GPU Cloud Pricing by Region: US, Europe and Global Marketplaces
The same GPU does not cost the same everywhere. An H100 rented in a US data centre, one in Frankfurt or Paris, and one on a marketplace whose host happens to sit in Iceland can differ by 20–50% per hour, and hyperscaler prices add their own regional multipliers on top. This page groups every on-demand listing in the dataset by region, explains what drives the gaps, and gives a short procedure for choosing a region when latency and data-residency rules leave you a choice.
On-demand GPU rates by region
USD per single GPU per hour, on-demand list price, grouped by where the provider hosts the listing. "Listings" is the number of providers in the dataset offering that GPU in that region; the range runs from the cheapest to the most expensive of them.
| GPU | Listings | $/hr range | Cheapest (location) | Most expensive |
|---|---|---|---|---|
| United States | ||||
| H100 80GB | 8 | $2.39 – $4.99 | Vast.ai (US-East) | AWS |
| H200 141GB | 7 | $3.58 – $10.60 | Crusoe (US-West) | AWS |
| B200 192GB | 4 | $5.89 – $7.99 | Crusoe (US-West) | Azure |
| A100 80GB | 7 | $1.79 – $3.67 | Vast.ai (US-East) | Azure |
| MI325X 256GB | 1 | $2.25 | Vast.ai (US-East) | — |
| Europe | ||||
| H100 80GB | 5 | $2.79 – $3.49 | RunPod (EU-West) | OVHcloud |
| H200 141GB | 2 | $4.95 – $5.39 | Nebius (Finland) | Scaleway |
| B200 192GB | 1 | $6.89 | RunPod (EU-West) | — |
| A100 80GB | 4 | $1.89 – $3.19 | RunPod (EU-West) | Scaleway |
| MI300X 192GB | 1 | $2.89 | OVHcloud (France) | — |
| MI325X 256GB | 2 | $2.35 – $2.49 | Nebius (Finland) | RunPod |
| RTX 4090 | 2 | $0.69 – $0.79 | RunPod (EU-West) | Genesis Cloud |
| Global marketplaces (host location varies) | ||||
| H100 80GB | 1 | $2.29 | FluidStack (Global) | — |
| A100 80GB | 1 | $1.65 | FluidStack (Global) | — |
| MI300X 192GB | 1 | $1.99 | FluidStack (Global) | — |
| RTX 4090 | 2 | $0.65 – $0.74 | FluidStack (Global) | Vast.ai |
Region labels follow each provider's own naming: US-East/US-West/US-Central for US clouds and marketplaces, EU-West for RunPod's European hosts, and country or city names (Finland, France, Paris, Iceland) for European specialist clouds. Spot and reserved tiers are excluded — see spot vs on-demand pricing.
Why the same GPU costs more in some regions
- Electricity. An H100 server draws close to 10 kW under load. Regions with cheap power — the US Pacific Northwest and Texas, Iceland and the Nordics with hydro and geothermal — can undercut Frankfurt, London or Singapore by a margin that shows up directly in the hourly rate. This is why Nordic and Icelandic listings sit at the low end of the European table.
- Supply. Most specialist GPU clouds built their first and largest sites in the US, so US regions have the most listings and the deepest price competition. Europe has fewer H200/B200 listings, and scarce SKUs carry a premium wherever supply is thin.
- Hyperscaler regional multipliers. AWS, Azure and Google Cloud publish per-region prices for the same instance type; European and Asia-Pacific regions typically run 10–30% above the cheapest US region, and some Asian regions higher still. The hyperscaler rows in the table are US-region prices.
- Currency and tax presentation. European providers such as OVHcloud and Scaleway price in euros; the USD figures above move with the exchange rate. Prices shown to EU customers may also include VAT, which the dataset excludes.
- Data-residency demand. Regulated European workloads must stay in-region, which lets European sites charge a premium that a latency-indifferent training job never needs to pay.
Europe: Frankfurt, Paris, London and the Nordics
Within Europe the dataset shows a clear split. Marketplace and specialist hosts in the EU-West pool, Finland and Iceland price the H100 at $2.79–2.99/hr — within 20–30% of the cheapest US listing. Sovereign-cloud providers hosting in France (OVHcloud, Scaleway) list the H100 at $3.39–3.49/hr and the A100 at $2.89–3.19, a 40–70% premium over the cheapest European marketplace rate, in exchange for EU-jurisdiction hosting and enterprise support. Hyperscaler Frankfurt and London regions sit above both: expect their US list price plus a regional uplift, and confirm on the provider's per-region page before budgeting.
Asia-Pacific: Singapore, Malaysia, Japan, India
The dataset currently has no on-demand listings hosted in Asia-Pacific, so no figures are shown rather than estimated ones. What can be said from provider pricing pages: hyperscaler GPU instances in Singapore, Tokyo and Sydney generally carry the largest regional uplift of any geography over US rates, driven by power cost and constrained supply; a growing set of regional providers (in Singapore, Malaysia, Japan, India and Australia) undercut the hyperscalers locally but rarely reach US marketplace prices. If your workload has no residency requirement, the cheapest option for an Asia-based team is usually a US or Nordic region plus the egress cost of moving results back — check the hidden cloud GPU costs page for the egress line. Asia-Pacific listings will be added to the index as providers are reviewed.
How to choose a region
- Apply the hard constraints first. Data residency (GDPR, sector regulation, customer contracts) and end-user latency for inference rule regions in or out before price matters.
- For training, ignore geography and buy the cheapest capacity. A training job does not care where it runs; the only regional costs that matter are the hourly rate and the egress fee to bring checkpoints home. Global marketplaces and Nordic hosts are usually the floor.
- Check capacity, not just price. A cheap region with "limited" availability is not cheap if you cannot get the instance. Several European H100/H200 listings in the table are marked limited.
- Price the whole bill. Storage and egress rates also vary by region, and the difference can be larger than the GPU-rate difference for data-heavy jobs.
- Use spot in the region with the most idle capacity. Spot availability is a regional property; secondary US regions and off-peak European regions tend to have the deepest spot pools. See the spot GPU guide.
Related reading
- GPU rental price index — the same listings summarised as low / median / high per accelerator, regardless of region.
- GPU cloud providers compared — who hosts where, with a region filter.
- Spot vs on-demand GPU pricing — the other big price lever.
- GPU carbon footprint calculator — region also sets the grid carbon intensity of a run.