OVHcloud AI Endpoints Review: Open Model Catalog, Token Pricing and European Infrastructure

OVHcloud AI Endpoints delivers serverless inference through OpenAI-compatible APIs for open-weight models. Launched in April 2025 and hosted in Gravelines, France, the service targets developers seeking European data residency and GDPR compliance. It currently lists around 22 models after a reduction from more than 40 at launch.

OVHcloud AI Endpoints Review: Catalog, API Features and Infrastructure

The catalog includes text-generation models such as Qwen3-8B through Qwen3.5-397B variants, Meta-Llama-3.3-70B-Instruct, and gpt-oss-120b. Vision and multimodal options cover Qwen2.5-VL-72B, while speech services include Whisper variants for transcription and NVIDIA Riva voices for text-to-speech. Image generation uses Stable Diffusion XL, and several moderation models remain available at no cost.

API access uses the base URL https://oai.endpoints.kepler.ai.cloud.ovh.net/v1 and supports chat completions, streaming, function calling, structured outputs and vision on compatible models. A newer Responses API and batch mode with up to 50 percent discount and 24-hour turnaround have been added recently. Rate limits stand at 2 requests per minute for anonymous users and 400 requests per minute once authenticated under a Public Cloud project. Payload size is typically capped at 2 MB.

Infrastructure runs exclusively in OVHcloud’s Gravelines data center under EU jurisdiction. Certifications include ISO 27001, SOC 2 and HDS. The platform is reachable from Europe, APAC and Canada, with zero data retention beyond billing records and no use of inputs for model training.

OVHcloud AI Pricing, Ease of Use and Security Details

Pricing follows a strict pay-as-you-go model per million tokens. Examples include gpt-oss-120b at €0.08 input and €0.40 output, Qwen3.5-9B at €0.10/€0.15, Llama-3.3-70B at €0.67/€0.67 and Qwen2.5-VL-72B at €0.91/€0.91. Speech-to-text is billed per second. Several models such as Qwen3Guard variants and Stable Diffusion XL remain free. Batch jobs receive an additional discount.

Anonymous access allows immediate testing of free models at 2 requests per minute without an account. Authenticated use requires a payment method on a Public Cloud project. The web playground supports prompt testing and model selection by specification. Documentation and OpenAI SDK compatibility reduce migration effort for existing codebases.

Security centers on EU sovereignty, GDPR alignment and explicit zero-retention policy. No training occurs on customer data. Known limitations include upcoming retirement of selected Mistral variants on 4 November 2026 and a leaner catalog compared with the initial launch.

OVHcloud AI Endpoints Limitations, Comparisons and Final Verdict

Public user feedback remains limited in independent forums, with most comments focused on straightforward OpenAI SDK integration and occasional requests for lower latency tiers. No widespread complaints about billing accuracy or uptime appear in current sources. The service positions itself against larger providers by emphasizing data residency rather than raw model count or global edge presence.

Compared with similar offerings, OVHcloud AI Endpoints charges lower rates on several Qwen variants than equivalent US-hosted endpoints, yet offers fewer total models than platforms with broader proprietary catalogs. European-only hosting contrasts with services that maintain multi-region footprints, which can affect latency for users outside Europe. Batch mode pricing aligns closely with competitors offering asynchronous discounts.

OVHcloud AI Endpoints supplies a focused European alternative for developers who prioritize sovereignty and predictable token costs over the widest model selection. The platform’s recent additions of batch processing and an expanded Responses API address practical workflow needs, while the streamlined catalog reflects a deliberate emphasis on high-performing open-weight options. Continued additions of newer Qwen releases and potential performance tiers remain the main areas to watch.