QVAC Marketplace: A Decentralized P2P AI Inference Market Built on Solana

QVAC Marketplace is a decentralized peer-to-peer AI inference platform built on Solana. Users can run open-source large language models on strangers' machines using SOL tokens, with end-to-end encryption throughout and no need for API keys. This platform aims to address issues with centralized AI inference services such as high costs, privacy data leaks, reliance on a single service provider, and cumbersome identity verification. It also allows users with GPU resources to become inference service providers and earn SOL rewards.

Most developers and users currently rely on centralized cloud services like OpenAI and Anthropic to run large language models, facing issues such as high API costs, privacy data being held by platforms, service availability dependent on a single provider, and cumbersome identity verification processes. QVAC Marketplace proposes a decentralized P2P inference market solution: anyone with a GPU can become a service provider, and users pay directly with SOL to run open-source large language models in an end-to-end encrypted environment.

QVAC follows the principle of "minimal trust" and its core architecture consists of four components: 1. Solana Smart Contract Layer: As a trust anchor, it handles provider registration, task hosting, and fund settlement. The contract address is publicly auditable (Program ID: 6rbgdrQdxziVC25kt1Xmtz36ApiLdUVGpdyDcssmgoec); 2. QVAC Web Server: Hosts the web interface (qvacmarketplace.io) to process metadata without touching inference content; 3. QVAC Bridge: A local WebSocket bridge responsible for building transactions, communicating with Phantom wallets, and establishing P2P connections. Private keys are always stored in the wallet; 4. QVAC Provider: GPU machines can run inference nodes, supporting open-source models like Qwen and Llama. Providers set their own prices and earn SOL rewards.

QVAC's privacy protection design includes: 1. P2P Encrypted Communication: Inference traffic is transmitted via the Holepunch HyperDHT protocol, using the Noise protocol for end-to-end encryption. The web server and Solana only see public keys and 32-byte hashes, with no access to actual conversation content; 2. No-Logging Commitment: Neither the market nor Solana views inference content. Prompts and model responses are encrypted throughout; 3. Local Bridge Security: Listens only on 127.0.0.1:3000 and uses a source whitelist to restrict external access.

Payments use an escrow model to protect both parties' interests: 1. Fund Locking: When a user initiates a request, the quoted amount is transferred to a Solana Job account and locked; 2. Inference Execution: After an encrypted connection is established, model tokens are streamed, and the provider submits the SHA-256 hash of the response to the chain; 3. Confirmation and Release: Funds are transferred to the provider after user confirmation. If no delivery occurs within the timeout (600 seconds), users can call refund_job to retrieve funds. Provider Ecosystem: No KYC verification, no platform fees, self-set pricing, support for multiple open-source models, and registration via cloning the code repository. Tech Stack: Smart contracts use Rust + Anchor 0.30.1, client SDK is TypeScript, bridge requires Node.js ≥22.17, P2P network uses Holepunch HyperDHT, and includes 39 integration tests.

MVP Version Limitations: Inference trust issues (malicious providers may not deliver real content; users need to rely on on-chain jobs_completed counts to avoid choosing them again). V2 Improvement Directions: Introduce a formal dispute resolution path, possibly using zero-knowledge proofs or TEE (Trusted Execution Environment) to enhance inference verifiability.

QVAC combines blockchain escrow mechanisms with P2P privacy protection, representing an important exploration direction for decentralized AI services. It provides end-to-end encryption for privacy-focused users; lower-cost options for developers; and a way to monetize idle computing power for GPU owners. Despite its limitations, its technical architecture and design philosophy offer references for the industry, and it may become a supplement or alternative to centralized cloud services in the future.