French exits for francophone corpora and Mistral-era evaluation

Why France has its own AI collection surface

France sits at the centre of a distinct European AI research cluster — Mistral AI, Kyutai, Hugging Face's Paris presence, LightOn, plus the Saclay research corridor (CEA List, Inria, Polytechnique). For an AI team doing:

• pre-training corpus work on francophone content (France, French Belgium, French Switzerland, Quebec, Maghreb)
• evaluation against French-speaking knowledge benchmarks (BLIMP-fr, Meta's FrenchBench, CNIL-aligned safety tests)
• competitive intelligence against Mistral or other francophone model APIs

...a French-resident exit materially changes the corpus composition and model responses you collect.

What the FR pool routes through

• Datacenter — AWS eu-west-3 (Paris), GCP europe-west9, OVHcloud (Strasbourg, Gravelines, Roubaix), Scaleway DC3 (Paris).
• Residential — Orange, Free (Iliad), SFR, Bouygues with département-level targeting in Île-de-France and Lyon.
• Mobile — Orange, Free Mobile, SFR, Bouygues 4G/5G.

Good fit

• French-language training-corpus collection from .fr, .qc.ca, .be, .ma, and .tn surfaces
• Scraping of French government open-data (data.gouv.fr) and Eurovoc multilingual alignment resources
• Evaluation of French-tuned models from an FR-resident vantage
• RAG ingestion of French institutional content (Cour de cassation, Légifrance, Insee)

Not a good fit

• Scraping of personal contact data under CNIL's 2023 guidelines (fines in the low six figures are routine at this point)
• LCEN-regulated content categories (terrorism, CSAM, illegal gambling)
• Circumvention of robots.txt or machine-readable opt-outs on publishers that have declared TDM reservation

ASN depth in the French residential pool

French broadband concentrates around the four historic operators:

• Orange (AS3215) — largest residential footprint nationally, including FTTH and legacy ADSL.
• Free (Iliad / AS12322) — second-largest, strong urban FTTH with distinctive subscriber profile (technically-literate, younger demographic skew).
• SFR / Altice (AS15557) — cable + FTTH, strong in specific metro footprints.
• Bouygues Telecom (AS5410) — FTTH + DSL, smaller but material share.

Mobile: the same four operate the 4G/5G networks, with Free Mobile notably disruptive on pricing post-2012. The mobile pool samples across all four.

Top French metros

• Paris / Île-de-France — dominant. Government, press (Le Monde, Le Figaro, Libération), AWS eu-west-3 home region, financial services (La Défense). Default origin for most FR-anchored AI work.
• Lyon — second metro, industrial / pharma / research, regional press (Le Progrès).
• Marseille — Mediterranean connectivity hub (major landing station for Mediterranean cables), regional press (La Provence).
• Toulouse — aerospace cluster (Airbus), research (LAAS-CNRS), regional press (La Dépêche).
• Bordeaux, Lille, Nantes, Strasbourg — secondary metros with meaningful regional content footprints.

French legal landscape for AI data collection

• RGPD (French GDPR) + Loi Informatique et Libertés — CNIL is the supervisory authority and one of the most active enforcement bodies in the EU. CNIL's 2023-2025 guidance on AI training data sets concrete expectations: documented legitimate-interest analysis, demonstrable technical and organisational measures, and opt-out mechanisms for data subjects.
• Code de la Propriété Intellectuelle (Art. L.122-5-3) — TDM exception with an explicit opt-out mechanism for rightsholders. The opt-out can be machine-readable or contract-based. SquadProxy supports gateway-level opt-out honoring.
• LCEN (Loi pour la Confiance dans l'Économie Numérique) — regulates illegal content online. Certain content categories (terrorism, CSAM, illegal gambling) are criminally restricted regardless of training purpose.
• EU AI Act enforcement — France enforces via DGCCRF + CNIL coordination. French implementation particularly strict on foundation-model provider obligations.

Latency and edge routing

FR edge routes through AWS eu-west-3 (Paris). Latency:

• OpenAI API: ~25-40ms p50
• Anthropic API: ~30-45ms p50
• Google Gemini (europe-west9): ~15-30ms p50

Residential adds 50-200ms; mobile 80-250ms.

Related

• French residential proxies for AI
• French 4G mobile proxies
• Residential vs datacenter routing