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)
Inference hosting compared: serverless and dedicated GPUs
We may earn commissions when you shop through the links below. Learn more
Serve open models on serverless GPUs that scale to zero, on dedicated instances or on managed endpoints. Compare verified per-GPU-hour prices, billing increments and regions.
Inference hosting providers
Serverless or dedicated?
- Serverless GPUs
Billed while requests run and able to scale to zero. A good fit for bursty or low traffic; expect cold starts.
- Dedicated GPUs
Billed for every hour the instance runs. A good fit for steady traffic and predictable latency.
- Billing increment
Per-second billing matters for short jobs and autoscaling; hourly billing rounds partial hours up.
- Memory fit
The model weights plus the KV cache for your context length and concurrency must fit in VRAM.
- Per-token APIs
Some platforms sell model access per million tokens instead of GPU time. This page compares GPU-hour pricing; per-token prices are not mixed into it.
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.
Explore related hubs
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 does an inference server do?
It loads a model into GPU memory and answers requests over an API, batching them to keep the GPU busy. vLLM and Ollama are common open-source examples.
How much does inference cost?
It depends on how the model is served: per-token APIs charge per million tokens, dedicated endpoints and GPU instances charge per GPU-hour while they run, and serverless GPUs charge for the seconds they work. The results list verified per-GPU-hour prices; the GPU Cost Calculator turns them into a monthly estimate.
Is a serverless GPU cheaper than a dedicated GPU?
It depends on utilisation. Serverless bills only while requests run, so it suits bursty traffic; a dedicated GPU bills every hour it runs, so it suits steady load. Filter by billing to compare both for the same GPU.
What does per-second billing change?
You pay for the seconds a worker runs instead of rounding up to the hour, which lowers the cost of short jobs and scale-to-zero endpoints.
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.