Mistral opens an EU-only inference lane
Teams worried about GDPR and the EU AI Act have long defaulted to US-based clouds for compliance. Mistral's newly generally available Regional Endpoints let you pin inference processing to the EU — but account data and billing aren't covered yet. Here's what actually changed, straight from the primary sources.
When "location"
wasn't a choice
Mistral AI is headquartered in Paris, but until now its API gave no explicit way to choose a processing region. Teams facing GDPR (the EU's General Data Protection Regulation) or the EU AI Act's high-risk obligations, which are now moving into full force, tended to default to US-based clouds like Azure or AWS anyway — simply to avoid the risk of not knowing where their data was actually processed.
In finance, healthcare, and government, where cross-border data transfers are themselves subject to audit, this "unknown processing location" problem was a real compliance-explanation burden. Working through Standard Contractual Clauses (SCCs) or adequacy decisions case by case became one of the friction points slowing AI adoption.
| Before | After Regional Endpoints |
|---|---|
| No way to choose a processing region | Explicit choice of EU or US |
| Default processing location undisclosed | Pinned to the EU via api.eu.mistral.ai |
| Cross-border explanation was on the customer | Inference designed not to cross the region boundary |
| Flat pricing | Regional pinning costs 1.1x standard rate |
Only inference sits inside the boundary
"EU-only" refers strictly to where inference is processed. Account management and billing — the control plane — sit outside the scope of this change.
In its official announcement, Mistral confirmed that Regional Endpoints are now generally available (GA). Instead of the default api.mistral.ai, customers can send requests to api.eu.mistral.ai or api.us.mistral.ai to pin input and output token processing to that region. Pricing runs at 1.1x the standard list rate (covering input tokens, output tokens, and cache reads/writes) — this is a feature you pay for based on where processing happens, not on latency.
What's easy to miss: account configuration, API keys, billing, access management, and usage logs — the "control plane" — are not covered by regional pinning. In other words, "EU-only" describes the inference processing location alone, not full data sovereignty over contracts and operational logs. Deploying without understanding this distinction risks a mismatch at audit time.
Regional Endpoints
by the numbers
Pricing, availability, and uptime — the published figures, as-is.
Announced alongside it, Priority Tier (see the official docs) is in public preview, offering custom rate limits and priority queuing for mission-critical workloads, backed by a 99.5% uptime SLA — roughly 3.5 hours of allowed downtime per month. But eligibility depends on the model, region, and available capacity, and onboarding requires contacting a Mistral account representative directly. This isn't a self-serve, sign-up-and-go feature.
Regional Endpoints themselves still have limited add-on support. Only function calling (external API invocation) works at regional endpoints — Agents, batch processing, and file management are not available through the regional addresses. Migrating part of a workflow doesn't mean the rest follows; anything using those unsupported features still runs through the default endpoint.
Who benefits, and how
The biggest impact lands on teams that own compliance accountability.
Business
For finance, healthcare, and government deals where EU compliance is a contract requirement, proposing this as-is — "processing location is covered, control plane isn't" — can create a mismatch at audit time. Spell out inference vs. control-plane scope explicitly in RFPs and DPAs (data processing agreements).
Engineer
Migration is as simple as swapping the endpoint URL (api.eu.mistral.ai), but anything relying on Agents, batch processing, or file management stays on the default endpoint. Build a feature-by-feature regional-support map before cutting over, and fold the 1.1x token-cost bump into budget estimates.
PM / Rollout lead
Don't repeat "EU-only" verbatim in messaging without clarifying what's covered (inference) and what isn't (control plane), internally and externally. If Priority Tier is part of the plan, build the account-rep application lead time into the schedule — it isn't instant.
"EU-only" still means inference only.
Sovereignty over contracts and operational logs remains the next assignment.
What to check next
Check what your plan actually covers
Only Regional Endpoints and function calling support regional pinning today — Agents, batch, and file management don't. Map each feature you use against regional support.
Brief legal and IT on the control-plane gap
Account data, billing, and usage logs are not pinned to the region. Cross-check this against your DPA and security register wording before relying on "EU-only" claims.
Apply early if you need Priority Tier
It's public preview and requires individual setup through a sales contact. If a mission-critical workload might need it, start the conversation now.
Alongside this announcement, Mistral also outlined multi-year compute-purchase agreements with several European enterprises to build up to 200 megawatts of European compute capacity by the end of 2027 and 1 gigawatt by the end of 2030. Regional Endpoints isn't a standalone feature drop — it's positioned as part of this broader European infrastructure push, and whether regional coverage eventually extends to the control plane looks like the next thing to watch.