SN9 enables large-scale AI model training using IOTA architecture

Training a large language model typically requires a warehouse full of GPUs, a seven-figure cloud computing bill, and the kind of organizational muscle only a handful of companies possess. Bittensor’s Subnet 9 is trying to flip that script with a new architecture called IOTA, short for Incentivised Orchestrated Training Architecture, which splits massive AI models across multiple machines so no single participant needs to hold the entire thing in memory.

From winner-takes-all to collective assembly line

Previous versions of SN9 operated on a competitive model. Miners essentially raced each other, and only top performers earned rewards. By August 2024, that setup had successfully pretrained large language models with up to 14 billion parameters.

But the winner-takes-all approach had a ceiling. It discouraged smaller contributors who couldn’t compete with well-resourced miners, and it created natural bottlenecks around what any individual machine could handle. IOTA, published on arXiv on July 16, 2025, rethinks the entire incentive structure.

Instead of isolated competitors, miners now function as nodes in a collaborative pipeline. The architecture integrates both pipeline parallelism and data parallelism, two techniques borrowed from how major AI labs already distribute training workloads internally. Rewards under IOTA are distributed proportionally among all pipeline miners based on their actual contribution, removing the primary disincentive for smaller GPU owners to participate.

Training AI models from your living room

The practical extension of this architecture showed up in February 2026 with the launch of “Train at Home,” a consumer application that lets Mac users contribute their GPU power to the training pipeline. The application works through an orchestrator that handles coordination across contributors. It distributes model layers evenly and manages the reward allocation so individual users don’t need to understand the underlying pipeline mechanics.

What this means for investors

Most “decentralized compute” projects in crypto have focused on inference, running already-trained models, rather than training new ones from scratch. Training is orders of magnitude harder because it requires tight synchronization, massive data throughput, and consistent uptime across all participating nodes.

IOTA’s pipeline parallelism approach sidesteps the memory constraints that have historically made distributed training impractical for billion-parameter models by splitting model layers across machines rather than requiring each participant to hold a complete copy. The prior track record of SN9 pretraining models up to 14 billion parameters provides at least a baseline proof that the subnet can handle meaningful workloads.

For TAO holders specifically, the shift from winner-takes-all to proportional rewards could meaningfully change mining economics on Subnet 9. Broader participation means more distributed demand for TAO staking, but it also means individual reward rates will compress as more miners join the pipeline.

A malicious or malfunctioning node in a training pipeline can corrupt gradient updates for the entire run. How IOTA handles Byzantine fault tolerance in practice will determine whether this architecture scales beyond proof-of-concept into production-grade training infrastructure.

Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our Editorial Policy.