Your Simple Guide to IOTA (SN9)
Training large AI models costs hundreds of millions of dollars. Only a handful of corporations can afford the GPU clusters, power, and engineering needed to build a single frontier language model. As a result, these companies control the development of the most capable AI systems. IOTA changes this dynamic by introducing decentralized swarm training on Bittensor. Specifically, it lets globally distributed participants train large language models together without any single entity owning all the infrastructure.
What is IOTA?
IOTA stands for Incentivised Orchestrated Training Architecture. It runs as Subnet 9 (SN9) within the Bittensor ecosystem. Moreover, it represents a fundamental redesign of decentralized AI training. Instead of miners competing individually in a winner-takes-all format, IOTA turns all participants into a single cooperating unit. In other words, everyone contributes compute toward the same model and earns rewards based on their contribution.
The architecture combines data parallelism with pipeline parallelism. In practice, IOTA (SN9) splits a large language model into layers. Then, different miners process different sections. Because of this, no single miner needs to fit the entire model on their hardware. For example, a miner with a consumer-grade GPU can handle one layer, while a data center operator can take several. Consequently, the hardware barrier drops significantly compared to centralized training.
IOTA launched on the Bittensor mainnet in June 2025. It quickly attracted attention from both the crypto and AI research communities. Notably, Forbes featured IOTA in 2025 in an article about swarm intelligence reshaping AI training. The team also published a technical primer on arXiv, which established academic credibility. Overall, the project marks a major step toward Bittensor’s original vision of building competitive, decentralized alternatives to proprietary AI models.
How IOTA works?
IOTA (SN9) uses three core roles to coordinate training. First, the Orchestrator sits at the center of a hub-and-spoke architecture. It assigns miners to specific model layers, triggers training and merging cycles, and maintains global oversight. Second, miners supply GPU compute, memory, and bandwidth. Third, validators monitor miners and verify that all submitted work is honest.
Each miner downloads the current weights for their assigned layer and then processes activations. Forward activations push data samples through the model to produce losses. Backward activations push gradients in the opposite direction to enable learning. The system alternates between training and merging stages. During merging, miners sync their local weights with peers on the same layer through a technique called Butterfly All-Reduce. As a result, many contributions merge into one global model without any central server.
Bandwidth presents one of the biggest challenges in distributed training. IOTA addresses this with a novel bottleneck transformer block. In preliminary experiments, this block achieved up to 128x activation compression on a 1.5B parameter model. Therefore, the data miners send between pipeline stages shrinks dramatically. This makes large-scale coordination more feasible, even over standard internet connections. The team expects this compression approach to scale well to larger models, though further validation on longer training horizons is ongoing.
To distribute rewards fairly, IOTA uses CLASP (Contribution Loss Assessment via Sampling of Pathways). CLASP draws on Shapley values from cooperative game theory. Instead of rewarding miners on simple metrics, it measures each participant’s marginal contribution to model improvement. Consequently, it detects free riders and bad actors. Meanwhile, honest contributors receive proportional rewards. Validators confirm work quality by rerunning portions of each miner’s training and checking results with cosine similarity.
Who is behind it?
Macrocosmos AI built and operates IOTA. Will Squires (CEO) and Steffen Cruz (CTO) co-founded the company. Both have deep roots in the Bittensor ecosystem. Specifically, Steffen previously served as CTO of the OpenTensor Foundation and was a core developer of the Subnet 1 (Apex) codebase. He holds a PhD in Subatomic Physics from the University of British Columbia. Will brings experience from major infrastructure projects and AI product strategy. He holds a Master of Engineering from the University of Warwick.
Macrocosmos now employs around 24 people. The team includes machine learning engineers, data scientists, and researchers. Additionally, they run multiple Bittensor subnets. These include Subnet 1 (Apex) for agentic AI inference, Subnet 13 (Data Universe) for web-scale data collection, and Subnet 25 (Mainframe) for scientific compute. Together, these subnets form a full-stack AI pipeline. Within this pipeline, IOTA (SN9) provides the core training layer.
All IOTA components are open-sourced on GitHub under the macrocosm-os organization. The technical primer lists Felix Quinque, Alan Aboudib, Szymon Fonau, Rodrigo Lopez Portillo Alcocer, Brian McCrindle, and Steffen Cruz as authors. Furthermore, the project maintains active communities on Discord, X (Twitter), and Substack. There, the Macrocosmos team publishes regular updates and technical deep dives.
Why it is valuable?
IOTA (SN9) solves a core problem in decentralized AI. Previously, Subnet 9 ran competitions where miners trained their own models independently. Only the best model received a reward. As a result, most compute went to waste. Barriers to entry stayed high, and the actual training process remained centralized. IOTA changes this entirely. Now, every participant contributes to the same model. Every valuable contribution stays in the system. And the network distributes rewards proportionally.
Because of pipeline parallelism, IOTA grows more efficient at scale. Unlike data-parallel methods, it does not require every node to hold a full copy of the model. Instead, IOTA spreads the model across participants. According to mining documentation, GPUs with as little as 16 GB of VRAM can technically participate, though miners with higher-end hardware (80 GB VRAM and above) will process updates faster and earn significantly more rewards. This means both consumer-grade GPUs and enterprise-level clusters can work on the same training task, though with different levels of output.
IOTA also tackles the rising cost of centralized AI training. When models double in capability, training costs often jump tenfold. Decentralized swarm training offers a way around these spiraling expenses. It taps into distributed idle compute worldwide. In addition, Bittensor’s token incentive mechanism organizes this compute at massive scale.
The Train at Home application lowers the barrier even further. Users simply download a desktop app and connect a wallet. Then, anyone with a compatible device can join the swarm and earn rewards automatically. Currently, the platform supports macOS (Apple Silicon with 8 GB memory) and Linux (16 GB VRAM and RAM). On top of that, a live dashboard lets participants track active miners, monitor training progress, and view validator metrics in real time.
The future of IOTA
The Macrocosmos team is pushing IOTA (SN9) toward commercial use. A key goal is to monetize training FLOPS directly. For example, the network could train models for external clients. Alternatively, it could lease the swarm for custom training runs. To support enterprise adoption, the team is also developing privacy-preserving model weight technology.
Integration with other Macrocosmos subnets opens up major opportunities. Data from Subnet 13 (Data Universe) can flow directly into IOTA training pipelines. Meanwhile, finished models can go live through Subnet 1 (Apex) for inference. This creates a full decentralized AI stack: data collection, training, and deployment. Bittensor’s incentive mechanism coordinates it all.
Additionally, the team is researching broader hardware support. The goal is to make IOTA compatible with more consumer-grade GPUs. Combined with the Train at Home app and ongoing compression improvements, this could let anyone with a modern computer help train frontier AI models.
The long-term vision is clear. IOTA wants to become the first distributed compute layer for model training that matches centralized performance. By combining Bittensor’s node coordination with novel compression and orchestration, IOTA aims to give organizations a real alternative. Ultimately, if it succeeds, it could change who gets to build frontier AI.
Sources:
https://docs.macrocosmos.ai/subnets/subnet-9-iota
https://github.com/macrocosm-os/iota
https://iota.macrocosmos.ai
