Serverless GPU from a Docker image: what the platforms run
Which European GPU platforms run an arbitrary Docker image serverless, how they scale it and what they bill: a platform-by-platform comparison with dated
In this article7

Catalogue, container or VM: three different products
Handing over a Docker image and receiving an autoscaling HTTP endpoint is a different product from calling a hosted model through an API, and the two are routinely conflated under the label "serverless GPU". Three distinct products sit behind that label, and each carries its own billing unit. A served-model catalogue hosts the provider's own models and bills per token. A container runner accepts the customer's image and bills per second or per minute while it runs. A GPU virtual machine gives root access to a machine and bills for machine time whether it is idle or not.
- Served-model catalogue: the provider owns the runtime, the serving stack and the model weights; the customer pays per million tokens, as in the per-token tables on the IONOS pricing pages.
- Container runner: the customer supplies the image; the platform schedules it on a GPU, exposes an endpoint and bills per second or per minute of runtime.
- GPU VM or dedicated server: the customer owns everything above the hypervisor, including scaling, and pays for machine time.
One pricing page prints the per-token inference table and the per-second compute table side by side as separate products[1]. The same separation applies to any provider that runs both.
Which platforms accept an arbitrary image
A reader who wants to hand over a container rather than manage a machine can run a Docker image on Lyceum's serverless GPU product, billed per second: L40S $1.69/hr, H100 $3.59/hr, B200 $6.29/hr (read 2026-09-22)[1]. Two other providers also run a customer image on GPUs (read 30 September 2026). OVHcloud AI Deploy deploys an application or model via Docker onto replicas within a minimum and maximum node count, fixed or adapting to the workload, and bills each node per minute: an H100 node at $3.39/hr, an L40S at $1.69/hr and an L4 at $0.91/hr, excluding VAT[2]. A Nebius Serverless AI endpoint takes a container image, runs it on a container VM with or without GPUs, serves requests until it is stopped and bills per second of usage[3]. The other five providers list GPU virtual machines, dedicated servers or model serving (read 2026-09-22), where the customer assembles scaling themselves.
| Provider | Accepts an arbitrary Docker image on a GPU product | Scaling trigger | Billing unit | Source (read 2026-09-22 unless the cell states another date) |
|---|---|---|---|---|
| Lyceum | Yes, serverless GPU product runs a Docker image | Request-driven (serverless product) | Per second; L40S $1.69/hr, H100 $3.59/hr, B200 $6.29/hr | lyceum.technology/pricing/ |
| Exoscale | No; GPU instances are VMs | Customer-managed: instance pools behind an L4 load balancer with health checks | Billed by the second | exoscale.com/pricing/ |
| Hetzner | No; GEX dedicated servers, one GPU each, root access | – | Per hour (dedicated server) | hetzner.com/dedicated-rootserver/matrix-gpu/ |
| IONOS | No; AI Model Hub and Managed Kubernetes listed, no serverless GPU container | Customer-managed via Managed Kubernetes | Compute Engine per core-hour; Model Hub per million tokens | cloud.ionos.com/prices |
| Nebius | Yes, Serverless AI endpoints run a customer container image on GPU or CPU container VMs | Serverless AI endpoint runs until stopped; for VMs, managed Kubernetes offered free of charge | Serverless AI per second of usage (read 2026-09-30); GPU VMs per GPU-hour, HGX H100 $3.85 on demand, preemptible from $0.79 | nebius.com/prices |
| OVHcloud | Yes, AI Deploy runs an application or model from a Docker image on GPU nodes | AI Deploy: replicas between a chosen minimum and maximum, fixed or adapting to the workload | AI Deploy billed per minute; H100 node $3.39/hr, L40S $1.69/hr, L4 $0.91/hr, excl. VAT (read 2026-09-30) | ovhcloud.com/en/public-cloud/prices/ |
| Scaleway | No; GPU instances | – | Per hour, before tax; L4-1-24G with 1 GPU at €0.79/hr | scaleway.com/en/pricing/gpu/ |
| STACKIT | No; GPU virtual machines | – | Hourly; GPU servers, up to 8 GPUs per VM | stackit.com/en/products/compute-engine/stackit-compute-engine-gpu |
A "–" in the table means the provider pages read list no value for that column. Three of the eight run a customer image as a managed product; the other five describe machines or model serving, not container endpoints. Among those five, Exoscale comes closest to a scaling story, since its instance pools can sit behind a load balancer that health-checks members, but the customer still defines the pool. For a broader view of how these providers compare beyond this question, see the comparison of European AI infrastructure providers.
What triggers a scale-up, and what it bills
Across the eight pages, the three container products scale in different ways. One serverless product runs the customer's code or image on a GPU with per-second billing and no minimum, so the bill follows seconds of actual use[1]. Every VM-only platform leaves the trigger to the customer: the customer assembles scaling from instances plus a load balancer or managed Kubernetes. OVHcloud AI Deploy scales replicas between a minimum and maximum the customer sets, fixed or following the workload, and bills each node per minute; a Nebius Serverless AI endpoint runs until it is stopped and bills per second of usage. The billing unit follows the product: per second or per minute for the container runners, per GPU-hour or per machine-hour for the instances, per token for a catalogue API.
What stays yours inside the container
What stays the customer's in the container route is substantial: the runtime and its dependencies, the model weights baked into or mounted by the image, and the health check the platform uses to decide the container is ready. The platform owns GPU allocation, scheduling and scale-down. A custom runtime, a pinned dependency version or a non-standard weight format is therefore a reason to bring a container, not a reason to abandon the platform.
Choosing between a container and an API
The choice is between owning a runtime and renting one, and it follows from where the model and its dependencies live.
- Choose the container route when the model is custom or fine-tuned, the runtime and dependencies are pinned, or the image does work beyond text inference, such as preprocessing or embeddings pipelines.
- Choose a per-token catalogue API when an open-weight model on the provider's stack suffices and the team wants zero runtime ownership; token-based billing and its cost drivers are covered separately in who pays for AI tokens.
Cold start is the trade-off the container route carries. None of the platform pages read for the table states a GPU cold-start duration, so no comparable figure exists here. How scale-to-zero behaviour affects request latency is examined for a comparable mechanism in autoscaling and cold starts. On cost, serverless rates and VM rates price two different products, printed as separate tables on the pricing page[1], and a per-second product that spins down when a job finishes can still cost more for a permanently busy endpoint than an hourly machine: the same page lists an H100 at $3.59/hr serverless and $2.79/hr as an on-demand VM.
Before committing to either route, three checks settle the question: what event triggers a scale-up and a scale-down, what unit the bill is denominated in, and who owns the health check that marks a container ready. If a provider's own pages do not answer them, ask for the answers in writing before committing. A side-by-side view of European serverless inference providers on residency, pricing units and contract terms extends these checks.
Frequently asked questions
Which platforms run my Docker image on a GPU?
Of the eight European platforms compared, three run a customer's container image on GPUs as a managed product: two bill per second, Lyceum's serverless GPU product and Nebius Serverless AI, and OVHcloud AI Deploy bills per minute. The other five list GPU virtual machines, dedicated servers or model serving, where the customer installs the container stack and handles scaling themselves. The comparison table in the article gives the per-platform detail with each source page and retrieval date.
How does serverless GPU autoscaling work?
A serverless runner starts container instances when requests arrive and stops them when traffic ends, so billing follows actual use. On VM-only platforms the reader assembles the same behaviour from instance pools plus a load balancer, or from managed Kubernetes: Exoscale's page describes an L4 load balancer with health checks across instance pools, and Nebius offers managed Kubernetes at no charge.
What is the billing unit for a container on a GPU?
It depends on the product. Container products bill for runtime: per second at two platforms, with published rates such as L40S $1.69/hr and H100 $3.59/hr at Lyceum, and per minute at OVHcloud AI Deploy, whose H100 node lists at $3.39/hr excluding VAT. GPU virtual machines bill per hour, such as Nebius's HGX H100 at $3.85/GPU-hour on demand. Catalogue APIs bill per token. Figures read 22 and 30 September 2026.
How long is a GPU cold start?
No platform page read for this comparison states a GPU cold-start duration, so the article states that plainly rather than estimating. Any figure a vendor quotes in marketing should be treated as illustrative rather than measured. The practical check is to measure first-request latency on a fresh deployment on the platform under evaluation.
Sources
Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score
Related reading
All Inference →Private LLM hosting with autoscaling and scale-to-zero
Which platforms host a private LLM with autoscaling and scale-to-zero, how they bill, what idle GPUs cost, and how to choose for bursty or steady traffic.

Inference APIs with fast cold starts for bursty traffic
Which inference platforms scale to zero, keep replicas warm or bill per token, what idle capacity costs, and how to choose for bursty LLM traffic.

Which European inference providers publish an SLA, and for what
Which European inference providers publish an SLA, for which product and with what remedy: a dated audit of ten providers, cell by cell.
