Coaching a big language mannequin usually requires a warehouse stuffed with GPUs, a seven-figure cloud computing invoice, and the form of organizational muscle solely a handful of firms possess. Bittensor’s Subnet 9 is attempting to flip that script with a brand new structure referred to as $IOTA, brief for Incentivised Orchestrated Coaching Structure, which splits large AI fashions throughout a number of machines so no single participant wants to carry all the factor in reminiscence.
From winner-takes-all to collective meeting line
Earlier variations of SN9 operated on a aggressive mannequin. Miners primarily raced one another, and solely high performers earned rewards. By August 2024, that setup had efficiently pretrained massive language fashions with as much as 14 billion parameters.
However the winner-takes-all method had a ceiling. It discouraged smaller contributors who couldn’t compete with well-resourced miners, and it created pure bottlenecks round what any particular person machine may deal with. $IOTA, printed on arXiv on July 16, 2025, rethinks all the incentive construction.
As a substitute of remoted opponents, miners now operate as nodes in a collaborative pipeline. The structure integrates each pipeline parallelism and knowledge parallelism, two strategies borrowed from how main AI labs already distribute coaching workloads internally. Rewards below $IOTA are distributed proportionally amongst all pipeline miners based mostly on their precise contribution, eradicating the first disincentive for smaller GPU house owners to take part.
Coaching AI fashions out of your front room
The sensible extension of this structure confirmed up in February 2026 with the launch of “Prepare at House,” a shopper software that lets Mac customers contribute their GPU energy to the coaching pipeline. The applying works by means of an orchestrator that handles coordination throughout contributors. It distributes mannequin layers evenly and manages the reward allocation so particular person customers don’t want to know the underlying pipeline mechanics.
What this implies for buyers
Most “decentralized compute” tasks in crypto have centered on inference, operating already-trained fashions, somewhat than coaching new ones from scratch. Coaching is orders of magnitude tougher as a result of it requires tight synchronization, large knowledge throughput, and constant uptime throughout all collaborating nodes.
$IOTA’s pipeline parallelism method sidesteps the reminiscence constraints which have traditionally made distributed coaching impractical for billion-parameter fashions by splitting mannequin layers throughout machines somewhat than requiring every participant to carry an entire copy. The prior observe file of SN9 pretraining fashions as much as 14 billion parameters offers not less than a baseline proof that the subnet can deal with significant workloads.
For $TAO holders particularly, the shift from winner-takes-all to proportional rewards may meaningfully change mining economics on Subnet 9. Broader participation means extra distributed demand for $TAO staking, nevertheless it additionally means particular person reward charges will compress as extra miners be part of the pipeline.
A malicious or malfunctioning node in a coaching pipeline can corrupt gradient updates for all the run. How $IOTA handles Byzantine fault tolerance in follow will decide whether or not this structure scales past proof-of-concept into production-grade coaching infrastructure.
