confidential AI.
verifiable by you.
we're building UK inference for open-weight models.
verify what runs before sensitive data reaches it.
your SDK checks fresh CPU and GPU evidence, the approved workload, and the endpoint’s identity. your sensitive request stays with you until verification passes.
your request travels over a verified TLS connection. the network routes ciphertext; TLS terminates at the attested gateway, inside the confidential environment.
the open-weight model works with plaintext in protected CPU and GPU memory. the hardware boundary is designed to prevent host administrators from reading that memory.
the response returns to your application. prompt and response payloads are not retained after processing. only allowlisted operational metadata has a separate retention schedule.
a ZDR policy promises that prompts and responses won't be retained. on its own, it doesn't let you verify which software handles them, whether logging is disabled, or who can access them during processing.
bring inference in-house and you also take on GPU procurement, installation, power, and capacity planning. security updates, maintenance, and idle capacity remain yours to manage.
use reserved capacity with a workload and endpoint you can verify. our lightweight, source-available SDK runs in your infrastructure, checks attestation, and forwards your OpenAI-compatible requests only when verification passes.
prompts and responses are processed in protected memory, then cleared when the request is complete.
we’re seeking design partners for confidential inference.