Deploy to Scaleway¶

Scaleway is the most "sovereign AI"-aligned of the EU-sovereign hosts. HQ in Paris, data centres in Paris, Amsterdam, Warsaw. Runs Mistral's own infrastructure. Good for a political optics story.

OSS v0.1 is not one-click

The instructions below are a manual flow. Commercial edition ships a tested Terraform module that provisions the whole stack in one terraform apply.

Step 1 — provision an instance¶

Use the Scaleway Console or the scw CLI.

Recommended:

Instance: PLAY2-PICO for dev (1 vCPU, 2 GB) or PLAY2-NANO for small production (2 vCPU, 4 GB) or PRO2-S (4 vCPU, 16 GB NVMe) for comfortable production.
Image: Ubuntu 24.04 LTS
Region: fr-par-1 (Paris) — pairs with Mistral's data region.
Security group: allow 22/80/443.

With scw:

scw instance server create \
  type=PRO2-S \
  image=ubuntu_noble \
  name=lex-custis-prod \
  zone=fr-par-1 \
  ip=new

Step 2 — same as Hetzner¶

The rest is identical to the Hetzner walkthrough: install Docker, clone, ./install.sh, Caddy + DNS for TLS.

See Deploy to Hetzner steps 2–7.

Scaleway-specific advantages¶

Managed services (production)¶

If you want managed Postgres + managed Redis instead of running them in compose:

# Managed Postgres, 2 CPU / 4 GB / 80 GB, EU-sovereign
scw rdb instance create \
  engine=PostgreSQL-16 \
  region=fr-par \
  node-type=DB-GP-S \
  volume-size-gb=80

# Managed Redis
scw redis cluster create \
  version=7.0 \
  zone=fr-par-1 \
  node-type=RED1-XS \
  user-name=lexcustis

Then edit .env to point DATABASE_URL + REDIS_URL at the managed endpoints and remove postgres + redis services from docker-compose.yml (or make them no-op).

Scaleway KMS (commercial preview)¶

Scaleway's Key Manager is how the commercial edition holds the AUDIT_HMAC_MASTER_KEY outside Postgres. Not integrated in OSS v0.1 — the core still reads the key from env.

Scaleway Object Storage for backups¶

EU-region, S3-compatible. fr-par bucket lifecycle policy: transition to Glacier Deep after 30 days, keep for 10 years (Art. 19 retention).

GPU instance for Ollama¶

GPU-3070 (1 × RTX 3070, 8 GB VRAM, €0.90/hr) runs Llama 3.1 8B at interactive speed. GPU-H100-1-80G for frontier models.

Common pitfalls¶

Default Scaleway Ubuntu image has ufw disabled — you get the security-group firewall from Scaleway, but double-check before opening extra ports.
Scaleway's default volume is the lazy-loaded Block Storage — fine, but NVMe is faster for Postgres workloads. PRO2-S comes with local NVMe by default.
Billing in EUR by default. Convenient.
Availability sometimes tight in fr-par-1 — fall back to nl-ams-1 or pl-waw-1 (still EU-sovereign per SCC).

Political / marketing angle¶

If your pitch is "EU-sovereign AI with no data leaving Europe", Scaleway Paris + Mistral + BGE-M3 local + self-hosted Qdrant is the cleanest story. Every piece of the stack runs on French infrastructure, run by a French company, under French law.

That's the exact story a French MSA wants to hear when a national bank deploys your product.