OVHcloud AI Endpoints is a pay-per-token API for open-weight models (gpt-oss, Llama 3.3, Qwen, Whisper, embeddings) run in OVHcloud's Gravelines data centre in France. OVHcloud says it doesn't store your prompts. Pick it when data must stay with a European provider and an open model is good enough. The catalogue is small, and models retire with three months' notice.
Other good fits
Cheap bulk jobs on open models, for example gpt-oss-120b at €0.08 per million input tokens.
Teams already on OVHcloud Public Cloud who want AI on the same invoice.
+ Pros
Read pros
Pick it when data should be processed in the EU by a European company. The infrastructure is in Gravelines, France, and OVHcloud states that data isn't stored or shared during or after use.
Pick it when an open model does the job. The catalogue covers chat and reasoning (gpt-oss-120b, gpt-oss-20b, Llama 3.3 70B, several Qwen models), speech-to-text (Whisper), text-to-speech, embeddings and an image model.
Pick it when you want to try before you sign up. Every model has a playground, and anonymous API calls are allowed at 2 requests a minute per model. With a key the limit is 400 a minute.
Pick it when your code already speaks the OpenAI API. The endpoints are OpenAI-compatible.
− Cons
Pick something else when you need Claude or GPT. Use the Claude API or a gateway such as OpenRouter.
Pick something else when you want a wider EU catalogue with Mistral Medium 3.5, GLM or DeepSeek. Scaleway serves them from Paris.
Pick something else when data must never leave your building. Run a model on your own machine with Ollama.
Avoid if
Nobody will read the retirement emails. Four models, three of them from Mistral, go offline on 1 October and 4 November 2026.
You can't add a payment method. Keys from a Public Cloud project in Discovery mode don't work.
Connection
Create a project-linked API access key, choose an active catalogue model and use its documented API route. Key creation and revocation have shared-project limitations, so establish who can manage a team key first.
Billing
Read billing
Prices checked 23 September 2026 (AI Endpoints catalogue):
Model
Input
Output
gpt-oss-120b
€0.08 per million tokens
€0.40 per million tokens
Meta-Llama-3.3-70B-Instruct
€0.67 per million tokens
€0.67 per million tokens
whisper-large-v3-turbo
€0.00001278 per second of audio
You pay per use. There are no servers to run. The running effort is elsewhere:
Keys. Admins can't revoke or audit keys that team members create. The only way to cut one off is to remove that person's access to the project.
Retirements. A retired model returns 404. Mistral-7B-Instruct-v0.3 and Qwen3-Coder-30B-A3B go on 1 October 2026; Mistral-Nemo and Mistral-Small-3.2 go on 4 November 2026. Notices go to the billing contact only, so add your developers to it.
Spend. There's no usage cap yet, only rate limits. The bill grows with tokens and audio seconds.
Data handling
OVHcloud's capability page describes its stated privacy and location scope. Review the applicable service conditions and your own handling before sending non-sample content.
First request
Documentation-based suggestion — not run here.
Transcribe and summarise a 10-minute team meeting.
In the OVHcloud Control Panel, open a Public Cloud project with a payment method, go to AI Endpoints and generate an access key.
Try whisper-large-v3-turbo in the playground with a short recording. Then send the full meeting file through the API. Ten minutes of audio costs under one cent.
Paste the transcript into gpt-oss-120b with: "List the decisions, owners and deadlines from this meeting. Quote the sentence each one comes from."
Check each item against its quoted sentence. Then look at the token and call counts in the Control Panel.
Links & evidence
Documentation-based · not hands-on · Sep 23, 2026
Based on OVHcloud's AI Endpoints catalogue and its getting-started, capabilities and billing guides, checked 23 September 2026. We did not run it for this review.
AI Endpoints catalogue · Sep 23, 2026 — Model list and prices: gpt-oss-120b €0.08/€0.40, Llama 3.3 70B €0.67/€0.67, whisper-large-v3-turbo €0.00001278 per second; rate limits 2 req/min anonymous, 400 req/min with key.
Capabilities and limitations · Sep 23, 2026 — Infrastructure in Gravelines, France; data not stored or shared; anonymous rate limit; no usage caps beyond rate limits; OpenAI compatibility.
Billing and lifecycle · Sep 23, 2026 — Decommission table (1 October and 4 November 2026), 3-month grace period, 404 after removal, notices to billing contact.
Getting started · Sep 23, 2026 — Payment method required, Discovery-mode keys can't use the service, admins can't revoke team members' keys.
Further review notes & alternatives
Alternatives and trade-offs
Scaleway wins on model choice and gives the first million tokens free. Its gpt-oss-120b costs more (€0.15 in / €0.60 out).
IONOS AI Model Hub wins when you want a German provider under a German contract.
Ollama wins when you have a suitable GPU and want no per-token bill at all.
Call open-weight models through a serverless inference API in European data centers.
Rent bare GPU instances when you want to run vLLM or Ollama yourself.
Use an OpenAI-compatible endpoint while keeping the deployment under French jurisdiction.