Bittensor Subnet Categories Explained

There's no single official taxonomy of Bittensor subnets — "category" isn't a field enforced on-chain, it's a description of what a subnet's incentive mechanism is actually built to reward. But in practice, the roughly 100+ active subnets cluster into a few recognizable groups.

Compute marketplaces

These subnets sell raw AI compute capacity, priced and allocated through the subnet's incentive mechanism rather than a traditional cloud contract. Chutes (netuid 64), the largest subnet by market cap as of this writing, positions itself as serverless compute for AI at scale; Targon (netuid 4) and lium.io (netuid 51) run similar incentivized compute marketplaces. These subnets tend to attract significant staking because their addressable demand — AI inference and training compute — is enormous and exists independent of Bittensor.

Model training and inference infrastructure

A step up from raw compute, these subnets coordinate the actual process of training or serving models across distributed, permissionless participants. iota (netuid 9) describes itself as the first permissionless pipeline-parallel training architecture — letting model training be split across independent miners rather than centralized GPU clusters. Affine (netuid 120) focuses specifically on reasoning-oriented mining.

Applied AI verticals

These subnets build toward a specific, often non-AI-native industry, using Bittensor's incentive structure to coordinate contributors:

• Score (netuid 44) — computer vision applied to camera feeds ("making every camera intelligent").
• NOVA (netuid 68) — accelerating drug discovery.
• Minos (netuid 107) — genomics infrastructure.
• Vanta (netuid 8) — decentralized liquidity and execution for trading firms.

These subnets typically have narrower, more specialized demand than general compute marketplaces, but correspondingly less direct competition within their specific niche.

Data and storage

Subnets built around collecting, structuring, or storing data rather than running models directly — for instance, subnets focused on web-scale data scraping or decentralized storage infrastructure. These tend to be judged less on model quality and more on coverage, freshness, and reliability of the underlying data or storage service.

LLM access and gateways

A smaller category of subnets that function as decentralized front doors to existing frontier models — providing drop-in API access to models like Claude, GPT, or Gemini through Bittensor's incentive layer rather than hosting original models themselves.

Why category matters when evaluating a subnet

Two subnets can have similar market caps and look similarly "successful" while facing very different competitive dynamics — a compute marketplace competing against several similarly sized rivals, versus a narrow applied vertical with little direct competition but a smaller total addressable market. How to evaluate a Bittensor subnet covers how to weigh that. And because category concentration shifts as new subnets register and old ones fade, the current market leaders are worth checking against this categorization rather than assuming it's static.

Related Articles

• Back to the complete Bittensor subnets guide
• How subnets earn their emissions in the first place
• Currently leading subnets, across categories
• How decentralized subnets compare to centralized AI infrastructure