How it works
ShareInference is an open-source inference network: people lend GPU time from a browser tab and earn credits, and credits pay for an OpenAI-compatible API.
- 1
Providers share a GPU
A contributor opens the Share page in Chrome or Edge. The browser downloads an open model once (about 2.4 GB, then cached) and runs it on the GPU through WebGPU, inside a Web Worker. The worker keeps a WebSocket open to the ShareInference server and waits for work.
- 2
Consumers call an API
A consumer creates an API key and sends OpenAI-style requests to /v1/chat/completions. Streaming and non-streaming both work, so existing OpenAI clients only need a new base URL.
- 3
The server routes, it does not compute
The server is CPU-only. It authenticates the request, queues it, picks one random idle provider that has the requested model loaded, streams the tokens back and settles credits. Consumers choose a model, never a provider.
- 4
Credits keep it fair
Every request costs credits (1 per input token, 5 per output token) and the provider who served it earns the same amount, so credits are only moved, never created or burned. Holders of credits are served first (priority lane); everyone else waits in a fair queue, which needs a GitHub account of at least 30 days.
Privacy, honestly
- Prompts are processed inside a contributor's browser. A technically skilled contributor could read them with the browser's developer tools. The interface never shows them, but the text does pass through that machine. Do not send secrets.
- Providers never see who you are: no account, API key or IP address is ever sent to them. Routing is blind, and the provider's public name is revealed to you only after the request has finished.
- ShareInference does not store or log prompt or answer text, only ids, token counts and timings.
- Contributors see their own usage live, such as tokens, speed and credits, never the content or the asker.
More detail on what is visible to whom.