Hyperstack
Single-GPU to 8-GPU VMs on A100 80GB PCIe, H100 PCIe and L40 for fine-tuning and inference, billed per minute with no data-transfer fees
A6000 48 GB · from:
$0.50≈ €0.447 (estimated conversion from $0.50)
vLLM hosting compared: templates, serverless endpoints and GPUs
We may earn commissions when you shop through the links below. Learn more
vLLM is an open-source inference engine with an OpenAI-compatible API server. Compare providers with official vLLM templates or documentation by GPU, VRAM and verified price.
vLLM hosting providers
What the official vLLM docs say
vLLM can be deployed as a server that implements the OpenAI API protocol, started with vllm serve followed by the model name.
GPU requirement: compute capability 7.5 or higher, for example T4, A100, L4, H100 or B200.
NVIDIA Blackwell GPUs (B200, GB200) need CUDA 12.8 or newer.
Tools for this decision
- VRAM / Model FitEstimate the GPU memory a model needs for inference or fine-tuning, and which GPUs fit.
- GPU Cost CalculatorEstimate monthly GPU spend from verified hourly prices, usage hours, instances and storage.
- Agent Hosting SelectorAnswer a few questions about your agent and get a best-fit hosting category and provider shortlist.
How we compare
Prices per GPU-hour.
GPU prices are USD per single GPU per hour from the provider's own pricing page or API. Multi-GPU nodes are divided by their GPU count, and a price only rented as a node says so (“rented as 8 GPUs”). Rows lead with the cheapest pay-as-you-go price (on-demand, secure, community or serverless). A spot, reserved or marketplace price leads only when the provider has no pay-as-you-go price here, or when you filter by that billing; it is labelled with its price type and listed in its own group after pay-as-you-go prices, never ranked against them. Prices exclude VAT and sales tax.Not verified means not verified.
When no official source confirms a value we show “Not verified” instead of an estimate. Where a provider's terms do not allow its prices to be republished, we link to its live pricing instead.Freshness on every row.
Updated today: checked on the current day (UTC). Fresh: up to 7 days. Stale: up to 30 days. Needs refresh: older than that.No paid placement.
Results only change order with the sort you pick. Affiliate relationships never affect order, filters or which providers appear.
Filters match a provider when one of its listed offers meets every GPU, VRAM, region, price and billing-increment condition together. Spot, reserved and serverless billing list each provider that offers that pricing and show only prices on that basis, or “not verified” when we have none. Read the full methodology and our affiliate disclosure.
Frequently asked questions
What GPU do I need for vLLM?
An NVIDIA GPU with compute capability 7.5 or higher, according to the vLLM docs, and enough VRAM for the model weights plus the KV cache for your context length and number of concurrent requests. The VRAM / Model Fit tool estimates the memory for common models.
What do the support levels in the results mean?
“Official template” means the provider publishes a ready-made vLLM image, template or worker. “Documented setup” means a guide explains the setup on that provider; when the guide is not the provider's own, the row names where it is published, such as “(vLLM docs)”. “Community guide” means a published community guide covers it. Providers where you would install vLLM yourself without such a guide are not listed.
Should I use Ollama or vLLM?
Ollama makes it simple to run a model on one machine for yourself or a small team. vLLM is built to serve many concurrent requests efficiently on GPUs, with an OpenAI-compatible server, so it suits production endpoints. Both are open source; you pay for the server or GPU they run on.
Are prices per GPU or per instance?
Per GPU. Every GPU price is shown in USD per single GPU per hour; offers for multi-GPU nodes are divided by the number of GPUs, and a price you can only rent as a whole node says so (for example “rented as 8 GPUs”). Storage, data transfer and taxes are not included unless the provider includes them.