Ambra runs open-weight models behind an OpenAI-compatible API. Prompts and completions pass through our systems to be answered and are not kept. We do not train on them, and we do not read them.
Zero data retention
The content of a request (the prompt, any system message, tool definitions and the completion the model returns) is held in memory only for as long as it takes to produce and deliver the response. It is not written to disk, to a database or to a log, by the gateway or by the inference servers behind it. There is no option to turn content logging on for an account, and no copy is taken for debugging, quality review or abuse monitoring.
The inference servers keep a short-lived cache of computed model state in GPU memory, which is what makes a repeated prompt prefix faster. It is volatile, is never written to storage, and is evicted as new requests arrive and cleared whenever a server restarts.
No training
Ambra does not use prompts or completions to train, fine-tune or evaluate any model, and does not sell, share or license them to anyone. The models we serve are published open-weight models, run as released by their authors.
What we do record
For billing, capacity planning and reliability we keep one metadata record per request. It holds the following and nothing else:
- the time of the request and a request identifier we generate;
- the name of the API key that made it (the key itself is stored only as a hash);
- the model, the endpoint, whether the response was streamed, and which server answered;
- prompt, completion and cached token counts, and the resulting charge;
- the HTTP status, the outcome, the time to first token and the total duration.
These records contain no request or response content and no end-user identifiers. Our gateway does not record the IP addresses of API callers. Usage records are kept for as long as they are needed for invoicing, accounting and tax purposes, and then deleted.
Where requests are processed
Inference currently runs in data centers in North America. The region of every model is published in our model list at /v1/models, and a request is answered in the region listed there. All traffic to the API is encrypted in transit with TLS, and the inference servers accept connections only from our own gateway.
Subprocessors
Ambra is moving inference onto its own compute nodes. Until that capacity is in service, requests are served on dedicated GPU instances rented from an infrastructure provider. The companies below can technically handle request content in transit or in memory; none of them stores it on our behalf, and none is permitted to use it for any purpose of its own.
| Subprocessor | Purpose | Location |
|---|---|---|
| RunPod, Inc. | GPU instances that run the models, and the instance that runs the gateway | North America |
| Cloudflare, Inc. | Network edge: TLS termination and the tunnel to the gateway | Global edge network |
Stripe, Inc. (United States) issues our invoices and processes payments. It receives billing contact details and invoice amounts, and never receives request content.
This list is kept current. A change of subprocessor or of processing region is published here before it takes effect.
Security
Access to the API is by bearer key over TLS. Keys are stored hashed and can be revoked at once. The model servers are not reachable from the public internet: they listen on a private interface behind a second, separately keyed connection from the gateway. Production access is limited to the founding engineering team and uses individual credentials. Because content is never stored, there is no content at rest to breach, export or disclose.
Ambra does not yet hold a SOC 2 report. If you find a vulnerability, or need a security questionnaire completed, write to contact@ambracompute.com.
Requests from authorities and from users
We cannot produce prompts or completions in response to a legal demand or a data subject request, because we do not have them. Requests that concern usage metadata or account details are handled under the privacy policy.
← Compliance