Prem AI

This is the base layer of the encrypted AI API: your application encrypts the request body before any network call. The gateway handles ciphertext plus operational metadata such as model name, size, timing, and billing. Only the processing runtime decrypts the payload to run inference. The practical effect is that a compromised gateway log or a curious operator doesn't surface your prompt text, and a local proxy variant gives the same protection to clients that only speak OpenAI's format. Prem never receives the key encryption key.

Each protected request can return attestation evidence: signed claims from the CPU and GPU, nonces to prove freshness, and policy results that say whether the measurement matched an approved reference. Prem publishes the verification code on GitHub, so a security team can run the check itself instead of trusting a dashboard badge. The evidence covers AMD SEV-SNP, Intel TDX, and NVIDIA confidential computing, including multi-GPU clusters where every GPU claim gets verified.

Confidential inference puts the whole path through encryption and a protected runtime, aimed at sensitive data and production workloads. Zero Data Retention is a plain OpenAI-compatible API endpoint that needs no key management and no proxy, where Prem routes the request to a partner under a no-retention contract. That second mode is the cheap way to test models, prototypes, and public data. Both modes expose different catalogs, and the API rejects a model sent to the wrong one with a clear 400 error.

The catalog leans on current open-weight models: Qwen 3.8, DeepSeek V4 and V4.1 variants, GLM 5.2 and 5.3, Kimi K3, and DiffusionGemma for decision tasks. Prices are per million tokens and published for both input and output, with cached prompt tokens billed at a discount. Confidential chat starts around $0.30 to $0.35 per million input tokens on the cheapest models, and audio transcription runs $0.05 per minute. You see the rate before you send anything.

Chat covers text, image, and video input depending on the model. A voice transcription route handles audio and returns Deepgram-compatible transcriptions. A decisions endpoint answers typed questions with probabilities, which suits classification and scoring jobs where you want a number rather than prose. One API key and one base URL reach all of them, which cuts down on the glue code a regulated pipeline usually needs.

The managed service runs in Swiss data centers with an uptime commitment near 99.98% and renewable power. For teams that can't use a hosted API, the Enclave hypervisor installs on existing GPU clusters, whether on-premises, in a public cloud account, or in a private cloud, and Prem says it can't read data processed there. Enterprise deployments go further: dedicated instances, custom model choices, and air-gapped setups.

Prem's broader pitch is sovereign AI: data, compute, and models that stay under your control, sold as three products around one account. Fluso is a no-code agent builder for teams that want to deploy agents against internal procedures without writing code. Prem Studio fine-tunes and compresses models on your own data, with exported weights you can self-host. Neither replaces the API, but they matter if you'd rather own a smaller model than rent a large one.