The democratization of AI, powered by the computers people already own.
P2P AI joins home machines into one network that runs large language models. Your prompt goes straight to a volunteer's computer over WebRTC and the reply streams back token by token. Borrow the pool's GPUs, or lend your own.
Connecting to the pool…
Some models are too big for one home computer, so they run split across several: each donor holds a slice of the layers. Each bar below is one model, layer by layer; a hatched gap is a part nobody serves yet.
Speeds are estimates from what donors report about their own computers. The hub does not measure them, and nothing here is a price.
The hub is busy, so this is its last known state.
We want to democratize access to large language models.
Today, access to LLMs runs through a handful of data centers, at a handful of companies' prices and rules. Meanwhile, millions of capable computers sit idle in homes every night.
Shard Superintelligence Lab is building the other way. People join their home machines into one peer-to-peer network, and that network becomes AI infrastructure owned by the people who run it.
Open a browser and ask. New accounts get starter credits, and a daily allowance is shared among the people who contribute.
A gaming PC or a laptop with a GPU can serve open models and earn credits for every token it generates.
Open models stay alive for as long as people choose to run them, not only for as long as one company keeps a server on.
An honest status: this is a beta. The pool serves small open models today, and the hub still keeps the ledger.
This browser makes its own Ed25519 key and keeps it here. The hub grants starter credits so you can chat right away; serving others earns more.
Answered by a volunteer's GPU, paid per token by receipt.
The volunteer who serves your prompt can read it. Don't paste secrets, passwords or private data.
Ask a question to start.
A donor's GPU answers over WebRTC, and every reply ends with its receipt: tokens in and out, what you paid, and the ticket.
Payments open soon. These are the prices; buying is not switched on yet.
Pro accounts are served first under congestion; donors who share compute can rank higher still. When donors are idle, everyone is served at once.
Four steps, each one signed. The hub never carries your tokens.
It escrows the worst-case cost of the reply and signs a ticket bound to your exact prompt, so a donor can't reuse it for anything else.
Your browser dials the donor over a WebRTC data channel. Tokens come straight from their GPU, and you sign a running receipt every 64 tokens.
You pay for the tokens you actually received. The donor is paid minus a small fee that funds the shared pool, and the rest of the hold comes back to you.
Busy donors serve higher fair-share weights first, from 0.25× to 8×. Serving the pool raises yours, and everyone is served eventually.
Models too big for one computer can be split. Each donor runs a slice of the model's layers, and your browser hands the activations from one to the next, token by token, with the same math as on a single machine.
The desktop app for Windows, and the donor for any computer with Python.
Chat, share your GPU and manage your account from a desktop app. Unzip the download and start the app inside.
The app is new and not yet code-signed, so Windows SmartScreen may warn that it is unrecognized. Check that the file's SHA-256 matches the one above (in PowerShell: Get-FileHash <file>), then choose “More info” and “Run anyway”.
Serves the pool in the background from any computer with Python. Install it with pipx:
Run the donor in the background on a PC with a GPU. It earns credits for every token it serves and keeps serving after this tab is closed.
pipx inject shard-sys1 torch "transformers>=4.56" huggingface_hub
newton join <handle>
Pick a new handle, or add this computer to an account you already have: make a link code in your account card, then run newton join --link <code> instead.
newton share --public
Already running Ollama, LM Studio or llama.cpp? The donor serves through it. To see what it found without serving, run newton share --dry-run.
When you serve, the person you serve connects to your computer directly over WebRTC and can see its public IP address. Their prompts reach your computer in plain text.
Shard Superintelligence Lab is a research lab working on peer-to-peer AI.
Researcher
Felipe Ranieri leads the lab's research. His work here spans peer-to-peer inference over WebRTC, fair sharing of donated compute, signed settlement between strangers, and AI agents that run on everyday machines.
He believes access to capable AI should not depend on owning a data center, and that the computers people already have can carry the load together.
Felipe on LinkedInShard is a small research lab. The pool runs on donated computers and one modest hub server. Donations pay for that server and for cloud GPUs so the lab can train open models that anyone can run on the pool.
Donations are gifts to the lab. Apart from an optional Supporter badge and a place on the supporters wall, they buy nothing: no credits and no priority. Credits are never redeemable for money.
Bitcoin (on-chain, Taproot)
Bitcoin address
bc1pychzp39lwakzez2t7lq6y99zwusdx227zgh5gdpyh6g807g0t7uq8yf7fr
Thank you. Your gift helps keep peer-to-peer AI running and free to try.
Raised so far
Be the first supporter.
No one has chosen to be listed yet.
Pick a handle first, above, to get a recognised supporter tag.
Get a personal tag, then send a donation ending in that many satoshis so the hub can match it to your account.
Recognition appears after one confirmation (about 10-60 minutes).
And what it doesn't. The hub publishes these terms itself, in English, live at
/api/stats.