Platform Overview

Traditional AI inference providers can access the data you share with the AI model. While your data is encrypted from the user to the provider via TLS/SSL, it is decrypted by the provider before being processed by the AI model. The provider has the ability to access, train on, or even sell your data, with only legal contacts like data processing agreements (DPAs) to protect you. We call this pinky-promise security.

With Tinfoil, all your data is encrypted directly to the GPU running the AI model, where it is securely processed. The AI model runs inside a secure enclave environment, ensuring end-to-end confidentiality. Nobody — not even Tinfoil — can access your data. Tinfoil runs each model inside a dedicated confidential computing GPU, ensuring that your data is not shared with third parties or even accessible by Tinfoil. But you don't take our word for it: these guarantees are verifiable thanks to the transparency of the Tinfoil platform.

Until recently, secure enclaves (like those on your phone to protect your biometrics) were restricted to CPU-only workloads and not powerful enough to run full-scale AI models.

With NVIDIA's recent release of confidential computing mode on their Blackwell and Hopper GPUs, it is now possible to couple secure enclaves with powerful GPUs to run full-scale AI models with the guarantees of traditional secure enclaves. These state-of-the-art GPUs in confidential-compute mode achieve the same performance as traditional (non-confidential) deployments.

Learn about GPU confidential computing