Enterprise AI adoption has a trust problem, and it's earned: the industry spent two years asking companies to send their most sensitive data to black boxes with vague retention policies. Our answer is an architecture where capability never requires custody — frontier models, SOC 2 Type II controls, and a mode where your data provably ceases to exist the moment your response is delivered.
Zero retention, by construction
With zero-retention mode enabled, prompts, files and outputs live only in volatile memory on the inference node. Nothing touches disk: no logs of request bodies, no caches that outlive the response, no "deleted after 30 days" asterisks. When the last byte streams back to you, the buffers are freed and the data is gone. We verify this the hard way — memory forensics on production nodes are part of our quarterly penetration test, and the pen-test summary is available to enterprise customers under NDA.
The trade-offs are real and we state them plainly: zero-retention disables conversation history, prefix caching and support's ability to replay a failed request. Enterprise customers choose per-project — most run product workloads with standard 30-day retention and route regulated workloads through zero-retention.
Your region, your keys
Region pinningguarantees inference happens in the jurisdiction you choose — EU, US, or APAC — enforced at the routing layer, not by policy document. A pinned request that can't be served in-region fails closed rather than silently crossing a border.
- BYOK encryption: data at rest is encrypted under keys you hold in your own KMS. Revoke the key and every byte we store becomes ciphertext we cannot read.
- Per-workspace isolation: tenants are separated at the compute layer, not just the database — dedicated inference clusters are standard on Enterprise.
- No training, ever: customer data is never used to train models on any plan. This is contractual, not a settings toggle someone can flip.
Compliance without theater
We hold SOC 2 Type II (report available under NDA), are GDPR and HIPAA ready with signed BAAs, and publish our subprocessor list with 30-day change notifications. Every workspace gets a full audit log — every request, key, actor and model — exportable to your SIEM. Compliance documents describe the system; the audit log proves it.
The best data-breach story is the one where there was nothing to breach. Zero retention isn't a policy promise — it's an architecture where the promise can't be broken quietly.
What's next
This quarter we're shipping customer-managed logging (your audit trail streams to your infrastructure in real time) and confidential-compute inference on AMD SEV-SNP for workloads where even our own operators must be outside the trust boundary. Security reviews move at the speed of your slowest question — if you have one, our team answers architecture deep-dives directly. Ask.