# AI and GPU workloads through OCRE

Research data that cannot leave European jurisdiction can still be trained on,
fine-tuned and served. Exoscale runs the GPUs, VSHN operates the platform around
them, and both are procured through the GÉANT OCRE 2024 framework with a 10%
discount on compute and storage.

## Which AI models are available?

OCRE procures cloud services, not models, so the framework itself has no model
catalogue and no model approval list. What you procure here is an open platform:
you choose the model, Exoscale and VSHN run it on European hardware.

- **Dedicated Inference** serves any model on Hugging Face, public, gated or
  private, and your own weights next to them. In practice that means the
  families research teams ask about: **Apertus**, the Swiss AI foundation
  model, alongside Llama, Mistral, Qwen, DeepSeek and Gemma, with Whisper for
  speech and open embedding models for RAG. Exoscale's own documentation names
  `openai/gpt-oss-20b`, `openai/gpt-oss-120b` and `Qwen3-Coder-480B` as
  examples of what teams deploy there. Neither list is a catalogue; both
  illustrate a service that runs anything on Hugging Face.
- **GPU servers** run whatever you build on top: a fine-tuning job, your own
  vLLM or llm-d deployment, a training run across several cards.
- What decides whether a given model fits is **GPU memory** in the zone you
  pick, from 16 GB on a V100 to 8x 288 GB on a B300 HGX, not an approval list.

Closed models are the exception. **GPT-4, Claude and Gemini** are not part of
this offering: their weights are not licensed for hosting outside their vendors
and those vendors' own clouds, so through OCRE they are reachable only via a
hyperscaler offering, with the CLOUD Act exposure the rest of this page exists
to avoid.

Send us the model or the model size, the expected load and the zone, and you get
a written estimate back.

## Dedicated Inference

Available now. Run **any Hugging Face model** as a production API endpoint on
dedicated European GPUs.

- **OpenAI-compatible API**, so existing tooling and SDKs work unchanged
- Public, gated or private models; weights cached in Exoscale object storage
- Exoscale operates the GPUs, the Kubernetes underneath, the load balancers and
  the model serving
- Manual replica scaling and scale-to-zero, so an endpoint costs nothing while
  idle
- Billed per GPU-second by GPU model, no per-token charge and no upfront fee,
  plus model storage
- **99.95% monthly endpoint availability SLA** from Exoscale, with service credits

Available in `at-vie-2`, `ch-dk-2`, `de-fra-1` and `hr-zag-1`.

Current list prices are on
[exoscale.com](https://www.exoscale.com/ai-cloud-infrastructure/dedicated-inference/);
the OCRE discount applies on top.

## GPU servers

Billed per second, with NGC containers, Terraform support, and GPU node pools
that attach to an SKS Kubernetes cluster so ML pipelines schedule onto GPUs like
any other workload.

| GPU | Memory | Generation | Zones |
|---|---|---|---|
| RTX Pro 6000 | 96 GB | Blackwell | de-fra-1, ch-dk-2, hr-zag-1 |
| A40 | 48 GB | Ampere | de-fra-1 |
| A30 | 24 GB | Ampere | ch-gva-2 |
| RTX A5000 | 24 GB | Ampere, liquid-cooled | at-vie-2 |
| Tesla V100 | 16 GB HBM2 | Volta | at-vie-1 |
| B300 HGX | 8x 288 GB | Blackwell Ultra, NVSwitch | ch-gva-2, on request |

The RTX A5000 machines in Vienna are liquid-cooled and their waste heat is
recovered.

## Vector search, today

Retrieval-augmented generation needs somewhere to put the embeddings, and both
options are managed services rather than something your team runs:

- **Managed PostgreSQL with pgvector**, in every zone, included in the plan price
- **Managed OpenSearch with vector search**, for semantic search at scale

Together with Dedicated Inference, that is a complete RAG stack inside EU and
Swiss jurisdiction, with no data leaving it.

## Managed Inference: announced, not yet available

Exoscale has announced a per-token managed inference service with a curated
catalogue of production-ready models behind a managed API: no GPUs to size and
no idle cost. Planned coverage includes assistants, coding models, small models,
embeddings, OCR and speech, deployed in Europe and Switzerland.

**The waitlist is open and the service is not yet generally available.** If it
would change your architecture, tell us and we will put you on it.

## Why this matters for research

Hyperscaler AI services raise the same Schrems II questions as any other US-owned
processing, and a grant condition or an ethics approval that requires European
data residency does not bend for a model endpoint. Every component here runs in a
zone you choose, operated by companies with no US parent.

Ask us for a written estimate for your workload, or a trial voucher to benchmark
your own model before committing.