Subnets
Subnets are the building blocks of Bittensor. Each subnet is an independent competitive market focused on a specific task, from text generation to protein folding. Together, they form a decentralized network of specialized AI services.
What is a Subnet?
A subnet is an independent network within Bittensor where miners compete to perform a specific task and validators evaluate the quality of their work. Each subnet has its own incentive mechanism, custom code written by the subnet owner that defines what "good work" means. On a text generation subnet, that might mean producing coherent, accurate responses. On an image generation subnet, it might mean creating high-quality images that match a prompt.
Every subnet is identified by a netuid, a unique integer. Netuid 0 is special: it's the root network, which doesn't perform AI tasks but instead coordinates emission distribution across all other subnets. Validators on the root network set weights on subnets (not individual miners), determining how much TAO each subnet receives.
Subnets operate largely independently. Each has its own set of registered miners and validators, its own tempo (how often consensus runs), and its own scoring criteria. Subtensor records the results on-chain (weights, registrations, stake) but the actual computation and evaluation happen off-chain between miners and validators.
Subnet Lifecycle
Creation. Anyone can create a subnet by calling register_network and locking a specified amount of TAO. The lock cost adjusts dynamically: it increases when subnets
are being created frequently and decreases over time when no new subnets appear. This mechanism
prevents spam while keeping the barrier responsive to demand. Once the TAO is locked and the transaction
is confirmed, the new subnet receives a netuid and the creator becomes its owner.
Operation. A running subnet is a continuous cycle. Miners register and begin serving requests. Validators send tasks, evaluate responses, and submit weight vectors to the chain. Every tempo blocks (the subnet's epoch length), Subtensor runs Yuma Consensus on those weights and distributes emissions. The subnet owner can adjust parameters like the maximum number of UIDs, the tempo, and registration difficulty.
Deregistration. Subnets can be removed from the network. If a subnet consistently fails to attract stake and participation, or if the network reaches its subnet capacity, the lowest-performing subnets face deregistration to make room for new ones. When a subnet is deregistered, its locked TAO is released and all registered UIDs within it are cleared.
Scenario: The Life of Subnet 42
Step 1 of 6Creation: Locking TAO to launch SN42
A team builds an AI model for real-time translation. They lock 500 TAO (the current dynamic cost) by calling register_network. Subtensor assigns netuid 42, and the team becomes the subnet owner with control over parameters like tempo and max UIDs.
The Ecosystem Today
The Bittensor ecosystem includes dozens of active subnets, each tackling a different problem domain. Some focus on language tasks like text generation, translation, and summarization. Others target vision, audio, time-series prediction, data scraping, or mathematical reasoning. The diversity is intentional: Bittensor aims to be a general marketplace for intelligence, not just one type of AI.
Each subnet develops its own culture and competitive dynamics. Some attract large-scale GPU operators running cutting-edge models. Others are accessible to smaller participants with more modest hardware. What they share is the common infrastructure: Subtensor handles registration, stake accounting, weight recording, and emission distribution, while the subnets handle the actual AI work.
The subnet landscape evolves constantly. New subnets launch to explore novel tasks, existing subnets refine their incentive mechanisms, and underperforming ones get replaced. This Darwinian process (subnets competing for stake and emissions, miners competing within subnets for rewards) is the engine that drives Bittensor's quality of intelligence upward over time.