Solana Confidential Compute is a developer beta for building AI applications that process prompts off-chain and publish hardware verification records on Solana. The design aims to make some AI interactions harder for infrastructure providers to read. It does not make users anonymous or prove that an AI response is correct.
- In beta on Solana devnet; no mainnet date announced.
- Supports Intel TDX and AMD SEV-SNP attestations. NVIDIA GPU attestation is still in development.
- Prompts are not intended for publication on-chain, but service providers can still see some request metadata.
- The Foundation says the release is unaudited and intended for development.
The Solana Foundation developed the open-source toolkit with Nito. It separates AI processing from blockchain verification: the model runs off-chain in a protected computing environment, while Solana records information applications can use to check the environment associated with an AI output.
How Solana’s verification is meant to work
The Foundation describes a confidential virtual machine that generates a signing key and keeps it inside protected hardware. The environment can use that key to sign an AI output, allowing an application to check which key produced the signature.
A hardware attestation, a report signed by the hardware, makes claims about the protected environment. In the Foundation’s described process, the report links the signing key to a recent Solana slot. A slot is a period in Solana’s block-production schedule. Linking the report to a recent slot places the attestation in a current blockchain context. The verification program checks the report, then records the key, hardware type, and expiration date on Solana.
An application can check the signature and consult the on-chain record before accepting an output. This can help verify that a response was signed by a key associated with a recorded environment. It cannot establish that the answer is true, that the model behaved safely, or that no software defect or other influence affected the result.
According to the Foundation, the beta supports attestations for Intel TDX and AMD SEV-SNP. NVIDIA GPU attestation remains in development. The Foundation says the release is unaudited and intended for developers testing applications, not a production-ready security guarantee.
What the privacy design hides, and what it does not
Solana’s confidential chat example, built with Next.js and React, is designed to have the browser verify the protected environment before encrypting a prompt to its key. If verification fails, the prompt should not be sent. That check depends on the webpage being trustworthy: users still have to rely on the code their browser receives and runs. A compromised or misleading page could undermine the intended verification and encryption process.
According to the template documentation, encrypted messages pass through the application server and Nito before reaching the protected environment. The model processes the prompt there, then encrypts the response before sending it back to the browser. The Foundation says verification records can be published without putting users’ prompts or conversations on Solana.
That does not hide all activity from every service involved. The template documentation says the application server can see a visitor’s IP address. Nito can see the account, selected model, request timing, and encrypted message size. Those details can reveal patterns even when prompt contents are encrypted.
The documentation also says users rely on the hardware manufacturer’s attestation and the software running inside the protected environment. Verification identifies a published software release, but does not show that the release is free of bugs. The materials do not provide a broader guarantee about what every operator or hardware provider can access.
Developers must also provide authentication, rate limiting, and bot protection. The example requires a funded Nito account and API key. The toolkit is best understood as a set of tools for building and checking a specific confidential-computing setup, not as anonymous AI with no intermediaries or trust assumptions.
Potential uses are still proposals
The Foundation has cited personal assistants, agents handling confidential business data, private trading strategies, and AI decisions for prediction markets or insurance claims as possible applications. It has also pointed to trading orders accepted only from verified machines and AI-assisted governance decisions.
These are proposed uses, not proof that the toolkit can make high-stakes AI decisions reliable or fair. Verifying the environment associated with an output does not fix model errors, poor inputs, or bad decisions. Anyone building these systems will still need to address those risks directly.
The Foundation has not announced a mainnet date. The available descriptions also give no timeline for NVIDIA attestation or an audit, and provide no performance benchmarks, pricing details, or metadata-retention policies. Those practical details will matter as much as the cryptographic design to developers deciding whether to use the system.
Key questions about Solana Confidential Compute
-
Does AI processing happen on Solana?
No. The Foundation says computation happens off-chain, while Solana records verification information.
-
Are prompts published on-chain?
The Foundation says prompts and conversations do not need to be published on Solana. The application server and Nito can see some request metadata.
-
Which hardware attestations does the beta support?
Intel TDX and AMD SEV-SNP. NVIDIA GPU attestation is still in development.
-
Is the toolkit available on mainnet or ready for production?
No. It is in beta on Solana devnet, and the Foundation says it is unaudited and intended for development. No mainnet date has been announced.
-
Does verification prove an AI answer is correct?
No. It can help an application check the key and software release associated with a signed output, but it does not prove the answer is accurate or the software is bug-free.
Solana Confidential Compute gives developers a way to pair off-chain AI processing with on-chain checks of the environment behind an output. The idea is worth watching, but the beta’s limits are clear: metadata remains exposed, users must trust parts of the software and hardware chain, and verification does not guarantee correctness or security.