Why 0G Private Computer

Most AI APIs ask you to trust a policy document. 0G Private Computer replaces policy with structure: hardware isolation, cryptographic proof, and an on-chain record you can read without asking anyone’s permission.

This page is for the person deciding whether a workload belongs here. Every claim below is followed by how you check it yourself.

Sovereignty over your own workload

Sending a prompt to a hosted model usually means handing over three things at once: the content, the record that you sent it, and any leverage you had over the provider.

On 0G, those stay separable. Your wallet is your account identity. Your balance sits in an on-chain payment contract, debited per request. The model catalog, the providers serving it, and the cost of each call are all readable data rather than a monthly invoice you reconcile after the fact.

How to verify. The model catalog is a public endpoint — GET /v1/models needs no authentication, and each entry carries a verifiability field stating which trust technology backs it. You can inspect the full set of available models and their guarantees before you hold an account at all.

Privacy: prompts that do not leave the enclave

Providers on the network run inside Trusted Execution Environments, but not every TEE gives the same guarantee. The distinction matters, and it is selectable per request through trust modes.

In private mode, requests route only to TeeML providers. The model itself runs inside the enclave. Your prompt enters encrypted, the response is signed inside the enclave, and neither 0G nor the operator of the hardware can read the plaintext.

In verified mode, requests route to TeeML and TeeTLS providers. Execution is provable — the response can be shown to come from the model you asked for.

The Router never silently downgrades. If no provider can meet the tier you asked for, the request fails with 503 and the code no_provider_for_trust_mode. You get an error, not a quieter guarantee.

How to verify. Filter the public catalog for verifiability equal to TeeML and confirm the model you plan to use is in that set. Then send a request pinned to private and confirm it either succeeds or fails loudly — there is no third outcome where it is served at a weaker tier without telling you. Trust mode can also be attached to an API key, so every request made with that key is forced to the tier regardless of what the calling code sends.

Verifiable inference

Each response carries an x_0g_trace block: the request ID, the on-chain address of the provider that served the request, and the exact billed cost of that call.

Ask for TEE verification and the Router synchronously checks the provider’s signature, reporting the outcome as tee_verified. A value of true means the signature verified; false means a signature was present and failed verification, and that response should be treated as untrusted.

How to verify. You do not have to take the Router’s word for it. Each response can be re-verified independently: read the provider’s service record from chain, fetch the signature for that response from the provider’s public signature endpoint, check the EIP-191 signature yourself, and compare the signed text against the content you received. The Router’s flag is a convenience; the underlying proof stands on its own. See Verification and TEE for the procedure.

Zero data retention

Prompt and completion content for text and audio inference exists only in memory for the lifetime of the request. There is no conversation table and no log archive. 0G does not train on your data.

Bounded exceptions exist where the format requires temporary storage:

Content	Retention
Text and audio inference content	None beyond the request lifetime
Uploaded multipart files	Deleted within 60 minutes
Image generation inputs and outputs	Deleted within 30 minutes

What is retained is billing metadata only: request ID, wallet address, model and provider, token counts, the trust tier that served the request, cost, and timestamp. None of it contains request content.

How to verify. The retained metadata is exactly enough to audit yourself. Usage data can be broken down by trust tier, so you can confirm how much of your traffic actually ran in the sealed tier rather than assuming it. Account and usage endpoints require a management key with the account:read scope; an inference key is refused there.

What this is not

0G Private Computer is an inference surface, not a full control plane. Permissioning for your own agents, audit logs of what your systems asked and received, and reversible workflows remain yours to build. This layer gives you the primitives — isolation, proof, and a readable cost record — and leaves the architecture above it to you.

The Router also does not attempt to restate how the underlying network works. For the protocol and network layer, see the 0G network documentation.

Next steps

• Quickstart — create a key and send your first request.
• Trust modes — the full comparison of standard, verified, and private.