P2P sharing: share your GPU, get any AI model

Short answer: open obscurify.ai/sharing, pick a model your graphics card can hold and press Start sharing. The tab runs an open model on your own GPU and answers other people’s questions with it. You earn 80% of what they pay, as credit you can spend on any model we carry, frontier ones included. There’s nothing to install and no email, wallet or token: the account is a random 16-digit number.

Why we built it

Private AI shouldn’t depend on one company’s data centre, including ours. Open models already run on the graphics cards people own, and research groups have shown you can train a large model across strangers’ machines: Covenant AI (formerly Templar) pre-trained a 72-billion-parameter model with more than 70 contributors who needed nobody’s permission to join.

Serving models that way, peer to peer, is the step that hasn’t landed yet. When it does, it needs a place where your compute and your identity stay apart. P2P sharing is that place, starting with models that fit in one browser tab.

How it works

  1. You share. A browser tab with WebGPU loads an open model (Qwen3.5 2B, 4B or 9B) onto your GPU. The weights download once, 1 to 6 GB, and stay in the browser’s cache.
  2. Someone asks. Anyone can pick a P2P model in the chat, or call p2p/qwen3.5-9b and the others through the OpenAI-compatible API. Obscurify sends the question to an idle tab.
  3. Your tab answers. It writes the reply on your GPU and streams it back through Obscurify, so neither side ever connects to the other.
  4. You earn. The person asking pays the listed price, and 80% of it is credited to your account. After 24 hours it’s added to your balance, which you can spend on any model, in the chat or through the API. Credit can’t be withdrawn as money.

Prices are per million tokens, input and output: 2B $0.02 and $0.06, 4B $0.03 and $0.10, 9B $0.05 and $0.18.

What each side can see

  • The people you serve never learn who you are. Every word is relayed by Obscurify, so they don’t get your network address either.
  • You never learn who they are. Your tab does process the text of their questions; that’s what running a model means. That’s why P2P models sit in their own clearly named category: nobody’s chat goes to a stranger’s browser unless they picked a P2P model.
  • Obscurify records which account served each job, the model, a token count and the credit, so it can pay you. Nothing about who asked.

No network has yet made prompts private from the GPU that answers them, at least not on consumer hardware. What we can do is keep the prompt apart from the person, the same line we draw with the large model providers.

Keeping it honest

A tab could save electricity by sending made-up text instead of running the model. So idle tabs now and then get a check question with a known answer. A tab that answers wrong loses everything it earned that hasn’t cleared yet, and its account is kept out for a day. That’s why earnings wait 24 hours.

Token counts come from Obscurify’s own count of what passed through, not the tab’s report, so a tab can’t inflate the bill or its pay.

What to expect

  • Earnings are small. A tab that’s busy the whole hour earns one or two cents of credit. Sharing pays for some of your own use; it isn’t income.
  • One request at a time per tab. A browser tab serves a single conversation at once. More tabs and more people means more capacity.
  • What you need: a browser with WebGPU (Chrome or Edge on Windows, macOS and ChromeOS, Chrome on Android, recent Safari; on Linux Chrome needs WebGPU switched on), a GPU with the memory shown for each model, and an account number.

What’s next

Bigger models, served across several people’s GPUs instead of one, are where this goes. When an open network such as Covenant AI’s starts serving that way, the plan is to connect P2P sharing to it, so the GPUs here can take part and everyone can use what it serves. We follow those projects on decentralized AI inference.

Start sharing, or try a P2P model in the chat.