on-demand rates today: NVIDIA H100 SXM $1.79 on Cudo Compute (no-kristiansand-1, GPU only); NVIDIA A100 80GB $0.87 on Vast.ai (floor); NVIDIA RTX 4090 $0.34 on RunPod (Community Cloud).
Data updated: