Trust AI

Private AI requests on secure GPUs

How it works

Confidential AI

AI is powerful. Your private data is the price unless you protect it.

Prompts often contain client names, contracts, source code, medical details, financial plans, unreleased products, and private files. Trust AI encrypts them on your device, sends only ciphertext through our server, and decrypts the answer only in your browser.

No plaintext prompts in our gateway Hardware-protected GPU execution Open client build with checksum verification
See how it works

Why it matters

A normal AI request can become a permanent copy of your sensitive work.

Your prompt may pass through application logs, debugging tools, queues, vendor accounts, support systems, and retention policies before a model answers. If that text is readable anywhere along the path, a private question can become discoverable business, legal, health, or personal data.

How it works

Privacy should not depend on trusting the server operator.

GPU confidential computing creates a protected place where AI work can run while the request stays isolated from cloud operators, administrators, and ordinary server software. In plain terms: your browser locks the request, the server routes the locked box, the protected GPU opens it only inside the secure environment, and the answer is locked again before it returns.

NVIDIA confidential computing GPU TEE deep dive GPU TEE deployment guide
01

Your browser locks the request first

The prompt is encrypted on your device before it leaves the page. Our gateway receives ciphertext, not the readable question.

02

The server routes and bills without reading

The backend checks balance, forwards the encrypted request, and records only operational metadata such as model, estimates, and proof hashes.

03

Protected GPU compute opens it only inside the secure environment

The model runs inside hardware-protected compute. Attestation is the proof that the expected secure environment exists before secrets are sent.

04

The answer comes back locked

The response is returned encrypted and decrypted only in your browser. The client also shows hashes and signature data so the path can be checked.

What changes

The server should route and bill. It should not read.

Protected while running

Confidential GPU execution protects AI code and data while the model is processing it, not only while files are stored or sent over the network.

Verifiable hardware

The client checks fresh attestation evidence and the model encryption key before it sends encrypted requests.

Open client path

The downloadable client is built from public source so users can compile it, calculate its checksum, and compare it with the published release.

Trust, but verify

Before secrets move, the client checks the secure environment.

The client fetches fresh attestation evidence, checks the model encryption key, encrypts locally, and shows proof hashes after the response. You can use the public client source to build the same user client and compare checksums.