Eschaton

Eschaton · protocol learning

Immanentize
the model.

Eschaton cuts a large language model into pipeline stages and trains it across permissionless GPUs on the open internet, so no machine ever holds the whole thing. Earn points for verified work. ESC trading fees pay the nodes; stake e/acc to run one.

Run a node →Watch the live run
live · connecting
loss
throughput
—
———
Nodes online
 
Tokens trained
 
Current loss
 
Epoch reward pool
trade fees · SOL
Epoch

The loop

Stake → compute → verify → distribute

Every part of the loop is checkable: stake and rewards live on Solana, and work is verified by spot checks before it counts.
  1. 01

    Stake e/acc

    Lock at least the minimum stake from your wallet. Stake is the Sybil filter: it makes every node accountable and slashable.

  2. 02

    Run a node

    Point a 16 GB+ GPU at the swarm with a separate hot key. The coordinator checks your stake and assigns you a pipeline stage.

  3. 03

    Co-train one model

    Your GPU serves a slice of the layers. Activations go forward, gradients come back, and nobody ever assembles the full weights.

  4. 04

    Fees follow verified compute

    Each epoch, trading fees in the reward vault are split by points for work that passed spot checks.

Thesis

Why model-parallel matters

Open models today get trained by whoever owns the biggest cluster. Protocol learning turns training into a network that anyone with a GPU can join, and lets ownership follow the work.
01

Split the model, not just the data

Data-parallel swarms need every node to fit the whole model. Pipeline stages let consumer GPUs train a model far larger than any one of them, and compressed activations make home internet links good enough.

02

Weights that never fully exist in one place

Each node only ever holds its own stage. No participant, including us, can walk away with the trained model. That's what makes contributor ownership meaningful.

03

Ownership follows contribution

Points for verified work settle on-chain every epoch. Lifetime points are the ledger of who built the model, and current-epoch points decide who shares that epoch's fees.

Right now

The swarm, stage by stage

Each square is a node. Hover for its address; click for its contribution history.

Pipeline

Fee flow

Trading fees fund the compute

No emissions and no inflation schedule. Rewards come only from the creator fees ESC trading actually produces: 90% goes to the reward vault and is split by verified work, and 10% goes to the e/acc lock, which buys e/acc on PumpSwap and locks it. The split is locked at launch, so the creator and team can't change it. e/acc is the stake, not the fee source, and ESC hasn't launched yet. See who controls what →
  1. 01

    Trades

    Swaps of ESC (a pump.fun launch, not live yet) pay a creator fee in SOL. e/acc is only the stake.

  2. 02

    Fee split

    pump.fun shares the creator fee: 90% to the reward vault, 10% to the e/acc lock. The launch transaction locks the split and revokes the sharing admin, so the creator and team can't change it.

  3. 03

    Vault and lock

    The reward vault is a program-owned SOL account: anyone can top it up, only epochs can spend it. The lock's SOL can only go to buy_and_lock, which buys e/acc on PumpSwap within on-chain price and budget limits and locks it in an account no instruction can move funds out of.

  4. 04

    Epoch settlement

    Each epoch's pool is capped at a share of the vault, split by verified points and committed as a merkle root.

  5. 05

    Claims

    After a challenge window, node owners claim with a merkle proof; points are added to the on-chain ledger. Unclaimed epochs expire.

Fee income depends on trading volume, so it can drop to zero. Staked e/acc is exposed to price risk and to slashing if a node fails verification. Nothing here is a promise of return.

Bring a GPU. Join the swarm.

16 GB of VRAM, a reachable port and a staked node key. Join or leave at any time; your points are kept on-chain.

Hardware & setupStake e/acc