Catalogue

ARK Cloud is hosted and operated in the European Union. Its terms say prompt and output data are not retained unless the customer consents or debugging requires it, and that external models are not trained on customer data. The privacy policy says content is deleted after inference by default; request logs redact prompt fields and do not log generated text. Stateful sessions and debugging are the explicit caveats.

Runs open models under Swedish jurisdiction with zero data retention, positioned against GDPR, NIS2 and Schrems II. Cached input tokens are billed at the same rate as uncached ones.

Prompts and outputs are never logged and never read by staff. Inference is stateless and nothing is written to persistent storage; no third-party sub-processors are used and processing stays in IONOS Cloud data centres in Germany.

Hosts the open-weight, non-autoregressive Laya decision model in Switzerland. Direct decision requests keep no request content, never train on it, and retain only metadata for 30 days. Swiss-only mode prevents failover outside Switzerland for a 15% surcharge. The optional chat agent uses a separate planning LLM, so this catalogue entry covers only the direct System One API.

Runs open-weight models on serverless GPUs in Paris and Finland with an OpenAI-compatible API. Its legal documentation says prompts and outputs are never written to disk or logged, request records contain usage metadata but no prompt or response text, and recent prompt prefixes exist only in GPU memory as a cache. Dedicated endpoint locations are agreed per workload, including a requested EU location. The DPA covers GDPR processing; ISO 27001 audit is in progress, not yet certified.

States that data sent through the API is not used for training, with a training opt-out that is a separate control from retention. Zero data retention is granted on request and reviewed case by case, and covers stateless calls such as chat, embeddings, OCR and speech, but not stateful products such as Batch and the Files API. Inference runs on GPU infrastructure Mistral manages, including third-party models such as GLM, and the EU endpoint pins processing to EU and EFTA data centres at a 10% upcharge.

CompactifAI is Multiverse Computing's managed serving product. We use its EU endpoint; inference stays in EU data centres. The API retains no prompts or completions and never trains on them; it keeps anonymised request logs for 30 days and minimal billing metadata. Requests are stripped of the user identifier and tagged only with a random request ID.

States that data is not stored or shared during or after model use. AI Endpoints inference is hosted in OVHcloud data centres in Gravelines, France.

Edgeless Systems' confidential-computing service encrypts prompts on the client and processes them only inside hardware-attested environments on Scaleway and Lyceum infrastructure in the EU. Request and response bodies are not retained after completion and are not used for training; operational metadata is kept for up to 90 days and token counts permanently for billing. Use the SDK or local encryption proxy for client-verifiable attestation: the hosted proxyless API provides memory encryption but not that verification.

Runs an explicit zero-retention policy: prompts and outputs are processed in memory, discarded after the response, and not logged, written to disk or reused for training. The policy has an independent zero-data-retention certificate; user-enabled storage is the stated exception. Inference runs in European data centres, including Lazio and Lombardy in Italy.

Part of the Schwarz Group and hosted exclusively in European data centres. Their FAQ states that no customer data from requests is stored and none is used to train models; shared models run in communal tenancy with no data exchange between projects or between customers. Billing follows named tiers rather than per-model rates.

Applies zero data retention by default. Error traces are kept up to two weeks, aggregated and anonymised usage up to six months, and data is held in Paris.

Strom is a native decision model that returns typed answers and calibrated probabilities without generating prose. Standard requests are processed in Helsinki and retained for 90 days, never for training. Its stronger zero-retention endpoint is enabled on request and costs 20% more; it excludes request content from application records and backups, although infrastructure diagnostics and crash dumps remain outside that guarantee.