What Miners Do

The Job

Every subnet defines one AI task. When a validator on that subnet sends a query, your miner receives it, does the work, and returns a response. That is the entire loop, and it runs continuously: you are operating a production AI service, not collecting a yield. Uptime is part of the job, because a miner that cannot be reached scores nothing for that epoch no matter how good its model is.

What the work involves depends on the subnet, and falls into roughly three shapes. Inference — serving a trained model, the most common and the most predictable to spec hardware for. Training contribution — supplying gradients or data to a collaborative training run, with longer cycles and a different hardware profile. Data work — scraping, cleaning, labelling or transforming, which ranges from CPU-cheap to heavier than inference.

What It Pays

Validators score your responses and write weights to the chain. Those weights become your share of the subnet’s emission for the epoch. The number that matters is therefore relative: not how good your miner is, but where it ranks among the miners on the same subnet. A top-quartile miner earns a multiple of a median one, and the bottom of the ranking earns close to nothing.

That makes expected pay a division problem before it is an engineering one. A subnet’s emission share divided by its active miners is the rough per-miner figure; a quiet subnet with modest emissions can pay better than a headline one with hundreds of optimised competitors. Registration also costs TAO, burned rather than refunded, so the first question is always which subnet — not how to tune.

Work out which subnet to mine → · Compare emission and miner counts →