One living model, trained on the world’s spare GPUs, with open weights on a blockchain — and it pays the people who train it.
Today’s best AI lives inside a handful of companies, on clusters worth hundreds of millions. Sestrian flips it: the model is the blockchain’s state, and it’s trained by a crowd of ordinary machines instead of one data center. Replay the chain from its first block and you rebuild the exact same weights, bit for bit. Contribute compute, get rewarded — just like mining, except the thing being mined is intelligence.
The model is the chain’s state. Anyone can read it, verify it, and rebuild it exactly — no gatekeeper, no API you can be cut off from.
Point your graphics card at the network and it trains the live model. A gaming PC, a workstation, even an Apple laptop — the client measures your speed and sizes each round to fit.
Every block mints a reward split between the miners who trained it, the block’s proposer, and the owners of the data it learned from.
Every decentralized-AI network so far pays its contributors in token inflation and hopes revenue shows up later. None of them pays training rewards out of inference income. That crown is unclaimed — and claiming it doesn’t require the world’s best model, only a model whose users pay more than its training costs.
Transformers solved cooperative machine cognition. Bitcoin solved adversarial bookkeeping. Neither solved a self-funding, openly-owned model that improves because improving it pays.
Better model → more paid inference → more revenue into training → better model. Issuance bootstraps it; fees are meant to carry it.
Others keep the incentive layer on-chain and the actual weights in someone’s S3. Here the weights are the chain state — there is no bucket to trust, revoke, or lose.
The one number that matters is the crossover: the first block where inference revenue exceeds issuance. Everything else is scaffolding for that.
Run the installer. It builds the node, makes your wallet, and reproduces the genesis model locally — deterministic, so nothing has to be trusted.
Your machine improves the current model on real data, entirely on your own hardware, in a round sized to what your GPU can finish in time.
It packages the improvement into a signed, compressed update — naming the data it learned from — and gossips it to the network.
The chain scores every update on held-out data, folds them into one new model, and pays out in proportion to what each actually improved.
A model is only as good as what it learns from, so data here isn’t free fuel — it’s owned, and it earns. You stake coins behind a corpus to put it in (skin in the game against a challenge market), and from then on it pays you: a share of every block trained on it, and a cut of the fees whenever the model actually leans on it to answer.
Every gradient names the corpus it trained on, and the network rejects any that doesn’t. Credit is recorded on-chain — it reaches the owner, not whoever scraped it.
Each answer is matched against what actually shaped the weights behind it, and the fee is split accordingly. Data that turns out to matter years later starts earning then — usage is the clock, not age.
The math behind every payout is recomputable from public data. A false claim is challengeable, and the stake behind it is slashable.
Floating-point math differs between an Apple GPU and an NVIDIA one — the same training step gives subtly different numbers. So the consensus boundary sits below the float: every machine quantizes its update to integers before committing it, and all the chain arithmetic is exact integer math. We run it live — a MacBook on Apple silicon and a 2080 Ti in another building, no coordinator between them, agreeing on the model down to the last bit.
Frontier AI shouldn’t be a private vault.
It should be public infrastructure — like the internet itself.
The money to train it comes from the network, not a boardroom. The data that goes in is credited and paid. The weights that come out belong to everyone. Bitcoin proved a currency could run with no central bank. Sestrian is the same wager, aimed at intelligence.
The devnet is live and training right now. If you have a spare graphics card and want in on public AI from the ground floor, this is where it starts.