Bittensor Subnet Licensing Analysis

Shared by @futuretrees

Updated Jul 8, 2026

Bittensor Subnet Licensing Decisions — Deep Research

Researched 2026-07-06. Data sourced from bittensor.ai subnet directory, taostats.io, GitHub LICENSE files (raw), and Bittensor documentation.

Executive Summary

The Bittensor ecosystem has 129 active subnets (as of July 2026), each operating as an independent "open-source startup" producing an AI-focused commodity. The overwhelming licensing pattern across the network is MIT License — the most permissive mainstream open-source license. A minority of subnets have no visible LICENSE file or use alternative approaches. No subnet in the top tier was found using copyleft licenses (GPL/AGPL). This has significant implications for anyone (like Ditto) considering a more restrictive licensing approach.

Key finding: The Bittensor culture strongly favors permissive open-source. Subnet codebases are treated as public infrastructure that anyone can read, fork, and build on — the moat comes from the incentive mechanism and on-chain emissions, not from code licensing.

How Bittensor Subnets Work (Licensing Context)

From the official Bittensor docs:

A subnet's incentive mechanism defines the work miners must perform, and the work validators must perform to evaluate the miners' work. The incentive mechanism is unique to the subnet, and maintained off-chain by the subnet creator in the form of a code-repository that defines the interface for miners and validators to participate.

This means:

  • The code repo IS the subnet — it defines how miners and validators interact
  • The repo is maintained off-chain by the subnet owner
  • The license of that repo determines who can participate, fork, or commercialize the subnet's code
  • The on-chain component (Yuma Consensus, emissions, alpha token liquidity pool) is protocol-level and separate from the repo license

Each subnet also has:

  • An owner cut (typically 18%) of emissions
  • A alpha token (dTAO) with its own AMM liquidity pool
  • Open or permissioned registration for miners/validators

Subnet-by-Subnet License Findings

Tier 1: Top Subnets by Emissions & Market Cap

SN Name GitHub Repo License Copyright Holder Notes
SN64 Chutes chutesai/chutes MIT Jon Durbin (2024) #1 emission (5.11%). Serverless GPU compute. Miner code split to rayonlabs/chutes-miner.
SN120 Affine AffineFoundation/affine Not found (404) — #2 emission (4.77%). Reasoning model competition. No LICENSE file at standard paths.
SN107 Minos (GitHub linked on-chain) Not verified — #3 emission (4.6%). Genomic variant calling. Partnered with Hippius for storage.
SN4 Targon manifold-inc/targon MIT Yuma Rao (2021), Manifold Labs (2023) #4 emission (4.32%). Confidential GPU compute with Intel TDX/AMD SEV-SNP.
SN51 Lium.io (GitHub on bittensor.ai) Not verified — #5 emission. GPU democratization marketplace.
SN8 Vanta (GitHub on bittensor.ai) Not verified — Decentralized trading network.
SN9 IOTA (Macrocosmos) (GitHub on bittensor.ai) Not verified — Distributed LLM pretraining.
SN56 Gradients (GitHub: rayon_labs) Not verified — Competitive training tournaments. Same team as Chutes.

Tier 2: Verified MIT Subnets

SN Name GitHub Repo License Copyright Holder
SN13 Data Universe macrocosm-os/data-universe MIT RusticLuftig (2023)
SN72 StreetVision natixnetwork/streetvision-subnet MIT Opentensor (2023)
SN44 Score score-technologies/score-vision MIT Score Technologies (2024)
SN85 Vidaio vidAio-subnet/vidaio-subnet MIT (badge in README) Not specified in file

Tier 3: No Visible LICENSE File

SN Name GitHub Repo Status
SN62 Ridges AI ridgesai/ridges 404 on LICENSE — no file found at standard paths
SN106 VoidAI v0idai/SN106 404 on LICENSE — no file found
SN120 Affine AffineFoundation/affine 404 on LICENSE — no file found

Note: "Not found" means no LICENSE file at LICENSE, LICENSE.md, or COPYING on the default branch. The repo may still have a license declared elsewhere (README badge, setup.py classifier, or a non-standard filename). It may also mean the repo is intentionally proprietary or the license is declared in a parent org document.

License Text Samples

Chutes (SN64) — MIT

MIT License
Copyright (c) 2024 Jon Durbin
[Standard MIT terms: commercial use, modification, distribution, private use permitted]

Targon (SN4) — MIT

The MIT License (MIT)
Copyright © 2021 Yuma Rao
Copyright © 2023 Manifold Labs
[Standard MIT terms]

Notable: Targon's copyright traces back to Yuma Rao (Bittensor's pseudonymous founder), then to Manifold Labs. This suggests the codebase originated from Bittensor's own template/boilerplate.

Data Universe (SN13) — MIT

MIT License
Copyright (c) 2023 RusticLuftig
[Standard MIT terms]

StreetVision (SN72) — MIT

MIT License
Copyright (c) 2023 Opentensor
[Standard MIT terms]

Notable: Copyright attributed to Opentensor (the Bittensor Foundation), not to NATIX. This may indicate the subnet was built on an Opentensor-provided template.

Score (SN44) — MIT

MIT License
Copyright (c) 2024 Score Technologies
[Standard MIT terms]

Pattern Analysis

1. MIT Dominance

Every subnet with a verifiable LICENSE file uses MIT. This is the most permissive license in common use — it allows anyone to:

  • Use the code commercially
  • Modify the code
  • Distribute the code
  • Sublicense the code
  • Use it privately

The only condition is preserving the copyright notice.

2. Bittensor Template Heritage

Several subnet LICENSE files copyright "Yuma Rao" or "Opentensor" — this traces back to the original Bittensor repository, which is itself MIT-licensed. Many subnets were built using Bittensor's official subnet template, inheriting the MIT license and sometimes the original copyright notice.

3. Code Is Infrastructure, Not the Moat

The Bittensor ecosystem treats subnet code as public infrastructure. The competitive moat comes from:

  • On-chain incentive design (Yuma Consensus weights, emission distribution)
  • Alpha token economics (dTAO liquidity pool, staking dynamics)
  • Network effects (miner/validator participation, holder count)
  • Operational execution (uptime, data quality, partnerships)
  • Owner cut (18% of emissions to the subnet creator)

Not from code licensing. Anyone can read the Chutes repo, but they can't replicate the on-chain emissions, the 12,686 holders, or the 5.11% network share.

4. Split Repositories (Miner/Validator Separation)

Some subnets split their code across multiple repos:

  • Chutes (SN64): chutesai/chutes (SDK/CLI, MIT) + rayonlabs/chutes-miner (miner code) + rayonlabs/chutes-api (validator/API code)
  • Affine (SN120): AffineFoundation/affine + AffineFoundation/affine-cortex + AffineFoundation/affinetes

This means the license of the main repo may not cover all operational code. Miner or validator code in separate repos may have different (or no) licenses.

5. No Copyleft Anywhere

Despite the ecosystem being deeply open-source-aligned, no top-tier subnet uses GPL, AGPL, or any copyleft license. This is a deliberate ecosystem norm:

  • MIT allows commercial entities to build on the code without obligation
  • Copyleft would create friction for enterprise adoption (a key Bittensor goal — PwC, Intel partnerships)
  • The TAO incentive mechanism already ensures contributions flow back to the network

Implications for Ditto (SN118 / Future Subnet)

Current Ditto Approach

From prior research, Ditto's harness codebase uses AGPL-3.0 + commercial dual licensing with a CLA. This is a significantly more restrictive approach than any verified Bittensor subnet.

Comparison

Dimension Bittensor Norm (MIT) Ditto Approach (AGPL-3.0 + Commercial)
Commercial use Free, no restrictions Requires commercial license for proprietary use
Modification Free, no obligations Must share modified source if network-used
Forking Allowed freely Allowed but copyleft applies
Enterprise adoption Zero friction AGPL creates legal review friction
Moat source On-chain incentives Code licensing + on-chain incentives
Ecosystem alignment Strong — matches Bittensor culture Departure — more restrictive than peers

Strategic Questions

  1. Does AGPL actually protect Ditto? The moat in Bittensor is on-chain (emissions, alpha token, network effects), not code. A competitor can't steal Ditto's TAO emissions by forking the code.
  2. Does AGPL hurt adoption? Enterprise partners (like PwC for SN44, Intel for SN4) may hesitate to integrate with AGPL-licensed code due to network-use disclosure obligations.
  3. Is dual licensing worth the complexity? MIT subnets generate revenue through emissions and owner cut. Ditto could do the same without the CLA/dual-license overhead.
  4. Does the Bittensor community expect MIT? The ecosystem norm is so strong that AGPL may signal a lack of alignment with the open-infrastructure ethos.

Recommendation Framing

  • If the goal is ecosystem alignment + enterprise friendliness: Match the Bittensor norm (MIT). The moat is the incentive mechanism and network effects, not the license.
  • If the goal is preventing proprietary forks: AGPL works, but consider that no successful Bittensor subnet has needed it. The on-chain economics already prevent freeriding.
  • If the goal is dual revenue (emissions + commercial licenses): AGPL dual licensing can work, but it's an untested model in the Bittensor ecosystem. The complexity may not be worth it unless Ditto has specific enterprise customers who need a commercial license.

Methodology & Limitations

  • Subnet directory: Sourced from bittensor.ai/subnets (129 subnets as of July 2026)
  • License verification: Checked raw.githubusercontent.com/{owner}/{repo}/HEAD/LICENSE for each subnet's linked GitHub repo
  • Sample size: Verified 7 subnet repos directly; identified GitHub repos for 15+ subnets from bittensor.ai pages and X/community posts
  • Limitations:
    • The bittensor.ai subnet directory page was truncated at ~26 subnets in the rendered output; not all 129 were individually checked
    • Some subnets have multiple repos (miner, validator, SDK) — only the primary repo was checked
    • "Not found" does not mean "proprietary" — the license may exist in a non-standard location
    • Subnet ownership and GitHub links can change (SN44 changed its on-chain GitHub repo during 2026)
    • This research did not check the Bittensor core protocol license (opentensor/bittensor) — that is separate from subnet-level licensing

Sources