Confidential AI

0 reviews

Confidential AI is infrastructure for running AI privately. Teams that need to run models on sensitive data, prompts, or proprietary weights face a gap: data is protected at rest and in transit, but exposed while the model runs. Confidential AI closes that gap with hardware-encrypted execution. Inference runs inside confidential VMs on Intel TDX and AMD SEV-SNP, with NVIDIA GPUs in confidential mode. Before the first token is sent, the client verifies the model, the code, and the hardware through attestation, and can check the live production cluster in the browser with no signup. The inference API is OpenAI-compatible. DeepSeek V4.1 Flash is live at $0.30 per 1M input tokens, $0.03 cached, $1.20 output. Confidential VMs are available as cloud GPU VMs (B300 $7.50, B200 $6.50, H100 $3.25 per GPU-hour), and the same stack is licensed to run on your own infrastructure, including confidential Kubernetes for model serving, private model weights, and private training and fine-tuning. Agents run on this infrastructure as a workload: teams bring their own agent and run it where the operator cannot read memory in use.