Choose an orderable configuration
Your model fits on one card, but an attractive quote comes from an eight-GPU node. Dividing by eight gives a unit cost. It does not reduce the bill to one eighth unless the provider lets you order a single card from that configuration.
First estimate weight memory, KV cache and concurrent requests. A model distributed across cards also needs suitable interconnects and runtime support. Adding the memory capacities does not by itself describe the performance of a multi-GPU server.
The following configurations were listed in the public catalogs reviewed on November 3, 2026. They describe raw compute for your own serving stack rather than a managed inference platform.
| Provider and SKU | Included GPUs | Ordering implication |
|---|---|---|
| Scaleway H100-1-80G | 1 × H100 PCIe, 80 GB | Single-GPU instance |
| Scaleway H100-SXM-2-80G | 2 × H100 SXM, 80 GB each | Price the complete pair |
| Scaleway B300-SXM-8-288G | 8 × B300, 288 GB each | Price the eight-GPU node |
| OVHcloud h100-380 | 1 × H100, 80 GB | Single-GPU instance |
| OVHcloud h200-1920 | 8 × H200, 141 GB each | Price the eight-GPU node |
Use the linked regional catalogs for the current euro quote on a French deployment. For a USD budget, obtain the applicable billing quote or apply an explicitly dated exchange-rate assumption to the complete instance total, including the relevant tax treatment. A US-region offer may have different prices and capacity.
Compare the complete monthly bill
Use the same term and configuration when comparing providers. Include persistent storage, snapshots, network addresses, transfers and operating work where applicable. Determine which resources remain billable after stopping an instance or deleting the GPU server. A retained storage volume can outlive the compute charge.
For a currency-neutral calculation, multiply the full instance hourly price by billed hours. In an illustrative USD budget, an eight-GPU node assumed to cost $40 an hour would cost $28,800 over a continuously rented 720-hour month. Its $5-per-GPU quotient would not make the whole node cost $3,600. These amounts are hypothetical and do not quote a provider.
Check commitment terms separately. A discount requiring a long reservation cannot be applied to a short prototype unless its contract allows that usage. The API versus GPU inference cost guide connects this full bill to useful production capacity and financial break-even.
Check region and capacity before committing
A public catalog can list an instance that is unavailable in your selected zone or behind an account quota. Confirm where it can actually run and whether the capacity can be retained for the required period. A product page is not evidence of stock allocated to your account.
Start with a short rental of the intended SKU and run your real request mix. Record the model, precision, software versions and test conditions, then measure memory use, errors, time to first token and latency under concurrency. The vLLM, llama.cpp and TensorRT-LLM comparison explains the serving choices.
For multi-GPU workloads, check links between cards and between machines. Public network bandwidth is not automatically the fabric carrying model communication. A lower hourly quote can lose its advantage if the complete workload needs more instances or misses its response-time target.
Review what European hosting actually covers
Location can help meet latency or data-residency requirements. Trace prompts, outputs, logs, backups and support access separately: a compute region in France does not establish the location of every associated service.
Review the contracting entity, subprocessors, data-processing terms, access arrangements and the scope of certifications for the exact offer. Where a project imposes a particular qualification or legal requirement, verify it in the applicable documentation and contract. The country of a data center alone does not establish every jurisdictional guarantee.
Apply the same process to other European suppliers: comparable quote, deployable capacity, a workload test and the applicable contract. Two providers’ catalogs are not enough to name a cheapest supplier across the entire European market.
Sources and method
Published November 3, 2026. Configuration checks: Scaleway GPU pricing with PAR-2 selected and OVHcloud Public Cloud GPU grid. The catalogs supply regional prices; the USD example illustrates billing arithmetic with explicitly hypothetical amounts. No hardware benchmark, stock reservation or legal opinion is implied.
