Rent a GPU: cloud prices compared
Key facts
An RTX 4090 rents from $0.55 an hour (Vast.ai, median of live offers) and an H100 from $2.20 (Vast.ai, median of live offers). You pay only while the machine runs.
Prices before VAT or sales tax, checked on each provider's own pricing page.
Renting a graphics card in a data centre runs the same open models as a PC at home, without buying one. Pick the GPU by how much memory your model needs, then the provider by price and where your data may go.
Updated
Danylo Chernov (NoctHost) benchmarked five Ollama models on Hetzner Cloud CPUs and rents A16 and A40 GPUs by the hour. Self-reported, community measurement, not ours. Read Danylo Chernov's interview.
Cheapest per GPU
The lowest hourly price per GPU, then the cheapest from a company based in the EU. A cheaper option is a Community Cloud host, a spot machine that can be stopped for someone else, or the single cheapest offer on a marketplace right now.
| GPU | Memory | Largest model that fits | Lowest per hour | Lowest from an EU company | Cheaper option |
|---|---|---|---|---|---|
| RTX 4090 | 24 GB | Qwen3.6-35B-A3B | $0.55 median of live offers at Vast.ai | None found | $0.15 cheapest live offer at Vast.ai |
| Other 24 GB cards (L4, RTX PRO 4000) | 24 GB | Qwen3.6-35B-A3B | €0.75 (about $0.84 at 1.1195 USD per EUR) at OVHcloud | €0.75 at OVHcloud | None listed |
| RTX 5090 | 32 GB | Recursive Flint | $0.80 median of live offers at Vast.ai | None found | $0.41 cheapest live offer at Vast.ai |
| 48 GB cards (L40S, RTX 6000 Ada, RTX A6000) | 48 GB | NVIDIA Nemotron 3.5 Bolt 74B-A9B | $0.40 median of live offers at Vast.ai | $1.38 at Verda (formerly DataCrunch) | $0.32 cheapest live offer at Vast.ai |
| A100 80GB | 80 GB | Ornith-1.5-122B | $0.66 median of live offers at Vast.ai | $2.09 at Verda (formerly DataCrunch) | $0.35 cheapest live offer at Vast.ai |
| H100 | 80 GB | Ornith-1.5-122B | $2.20 median of live offers at Vast.ai | €2.80 at OVHcloud | $1.73 cheapest live offer at Vast.ai |
| RTX PRO 6000 | 96 GB | Ornith-1.5-122B | $1.55 median of live offers at Vast.ai | $1.80 at Nebius | $1.00 cheapest live offer at Vast.ai |
| H200 | 141 GB | Qwen3.8-Flash-Next | $3.99 starting price at Hyperbolic | $5.40 at Nebius | $2.89 spot at Verda (formerly DataCrunch) |
| B200 | 180 GB | Qwen3.8-Flash-Next | $5.99 starting price at Hyperbolic | $8.19 at Verda (formerly DataCrunch) | $4.10 spot at Verda (formerly DataCrunch) |
Euro prices are the provider's own; the dollar figure next to them is our conversion at the exchange rate alone. A Hetzner server has a one-time setup fee, so it is not counted as the lowest hourly price.
RTX 4090 for rent
A gaming card with 24 GB. The cheapest way to run a small or mid-size model.
Holds 25 of our 42 models at standard quality with an 8,192-token conversation, by our estimate (23.4 GB usable). Largest: Qwen3.6-35B-A3B. Every model that fits in 24 GB.
| Provider | GPU as listed | Per GPU-hour | Cheaper option | Data centre | Source |
|---|---|---|---|---|---|
| Vast.ai | RTX 4090, 24 GB | $0.55 median of live offers | $0.15 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Runpod | RTX 4090, 24 GB | $0.89 | $0.34 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
Other 24 GB cards (L4, RTX PRO 4000) for rent
Data centre cards with 24 GB. Slower memory than an RTX 4090, often sold in the EU.
Holds 25 of our 42 models at standard quality with an 8,192-token conversation, by our estimate (23.4 GB usable). Largest: Qwen3.6-35B-A3B. Every model that fits in 24 GB.
| Provider | GPU as listed | Per GPU-hour | Cheaper option | Data centre | Source |
|---|---|---|---|---|---|
| Hetzner EU company | GEX45: RTX PRO 4000 Blackwell SFF, 24 GB | €0.34 or €214.00 a month; setup fee €209.00 once (about $0.38 at 1.1195 USD per EUR) | None listed | Helsinki, Finland (HEL1) | Hetzner, 21 Oct 2026 |
| OVHcloud EU company | l4-90 (L4), 24 GB | €0.75 (about $0.84 at 1.1195 USD per EUR) | None listed | Gravelines, France (GRA11) | OVHcloud, 21 Oct 2026 |
| Scaleway EU company | L4-1-24G | €0.79 (about $0.88 at 1.1195 USD per EUR) | None listed | Paris, France and Warsaw, Poland | Scaleway, 21 Oct 2026 |
RTX 5090 for rent
32 GB, room for a 32B model at standard quality.
Holds 26 of our 42 models at standard quality with an 8,192-token conversation, by our estimate (31.4 GB usable). Largest: Recursive Flint. Every model that fits in 32 GB.
| Provider | GPU as listed | Per GPU-hour | Cheaper option | Data centre | Source |
|---|---|---|---|---|---|
| Vast.ai | RTX 5090, 32 GB | $0.80 median of live offers | $0.41 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Runpod | RTX 5090, 32 GB | $1.19 | $0.69 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
48 GB cards (L40S, RTX 6000 Ada, RTX A6000) for rent
48 GB on one card: a 70B model at standard quality only just fits, with a short conversation.
Holds 29 of our 42 models at standard quality with an 8,192-token conversation, by our estimate (47.4 GB usable). Largest: NVIDIA Nemotron 3.5 Bolt 74B-A9B. Every model that fits in 48 GB.
| Provider | GPU as listed | Per GPU-hour | Cheaper option | Data centre | Source |
|---|---|---|---|---|---|
| Vast.ai | RTX A6000, 48 GB | $0.40 median of live offers | $0.32 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Runpod | RTX A6000, 48 GB | $0.59 | $0.33 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
| Vast.ai | RTX 6000Ada, 48 GB | $0.71 median of live offers | $0.57 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Vast.ai | L40S, 48 GB | $0.73 median of live offers | $0.43 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Runpod | RTX 6000 Ada, 48 GB | $0.99 | $0.74 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
| Lambda | NVIDIA A6000, 48 GB | $1.09 | None listed | Lambda has US regions and one in Germany, not listed per GPU | Lambda, 21 Oct 2026 |
| Runpod | L40S, 48 GB | $1.09 | $0.79 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
| Verda (formerly DataCrunch) EU company | 1x RTX 6000 Ada 48GB | $1.38 | $0.69 spot | Helsinki, Finland | Verda (formerly DataCrunch), 21 Oct 2026 |
| Nebius EU company | NVIDIA L40S (Intel), 48 GB | $1.55 starting price | None listed | Finland (eu-north1) | Nebius, 21 Oct 2026 |
| OVHcloud EU company | l40s-90 (L40S), 48 GB | €1.40 (about $1.57 at 1.1195 USD per EUR) | None listed | Gravelines, France (GRA11) | OVHcloud, 21 Oct 2026 |
| DigitalOcean | NVIDIA L40S, 48 GB | $1.57 | None listed | Toronto (TOR1) | DigitalOcean, 21 Oct 2026 |
| DigitalOcean | NVIDIA RTX 6000 Ada Generation, 48 GB | $1.57 | None listed | Toronto (TOR1) | DigitalOcean, 21 Oct 2026 |
| Scaleway EU company | L40S-1-48G | €1.47 (about $1.65 at 1.1195 USD per EUR) | None listed | Paris, France (PAR-2) and Warsaw, Poland | Scaleway, 21 Oct 2026 |
| Verda (formerly DataCrunch) EU company | 1x L40S 48GB | $1.79 | $0.90 spot | Helsinki, Finland | Verda (formerly DataCrunch), 21 Oct 2026 |
A100 80GB for rent
80 GB, the older data centre card. Cheaper than an H100 and slower.
Holds 35 of our 42 models at standard quality with an 8,192-token conversation, by our estimate (79.4 GB usable). Largest: Ornith-1.5-122B. Can I run Ornith-1.5-122B?
| Provider | GPU as listed | Per GPU-hour | Cheaper option | Data centre | Source |
|---|---|---|---|---|---|
| Vast.ai | A100 SXM4 80GB | $0.66 median of live offers | $0.40 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Vast.ai | A100 PCIE 80GB | $0.73 median of live offers | $0.35 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Runpod | A100 PCIe 80GB | $1.79 | $1.19 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
| Runpod | A100 SXM 80GB | $1.79 | $1.39 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
| Verda (formerly DataCrunch) EU company | 1x A100 SXM4 80GB | $2.09 | $1.04 spot | Helsinki, Finland | Verda (formerly DataCrunch), 21 Oct 2026 |
| OVHcloud EU company | a100-180 (A100 80GB) | €2.75 (about $3.08 at 1.1195 USD per EUR) | None listed | Gravelines, France (GRA11) | OVHcloud, 21 Oct 2026 |
H100 for rent
80 GB (94 GB for the NVL), the common card for serving a 70B model to several people.
Holds 35 of our 42 models at standard quality with an 8,192-token conversation, by our estimate (79.4 GB usable). Largest: Ornith-1.5-122B. Can I run Ornith-1.5-122B?
| Provider | GPU as listed | Per GPU-hour | Cheaper option | Data centre | Source |
|---|---|---|---|---|---|
| Vast.ai | H100 SXM 80GB | $2.20 median of live offers | $1.73 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Vast.ai | H100 PCIE 80GB | $2.33 median of live offers | $2.32 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Vast.ai | H100 NVL, 80 GB | $2.86 median of live offers | $2.33 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Runpod | H100 PCIe 80GB | $2.89 | $1.99 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
| OVHcloud EU company | h100-380 (H100 80GB) | €2.80 (about $3.13 at 1.1195 USD per EUR) | None listed | Gravelines, France (GRA9, GRA11) | OVHcloud, 21 Oct 2026 |
| Hyperbolic | Nvidia H100 SXM, 80 GB | $3.19 starting price | None listed | US and Europe | Hyperbolic, 21 Oct 2026 |
| Runpod | H100 NVL 94GB | $3.19 | $2.59 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
| Scaleway EU company | H100-1-80G | €2.87 (about $3.21 at 1.1195 USD per EUR) | None listed | Paris, France (PAR-2) and Warsaw, Poland | Scaleway, 21 Oct 2026 |
| Lambda | NVIDIA H100 PCIe, 80 GB | $3.29 | None listed | Lambda has US regions and one in Germany, not listed per GPU | Lambda, 21 Oct 2026 |
| Runpod | H100 SXM 80GB | $3.99 | $2.69 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
| Lambda | NVIDIA H100 SXM, 80 GB | $4.29 | None listed | Lambda has US regions and one in Germany, not listed per GPU | Lambda, 21 Oct 2026 |
| DigitalOcean | NVIDIA HGX H100, 80 GB | $4.41 | None listed | Several regions, Amsterdam (AMS3) included | DigitalOcean, 21 Oct 2026 |
| Verda (formerly DataCrunch) EU company | 1x H100 SXM5 80GB | $4.43 | $2.21 spot | Helsinki, Finland | Verda (formerly DataCrunch), 21 Oct 2026 |
| Nebius EU company | NVIDIA HGX H100, 80 GB | $4.50 | None listed | Finland (eu-north1) | Nebius, 21 Oct 2026 |
RTX PRO 6000 for rent
96 GB on one card, newer and often cheaper per GB than an H100.
Holds 35 of our 42 models at standard quality with an 8,192-token conversation, by our estimate (95.4 GB usable). Largest: Ornith-1.5-122B. Every model that fits in 96 GB.
| Provider | GPU as listed | Per GPU-hour | Cheaper option | Data centre | Source |
|---|---|---|---|---|---|
| Vast.ai | RTX PRO 6000 WS, 96 GB | $1.55 median of live offers | $1.00 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Hetzner EU company | GEX63: RTX PRO 6000 Blackwell Max-Q, 96 GB | €1.60 or €999.00 a month; setup fee €999.00 once (about $1.79 at 1.1195 USD per EUR) | None listed | Helsinki, Finland (HEL1) | Hetzner, 21 Oct 2026 |
| Nebius EU company | NVIDIA RTX PRO 6000, 96 GB | $1.80 | None listed | Madrid, Spain (eu-south1), UK and US | Nebius, 21 Oct 2026 |
| Verda (formerly DataCrunch) EU company | 1x RTX PRO 6000 96GB | $2.44 | $1.22 spot | Helsinki, Finland | Verda (formerly DataCrunch), 21 Oct 2026 |
| Runpod | RTX Pro 6000, 96 GB | $2.49 | $1.69 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
H200 for rent
141 GB on one card, for very large models or long conversations.
Holds 37 of our 42 models at standard quality with an 8,192-token conversation, by our estimate (140.4 GB usable). Largest: Qwen3.8-Flash-Next. Can I run Qwen3.8-Flash-Next?
| Provider | GPU as listed | Per GPU-hour | Cheaper option | Data centre | Source |
|---|---|---|---|---|---|
| Hyperbolic | Nvidia H200, 141 GB | $3.99 starting price | None listed | US and Europe | Hyperbolic, 21 Oct 2026 |
| DigitalOcean | NVIDIA HGX H200, 141 GB | $4.47 | None listed | New York (NYC2) and Atlanta (ATL1) | DigitalOcean, 21 Oct 2026 |
| Vast.ai | H200, 141 GB | $4.61 median of live offers | $3.42 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
| Runpod | H200, 141 GB | $5.29 | $3.59 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
| Nebius EU company | NVIDIA HGX H200, 141 GB | $5.40 | None listed | Finland, France and US | Nebius, 21 Oct 2026 |
| Verda (formerly DataCrunch) EU company | 1x H200 SXM5 141GB | $5.77 | $2.89 spot | Helsinki, Finland | Verda (formerly DataCrunch), 21 Oct 2026 |
B200 for rent
180 to 192 GB, the newest card in this list and the most expensive.
Holds 37 of our 42 models at standard quality with an 8,192-token conversation, by our estimate (179.4 GB usable). Largest: Qwen3.8-Flash-Next. Can I run Qwen3.8-Flash-Next?
| Provider | GPU as listed | Per GPU-hour | Cheaper option | Data centre | Source |
|---|---|---|---|---|---|
| Hyperbolic | Nvidia B200, 192 GB | $5.99 starting price | None listed | Not stated | Hyperbolic, 21 Oct 2026 |
| Lambda | NVIDIA B200 SXM6, 180 GB | $6.99 | None listed | Lambda has US regions and one in Germany, not listed per GPU | Lambda, 21 Oct 2026 |
| Runpod | B200, 180 GB | $7.99 | $5.98 Community Cloud | Several regions; Runpod has EU regions, not listed per GPU | Runpod, 21 Oct 2026 |
| Verda (formerly DataCrunch) EU company | 1x B200 SXM6 180GB | $8.19 | $4.10 spot | Helsinki, Finland | Verda (formerly DataCrunch), 21 Oct 2026 |
| Nebius EU company | NVIDIA HGX B200, 180 GB | $8.50 | None listed | US and Middle East only | Nebius, 21 Oct 2026 |
| Vast.ai | B200, 192 GB | $10.63 median of live offers | $6.25 cheapest live offer | Marketplace hosts, location varies | Vast.ai, 21 Oct 2026 |
The providers
Where the company is based decides which law it answers to. An EU company falls under the GDPR directly. A US company with an EU data centre keeps your data in the EU, but stays a US company. For personal data, check the provider's data processing agreement either way.
| Provider | Based in | EU data centres | Billing | Ready to run a model |
|---|---|---|---|---|
| Runpod | United States | EU data centre regions; Runpod says GDPR applies to data in them | Per second | Ready vLLM worker; Ollama runs on its PyTorch template |
| Vast.ai | United States | None stated | Per second, no minimum | vLLM template and an Open WebUI (Ollama) template |
| Lambda | United States | Germany (europe-central-1) | Per minute | Plain machine with drivers; guides for vLLM and Ollama |
| DigitalOcean | United States | Amsterdam (AMS3), H100 only | Per second, 5-minute minimum | Ready vLLM inference image |
| Hyperbolic | United States | Northern and Western Europe regions (countries not named) | Per hour | Plain Ubuntu with drivers and Docker |
| Hetzner | Germany (EU) | Helsinki, Finland (HEL1) | Per month or per hour, plus a one-time setup fee | Dedicated server; guides for Ollama |
| Scaleway | France (EU) | Paris, France and Warsaw, Poland | Per minute | Managed inference with an OpenAI-compatible API, at its own price |
| OVHcloud | France (EU) | Gravelines, France (and Paris for H200) | Per hour, or a monthly price | AI Deploy (Docker) and AI Endpoints, priced apart |
| Nebius | Netherlands (EU) | Finland and France for H100 and H200; Madrid for RTX PRO 6000 | Per second | Token Factory: dedicated endpoints with an OpenAI-compatible API |
| Verda (formerly DataCrunch) | Finland (EU) | Helsinki, Finland | Prepaid 10-minute blocks | Guides for vLLM and Ollama on its serverless containers |
If you buy through these links we may earn a commission. It doesn't change our ranking.
Also checked and left out: TensorDock: Its public price table says it was last updated in July 2024, and the live marketplace needs a login. Paperspace: Now part of DigitalOcean; its old price page mixes up hourly and multi-year prices. DigitalOcean is listed instead. Together AI GPU clusters: Sold as clusters, not as one GPU you start alone.
Rent a GPU or pay per token?
Services such as OpenRouter run the same open models and charge per token (about three quarters of a word). That is usually cheaper for light use, because you never pay for an idle GPU. A rented GPU wins when it stays busy, when you need a model or setting they do not offer, or when your data must stay on a machine only you use.
| Model | Input per million tokens | Output per million tokens | Source |
|---|---|---|---|
| Llama 3.3 70B Instruct | $0.1 | $0.32 | OpenRouter, lowest of 11 providers, 21 Oct 2026 |
| gpt-oss-120b | $0.03 | $0.15 | OpenRouter, lowest of 24 providers, 21 Oct 2026 |
| DeepSeek V3.2 | $0.2088 | $0.3096 | OpenRouter, lowest of 12 providers, 21 Oct 2026 |
For scale: one hour of the cheapest 48 GB card here ($0.40) costs the same as about 1.3 million output tokens of Llama 3.3 70B Instruct at $0.32 per million. Buy, rent or pay per use for your team works out the monthly cost.
What fits is our estimate from the model size, not a measured test. Prices are the providers' own, dated, before VAT, and change often; marketplace prices change by the hour. How we calculate
Common questions
- What does it cost to rent an RTX 4090 or an H100?
- On 21 Oct 2026 the lowest hourly figure we found for an RTX 4090 was $0.55 at Vast.ai (median of live offers), and for an H100 $2.20 at Vast.ai (median of live offers). Prices are per GPU per hour, before VAT or sales tax.
- Which GPU clouds are based in the EU?
- Hetzner (Germany), Scaleway and OVHcloud (France), Nebius (Netherlands) and Verda, formerly DataCrunch (Finland). The cheapest H100 among them was €2.80 per hour at OVHcloud. A US company with an EU data centre keeps the data in the EU, but it is still a US company.
- Which rented GPU do I need for a 70B model?
- Llama 3.3 70B at standard quality needs about 46.0 GB with a conversation of 8,192 tokens, by our estimate. The smallest GPU here that holds it is a 48 GB card. The cheapest one billed by the hour was RTX A6000 at Vast.ai, $0.40 per hour. More users or longer conversations need more memory, so an 80 GB A100 or H100 gives room to grow.