Confidential AI inference
Prompts are readable on your device and inside the enclave. The network, the host and our operators handle only ciphertext.
Enclave
The GPU runs in confidential computing mode, the CPU in SEV-SNP or TDX. Model, prompt and keys exist in plaintext only here.
Enclave
The GPU runs in confidential computing mode, the CPU in SEV-SNP or TDX. Model, prompt and keys exist in plaintext only here.
Host OS and hypervisor
Memory stays encrypted while in use. The host and the hypervisor see only ciphertext.
Operators and cloud provider
There is no privileged access path into the enclave. Alexandria operators hold no keys.
Network
TLS terminates inside the enclave. The network carries ciphertext only.
Measured at every layer
Before the first token, the client checks hardware, firmware, build and model measurements against published values.
Supports every open model
Trust Center
Documents for security, legal and procurement review, sent on request.
- Security whitepaper
- Architecture overview
- Attestation guide
- Data processing agreement
- Subprocessor list