iota
SN9Train massive AI models by connecting thousands of computers in parallel
Macrocosmos rebuilt subnet 9 as a pipeline parallel training network. The pitch: pretrain frontier-scale models on a swarm of unreliable, internet-connected GPUs instead of a billion-dollar datacenter.
// Pretraining as a swarm, not a fortress.
iota is a training subnet that splits a large language model into pieces and trains it across many independent machines on the open internet. The IOTA acronym stands for Incentivized Orchestrated Training Architecture. The team has been running subnet 9 since 2024, originally as a pure pretraining benchmark, and relaunched the subnet around the IOTA architecture in 2025.
The simple version: Imagine training a single huge AI model by stringing together a thousand consumer GPUs in different countries, with each machine handling one slice of the model and passing intermediate results along the chain.
Centralized equivalent: OpenAI or Anthropic running pretraining inside one operator's datacenter. iota is the decentralized version: same training objective, distributed across permissionless participants.
How it works:
- Miners are assigned a slice of the model by an orchestrator, process activations passed to them by upstream miners, periodically upload their local weights to a shared S3 bucket, and merge with peers using a Butterfly All-Reduce variant
- Validators spot-check miner work, watch for anomalies, score activation quality, and can trigger rollbacks to a prior stable epoch if the run is corrupted
- The problem it solves: Frontier pretraining is locked behind capex. Only operators that can build their own datacenters get to set the agenda for what large models look like.
- The opportunity: If a swarm of consumer GPUs can pretrain a real model, the addressable compute pool is orders of magnitude larger than what any single company controls.
- The Bittensor advantage: Token incentives let strangers contribute compute without trusting each other or a central operator. The orchestrator coordinates, the validators police, the chain settles.
- Traction signals: The team shipped a "Train at Home" macOS app that lets non-technical users contribute compute. The repo shows 265 commits, 9 contributors, with the most recent main branch merge dated April 16, 2026 and branch activity within the past week.
Category: Distributed Training | Centralized Competitor: OpenAI, Anthropic, Meta AI in-house training
Subnet 9 is one of the longest-running subnets on Bittensor. According to the IOTA paper on arXiv, SN9 was the first to demonstrate that an incentivized permissionless network could each pretrain foundation models. The relaunch around the IOTA framework is the team's attempt to push beyond benchmarking and run a single coordinated training job across the swarm.
Mechanism:
According to the iota repo README, the orchestrator distributes model layers across heterogeneous miners and streams activations between them. All network communication is routed through the orchestrator, and a shared S3 bucket stores activations and layer weights. Miners compete to process as many activations as possible during the training stage, periodically upload their local weights, and merge activations using a variant of Butterfly All-Reduce. Validators spot-check that work was actually performed and, per the team's TAO.app brief, can rollback the run to a prior stable epoch if adversarial behavior corrupts training.
The current run is a 1.5B parameter Llama-inspired architecture split across three layers (one head, one body, one tail), per the repo. The stated roadmap is to scale toward 15B, 50B, and eventually 100B parameter models. Compute requirements are deliberately modest: the README recommends a CUDA GPU with 16GB VRAM (an RTX 4090, for example) and Ubuntu 22.04.
The economic picture is unusual right now. The pool holds 47,855 TAO with a market cap near 116,000 TAO and the alpha token trades at 0.02262 TAO. Volume is light at about 1,571 TAO over the last 24 hours. The subnet's emission share is currently 0%, and Taoflow drives that share off net staking flows smoothed with a roughly 30-day half-life, so even a positive 7-day inflow (around 377 TAO net) does not immediately translate into network emissions. On top of that, the subnet is configured to burn roughly 94% of miner emissions according to TaoSwap. Together those settings mean the subnet is currently funding very little miner reward through the protocol, which is consistent with a relaunch and onboarding phase rather than steady-state operation.
The lead author on the architecture, CTO Steffen Cruz, is named publicly on the TAO.app about page. The repo's commit history shows a small, consistent cluster of contributors, with one engineer responsible for the majority of merges. Macrocosmos also operates other subnets in the Bittensor ecosystem, which means iota benefits from organizational scale even though the IOTA repo itself is focused.
- Emission share at zero: Subnet 9 currently receives 0% of network emissions under Taoflow. Restarting emissions requires sustained net staking inflows long enough to shift the smoothed average back into positive territory. Until that happens, miner rewards from protocol emissions are minimal regardless of the subnet's internal split.
- Execution risk on the scaling roadmap: The current run is at 1.5B parameters. The path to 15B, 50B, and 100B per the README is technically ambitious for any decentralized training setup, and progress on that ladder is the main signal of whether the architecture works at scale.
- Liquidity: With about 1,571 TAO in 24-hour volume against a 116,000 TAO market cap and a 47,855 TAO pool, position sizing matters. Slippage in and out is material at size.
- Competition: Decentralized training is a contested category on Bittensor and beyond. iota is competing both with other Bittensor training subnets and with centralized labs that operate at vastly larger compute scales.
Into the next one.
