// USE-CASE EXAMPLES

What people actually run here

Four realistic scenarios from the teams using SAIG Compute — from fine-tuning a domain model to serving a GDPR-auditable API for a Dutch municipality.

Every workspace starts the same way: pick a GPU class, attach your data, choose how to connect. The use case shapes which defaults make sense. Below is what that looks like in practice.

Research & training

Fine-tuning a domain-specific model on proprietary corpora

A university AI lab needs to fine-tune Mistral-7B on 40 GB of Dutch legal texts. The run is expected to take 18–24 hours. The corpus contains personally identifiable information — it cannot leave the EU, and the institution's DPO requires a contractual data-processing guarantee under EU law.

Sovereignty angle

Data and training run under full EU jurisdiction. Nixpay B.V. acts as data processor — GDPR Article 28 compliant. The model weights and training data never leave the EU.

What you provision

GPU class: high-memory (A100-class) · Image: PyTorch + Hugging Face · Access: SSH + Jupyter · Region: Netherlands · S3 bucket for training data and checkpoints

How

Upload your corpus to the S3 bucket, mount it into the workspace at launch, run your training loop via SSH or Jupyter. Checkpoints land back in S3 automatically. Tear down the compute when training completes — the checkpoint survives.

Getting-started guide
Start a workspace
Product / inference

Serving an OpenAI-compatible API for a production application

A Dutch legal-tech startup needs to serve Llama-3-8B-Instruct behind an OpenAI-compatible endpoint. Their product integrates via the `/v1/chat/completions` path. They need reliable uptime, a fixed API key, and assurance that no request data flows through US-cloud infrastructure.

Sovereignty angle

Every inference request is handled on European compute under EU jurisdiction. No US cloud, no CLOUD Act exposure. Your users' prompts and completions stay in the EU.

What you provision

GPU class: mid-range (L4-class) · Image: vLLM · Access: Inference API endpoint (OpenAI-compatible) · Region: Netherlands · Persistent volume for model weights

How

Choose the vLLM image and the Inference API access method in the wizard. Your endpoint is ready within minutes. Point your application at the provided URL with your bearer token — zero changes to your existing OpenAI SDK calls.

API inference guide
Start a workspace
Public sector / compliance

Auditable AI inference for a Dutch municipality

A municipality wants to pilot an AI assistant for citizen-facing services. The procurement officer requires: GDPR-compliant data processing, data residency exclusively in the Netherlands, and an audit trail of infrastructure decisions. The team has no in-house GPU expertise.

Sovereignty angle

Nixpay B.V. signs a Data Processing Agreement (Article 28 GDPR). Infrastructure is Netherlands-only. NIS2 alignment through a Dutch-owned operator with a documented, auditable supply chain.

What you provision

GPU class: entry-level · Image: vLLM or Jupyter · Access: private network (WireGuard) + Inference API · Region: Netherlands · Guided setup intake recommended

How

Start with a free guided-setup intake call. An engineer scopes the deployment, configures the WireGuard private network to your existing IT environment, and produces a written infrastructure description you can share with your DPO and auditors.

Getting-started guide
Talk to an engineer
Creative / media generation

Image and video generation for a European media studio

A creative studio generates AI imagery and short video clips for advertising campaigns. Runs are bursty — heavy for a few hours, then idle. The studio needs EU data residency because some generated assets include client-proprietary brand assets under NDA.

Sovereignty angle

All generated assets and prompt inputs stay in the EU. Egress is free — download your finished assets without paying for bandwidth. Client NDAs are enforceable under EU contract law with a Dutch counterparty.

What you provision

GPU class: mid-range (L4-class) · Image: ComfyUI or custom PyTorch · Access: Jupyter + SSH · Region: Netherlands or Germany · S3 bucket for assets

How

Top up GPU credit for the burst session, spin up the workspace, generate, export assets to S3, tear down. Billing stops the moment you stop the machine. The hard stop means no runaway meter if a session runs longer than expected.

Getting-started guide
Start a workspace

What people actually run here

Every workspace starts the same way: pick a GPU class, attach your data, choose how to connect. The use case shapes which defaults make sense. Below is what that looks like in practice.