The comparison table shows €0.45/h for an A100. The data scientist approves the budget. Monday morning: quota exhausted, 36-hour queue, 2 TB dataset ingress billed as egress, CUDA 11.x image incompatible with the chosen PyTorch stack. Wednesday finally a pod — two days lost. The hourly price was accurate; wait time cost appeared nowhere in the sales deck.
Comparing AI GPUs without measuring availability is comparing flight prices without departure dates. The European market advertises attractive rates on paper, but GPU capacity remains scarce during rush periods. This inquiry lists what catalogs omit and how to build an honest comparison before signing.
Beyond hourly price: five forgotten cost lines
The €/h rate is only one line on a larger invoice. Five items consistently blow budgets.
Availability and quotas. On public cloud, GPU capacity is shared. During hype periods (model launches, academic terms), queues can stretch to several days. A GPU "available" in the catalog may be unavailable in practice.
Dataset storage. Local NVMe, object storage, provisioned IOPS: each GB/month adds to compute. A multi-terabyte dataset moving between regions also generates egress — see our inquiry on egress fees.
Egress. Exporting a trained model, logs, or checkpoints to another region or provider can represent 20–40% of total budget if nobody modeled it upfront.
Setup and images. CUDA drivers, ML frameworks, Jupyter notebooks: engineer hours making the environment usable do not appear on the GPU invoice, but they cost real money.
Idle GPU. An instance left running between epochs or forgotten over the weekend leaks budget as surely as an unprovisioned traffic spike.
| Criterion | TCO impact | Question to ask |
|---|---|---|
| €/h GPU | Direct | Exact card, how much VRAM? |
| Queue | Project delay | Is there an allocation SLA? |
| Egress | +20–40% possible | Is exit region fixed? |
| Reservation | −30% sometimes | Is a 1–3 month commit required? |
Availability: billed GPU is not obtained GPU
Catalogs list prices for capacity queues sometimes make fictional during rush. Before any commitment, request an allocation trial: launch an instance, time the delay until a CUDA-ready environment, repeat over ten business days.
Also measure median delay over two weeks, not the best case observed on a quiet Tuesday. Teams promising a model delivery date to leadership without this test regularly explain an "infrastructure" delay.
On reserved offers (1–3 month commit), availability is usually more predictable — at the cost of reduced flexibility. For 30 days of continuous training, reservation often beats on-demand; for twenty fine-tuning hours per month, the opposite holds.
The cheapest GPU is the one running when you need it — not hero-priced on a static comparator at midnight.
Scenarios A and B: experimentation vs continuous training
Scenario A — Experimentation. Twenty GPU hours per month, 7B fine-tuning. Priority: instance in under one hour, small local disk, low egress. On-demand cloud fits; long reservation would overpay.
Scenario B — Continuous training. Two GPUs 24/7 for thirty days. Priority: reserved price, no queue, responsive support. Dedicated GPU or long reservation beats on-demand once load is stable and predictable.
Compare OVHcloud and Scaleway on reserved grids, and keep Hetzner as CPU-only fallback if GPU stays unavailable — so data prep pipelines do not stall entirely.
The summit: €/h comparison hides real capacity
Decide and move forward without blind spots
Start by requesting a GPU allocation trial from two or three shortlisted providers, and time provisioning until a CUDA-ready environment. Calculate realistic dataset egress before locking a region. Confront hourly cost with the cost of a week lost in queue — that is often where the real trade-off lies. Document results in a twelve-month TCO spreadsheet, then cross-check offers via our directory and comparator.
Frequently asked questions
Why is hourly GPU price misleading?
Because it ignores real availability, setup time, storage, egress, and hours the GPU runs without producing output. On a fine-tuning project, effective cost can double versus the advertised rate. Include these lines before comparing two quotes.
How do you compare two GPU cloud offers?
Align the same card model, region, included disk, egress, ML image, and allocation delay. Test allocation over two weeks and note median delay, not best case. Only this protocol compares offers that look equivalent on paper.
GPU cloud or dedicated GPU server?
Cloud suits spikes and experimentation. Dedicated pays off once GPU load runs continuously with validated economics — often between two and four weeks of intensive use. Run the math on your real profile, not a theoretical average.
Which European players offer GPUs?
OVHcloud, Scaleway, and Hetzner (limited offer) are visible European players, alongside hyperscalers in EU regions. Availability varies sharply: test allocation before any long project and plan a CPU fallback if queues slip.
Low GPU price? Measure hours between "launch" and "CUDA ready" — that is where budget burns, not on the quote's €/h line.