Bittensor

RawGraph
FieldValue
TypeDecentralized machine-learning network with blockchain-based incentives
Native tokenTAO (supply capped at 21 million)
FoundersJacob Steeves and Ala Shaabana
Core organizationOpentensor Foundation
First network"Kusanagi", launched January 2021
Current chainSubtensor ("Finney"), built with Substrate in Rust
Incentive systemYuma Consensus; dynamic TAO (dTAO) since February 13, 2025
SubnetsMore than 120 active as of 2026
Websitebittensor.com

Bittensor is a decentralized machine learning network that uses blockchain-based incentives to pay independent contributors for producing digital commodities such as model inference, training, data, and raw compute. The network runs on its own blockchain, called subtensor, and is organized into competitive marketplaces called subnets, each with its own task and scoring rules; participants who perform well earn the network's cryptocurrency, TAO, whose supply is capped at 21 million in imitation of Bitcoin [1][2]. Founded by Jacob Steeves and Ala Shaabana and stewarded by the Opentensor Foundation, Bittensor grew from a single experimental chain in 2021 into an ecosystem of more than 120 specialized subnets by 2026, spanning large language models, GPU rental, deepfake detection, and robotics [3][23][24].

How it works

Bittensor's design separates the blockchain from the machine-learning work. The subtensor chain, a Substrate-based blockchain written in Rust, acts as the coordination and payment layer: it runs consensus, records subnet and participant state, handles staking, and moves value. The actual AI workloads run off-chain inside subnets [4][2].

Each subnet is an independent incentive market defined by its creator. Four roles interact in it [2][5]:

  • Miners produce the subnet's commodity, for example answering language-model queries, renting out GPU capacity, or training model checkpoints.
  • Validators repeatedly test miners and submit weight vectors ranking the value of each miner's work.
  • Subnet creators define the task and the scoring code, and receive a share of the subnet's emissions.
  • Stakers delegate TAO to validators and share in their rewards.

Validator rankings are resolved by Yuma Consensus, an on-chain algorithm that converts the matrix of validator-submitted weights into emission payouts for miners and validators. The algorithm weights the input of higher-staked, more consistent validators more heavily and "clips" outlier evaluations, a defense against collusion in which validators could conspire to over-reward particular miners. Validators are thereby pushed toward honest, convergent scoring, and miners toward work that the whole validator set values [6][1]. The project's whitepaper, published under the name Yuma Rao and titled "Bittensor: A Peer-to-Peer Intelligence Market", presented this peer-ranking market as an alternative to centralized benchmark-driven ML development and claimed resistance to collusion of up to 50 percent of network weight [1].

At the end of each 360-block cycle, a subnet's accumulated emissions are split among participants: as of mid-2025 the documented split was 18 percent to the subnet owner, 41 percent to miners, and 41 percent to validators and their stakers [5]. The chain has also added Ethereum Virtual Machine (EVM) compatibility for smart contracts [6].

Bittensor's approach differs from federated learning, which coordinates training of a single model across trusted parties without payment, and from centralized distributed training clusters: Bittensor does not prescribe how work is done, only how it is scored and paid, and any subnet team can define a new market on the chain [1][2].

History

Accounts of Bittensor's origins vary in detail. The market-intelligence firm Messari records that the project was founded by Jacob Robert Steeves and Ala Shaabana, that the data provider CB Insights dates the company's founding to 2019, and that Steeves describes himself as having worked on Bittensor since 2016 [3]. The Opentensor Foundation launched the first version of the network, "Kusanagi", in January 2021, halted it in May 2021 to fix early consensus problems, and relaunched it in November 2021 as "Nakamoto" via a hard fork; according to network data cited by Messari, 546,113 TAO mined on Kusanagi, roughly 2.61 percent of the eventual total supply, were migrated to the new chain [3].

A testnet called "Finney" went live on January 10, 2023 to trial subnets and delegated staking, and on March 20, 2023 the Nakamoto mainnet was forked into the Finney mainnet, which remains the live subtensor chain. Subnets went live on Finney on October 2, 2023, converting Bittensor from a single machine-learning market into a container for many task-specific markets [3].

In February 2026 Steeves stepped down as chief executive of the Opentensor Foundation, a move framed as reducing the network's dependence on key individuals, and the foundation published a decentralization roadmap targeting December 2027 [32]. During 2026 the project's code repositories also moved from the foundation's longtime "opentensor" GitHub organization to a new "RaoFoundation" organization, and project materials began crediting a Rao Foundation, echoing the whitepaper's pseudonymous author [4][34].

The TAO token

TAO is the network's native asset. Its monetary design deliberately copies Bitcoin: a 21 million hard cap, block rewards that halve over time, and issuance only through network participation. Blocks are produced roughly every 12 seconds, and until late 2025 each block minted 1 TAO, about 7,200 TAO per day [9][10]. The project and ecosystem researchers describe the launch as having had no premine, no initial coin offering, and no investor or team allocation, with every token earned by mining, validating, or staking [9]; the small Kusanagi-era migration noted above predates the current chain [3].

Unlike Bitcoin's calendar-like block-height schedule, Bittensor halvings trigger when cumulative issuance crosses supply thresholds: the first at 10.5 million TAO issued, the second at 15.75 million, the third at 18.375 million. Because TAO spent on subnet and miner registration is "recycled" (returned to the unissued pool), halving dates cannot be predicted exactly [10]. The first halving took effect in mid-December 2025, cutting the block reward from 1 to 0.5 TAO and daily issuance from about 7,200 to about 3,600 TAO; trackers and exchanges had projected December 14, 2025, and recorded the switch taking effect around December 14-15 [10][11][12].

Most circulating TAO is staked: research firm Presto reported in June 2025 that over 72 percent of the roughly 8.8 million circulating TAO was staked back into the network [9]. TAO trades on major crypto exchanges; The Block reported the token trading between 150 and 200 dollars in early February 2026 [23].

Dynamic TAO (dTAO)

Bittensor's original emission governance concentrated power in a "root network" of the 64 largest validators, whose votes decided how much TAO each subnet received. The dynamic TAO (dTAO) upgrade, activated on mainnet on February 13, 2025 after more than a year of development, replaced that arrangement with a market mechanism [8][7].

Under dTAO, every subnet has its own token, informally called an alpha token, paired with TAO in an on-chain constant-product liquidity pool. Staking TAO "into" a subnet now means swapping it for that subnet's alpha token, and each subnet's share of TAO emissions is proportional to its alpha token's price relative to all other subnet tokens. The dTAO whitepaper describes these pool prices as "a view into how the Bittensor network values each subnet", shifting emission allocation from a small validator committee to open market participation [7][5]. Alpha tokens, not raw TAO, became the reward instrument inside each subnet [7].

The upgrade coincided with rapid expansion: an earlier 32-subnet cap had already been lifted, and the network grew from roughly 64 subnets around the time of the upgrade to more than 120 active subnets by early 2026, with subnet registrations reaching network IDs above 120 [23][24]. It also changed what determines a subnet's income, from validator judgment to token markets, a shift that has drawn its own criticism (see Criticism below).

Subnets

Each subnet targets a specific commodity or task. Well-documented examples as of 2026 include:

NetuidSubnetOperatorFocus
1ApexMacrocosmosOpen competitions for algorithmic and agentic optimization, described by Macrocosmos as a routing layer for intelligence [25]
4TargonManifold LabsConfidential compute cloud; GPU and CPU rental secured with Intel TDX and NVIDIA confidential computing [27]
13Data UniverseMacrocosmosDecentralized scraping and storage of training data from platforms such as Reddit and X [26]
34BitMindBitMindDeepfake and AI-generated-content detection, with miners competing in an adversarial generate-and-detect loop [29]
56GradientsRayon LabsModel training and fine-tuning marketplace [28][24]
64ChutesRayon LabsServerless inference for open-weights models, including DeepSeek, Qwen, and Kimi models [28]
80OpenRobotoOpenRobotoOpen competition for continuously improving robotics models [24][30]

Chutes has been the network's most commercially visible subnet: Rayon Labs describes it as a decentralized serverless compute platform serving trillions of tokens per month and offering new open-weights models shortly after release [28]. Targon's operator Manifold Labs, an Austin-based AI infrastructure company, has partnered with Intel on confidential computing for running sensitive AI workloads on untrusted hardware [27]. On August 6, 2026, the robotics project OpenRoboto launched as subnet 80, saying it would coordinate open weights, open data, and open benchmarks through Bittensor to improve robotics models [30][24].

Subnets can also leave. In April 2026 Covenant AI, which ran a distributed-training effort that produced the Covenant-72B model, exited the network entirely and its subnets were marked deprecated (see Criticism below) [31][24].

Ecosystem, investors, and market access

Bittensor is among the oldest and most prominent networks in the decentralized AI segment of the crypto industry, and researchers have compared its monetary design directly to Bitcoin's [9]. Other open-source AI and decentralized training efforts, such as Prime Intellect and Nous Research, run distributed training projects without a Bittensor-style on-chain token market.

Digital Currency Group (DCG), the crypto conglomerate founded by Barry Silbert, has been the most prominent corporate backer. DCG says it first invested in Bittensor in 2021, and in November 2024 it launched Yuma, a subsidiary led by Silbert that provides capital, infrastructure, and technical support to teams building Bittensor subnets [20]. Yuma later added an asset-management arm seeded with 10 million dollars, and its Yuma Composite fund, launched September 4, 2025, tracked a basket of subnet tokens; The Block reported the fund down 31.9 percent from launch through early March 2026, against a 46.1 percent decline in TAO over the same period [23]. Yuma's March 2026 "State of Bittensor" report counted over 120 operating subnets and put the aggregate market value of subnet tokens at 27 percent of TAO's market capitalization [23].

Grayscale Investments launched a Decentralized AI Fund in July 2024 holding TAO alongside other AI-linked tokens, and on August 7, 2024 announced the Grayscale Bittensor Trust, a single-asset TAO fund for accredited investors [18][19]. On December 11, 2025, days before the first halving, the trust began public quotation on OTCQX under the ticker GTAO, which Grayscale described as the first publicly quoted US investment product providing TAO exposure, in remarks reported by The Block [12].

Public companies have also adopted TAO treasury strategies. xTAO, a Bittensor-focused firm that raised 22.8 million dollars from investors including DCG and Animoca Brands, listed on Canada's TSX Venture Exchange in July 2025 [21]. The same month, TAO Synergies, a Nasdaq-listed company formerly known as the biotech Synaptogenix (ticker TAOX), announced a 10 million dollar TAO treasury purchase [22]. Presto Research estimated in June 2025 that DCG's Yuma held over 500,000 TAO, about 2.4 percent of supply, with Polychain Capital holding roughly 200 million dollars worth [9].

Security incidents

Bittensor suffered a string of thefts in mid-2024. In early June 2024, about 11.2 million dollars in TAO was stolen from a large holder [17]. Then on July 2, 2024, attackers drained at least 32,000 TAO, worth about 8 million dollars at the time, from user wallets; TAO's price fell as much as 15 percent on the news [13]. The Opentensor Foundation responded by putting the chain into "safe mode", producing blocks but blocking all transactions, while it investigated [13][14].

The foundation's postmortem identified a supply-chain attack: a malicious version 6.12.2 of the official Bittensor package had been uploaded to the Python package index PyPI, and when users with the compromised package performed operations requiring their decrypted coldkey, such as staking, unstaking, or transfers, the malware exfiltrated the key, letting attackers empty the wallets. The team removed the package, audited its GitHub repositories, and restarted transactions after an upgrade [14][15]. Investigator ZachXBT later tallied more than 28 million dollars in unauthorized transfers from 32 TAO holders between May and July 2024 connected to the incident [16].

Criticism

The 2024 halt itself became a criticism: observers noted that a supposedly decentralized network had been paused by its founding team, which they argued undermined Bittensor's decentralization claims, drawing comparisons to BNB Chain's 2022 halt [17].

Governance concentration has been a persistent complaint. Chain governance has run through a small "triumvirate" of decision-makers, and critics have argued that the core team's control over the incentive layer contradicts the project's positioning [32]. The sharpest rupture came on April 9, 2026, when Covenant AI, a subnet operator whose distributed-training work produced the Covenant-72B model, quit the network. Founder Sam Dare called Bittensor's governance "decentralization theatre" and accused Steeves of keeping effective control of the triumvirate, suspending emissions to Covenant's subnets, removing the team's moderation rights over its community channels, and deprecating its infrastructure unilaterally; TAO fell 15 percent, from 338 to 285 dollars, within two hours of the announcement [31]. The foundation's decentralization roadmap, published after Steeves resigned as chief executive in February 2026, targets December 2027 and includes wider validator competition and conviction-based voting for alpha token holders, with votes weighted by how long holders commit their tokens [32].

A further line of criticism concerns whether the incentive mechanism actually produces useful machine learning. Because dTAO ties a subnet's income to its alpha token's price rather than directly to the usefulness of its output, capital can flow to well-marketed subnets over technically productive ones; commentators at the crypto outlet ChainCatcher argued in 2025 that alpha-token speculation risked decoupling Bittensor's rewards from real AI progress [33]. Supporters respond that the market defunds weak subnets over time. Yuma's 2026 report highlighted subnets selling real services, including Targon's 2 dollar per hour H200 GPU compute and Score's computer vision for video processing, and found subnet value becoming less concentrated, with the top fifth of subnets' share falling from 82 percent to just over 55 percent [23].

See also

References

  1. ^Bittensor: A Peer-to-Peer Intelligence Market (whitepaper), Yuma Rao, Bittensor. bittensor.com/whitepaper
  2. ^Bittensor Documentation, Bittensor/Latent Holdings, 2026. docs.learnbittensor.org
  3. ^Bittensor Profile, Messari. messari.io/...profile
  4. ^subtensor repository (Substrate-based Bittensor blockchain), Rao Foundation (formerly Opentensor), GitHub. github.com/...subtensor
  5. ^Emission, Bittensor documentation, June 11, 2025 (archived July 17, 2025). web.archive.org/...emissions
  6. ^Yuma Consensus, Bittensor documentation, June 18, 2025 (archived August 11, 2025). web.archive.org/...yuma-consensus
  7. ^dTAO Whitepaper: subnet token economics, Yuma Rao, Dr. Nick, and 0xcacti, Bittensor. bittensor.com/dtao-whitepaper
  8. ^Bittensor activates dynamic TAO, restructuring emission model around real-time staking flows, Crypto Briefing, February 2025. cryptobriefing.com/...r-dynamic-tao-emission-model
  9. ^From Bitcoin to Bittensor: The Next Monetary Primitive, Min Jung, Presto Research, June 25, 2025. prestolabs.io/...ensor-the-next-monetary-primitive
  10. ^Halving, Taostats documentation. docs.taostats.io/...halving
  11. ^Bittensor Halving: All You Need to Know, Crypto.com, December 2025. crypto.com/...bittensor-halving-all-you-need-to-know
  12. ^Grayscale's Bittensor Trust hits public markets days before first halving event, The Block, December 11, 2025. theblock.co/...ets-days-before-first-halving-event
  13. ^Bittensor's TAO Slides 15% After $8M Wallet Drain Attack, CoinDesk, July 3, 2024. coindesk.com/...es-15-after-8m-wallet-drain-attack
  14. ^Bittensor Community Update, July 3, 2024, Opentensor Foundation, Medium. blog.bittensor.com/...ate-july-3-2024-45661b1d542d
  15. ^Explained: The Bittensor Hack (July 2024), Halborn, July 2024. halborn.com/...lained-the-bittensor-hack-july-2024
  16. ^Thread on Bittensor thefts May-July 2024, ZachXBT, X, October 2025. x.com/...1978465694644899917
  17. ^'Decentralized' AI network Bittensor halted in response to $8M hack, Protos, July 2024. protos.com/...tensor-halted-in-response-to-8m-hack
  18. ^Grayscale Investments Launches Two New Product Offerings: Grayscale Bittensor Trust and Grayscale Sui Trust, GlobeNewswire, August 7, 2024. globenewswire.com/...Trust-and-Grayscale-Sui-Trust
  19. ^Grayscale Introduces Crypto Funds for Decentralized AI Project Bittensor's TAO and Layer-1 Network Sui's Token, CoinDesk, August 7, 2024. coindesk.com/...tao-and-layer-1-network-suis-token
  20. ^DCG launches new Bittensor-focused subsidiary Yuma, doubling down on decentralized AI investments, The Block, November 2024. theblock.co/...own-on-decentralized-ai-investments
  21. ^Bittensor-focused firm xTAO to list common shares on TSXV; raises $22.8 million from DCG, Animoca Brands, The Block, July 2025. theblock.co/...xtao-bittensor
  22. ^Nasdaq-listed TAO Synergies buys $10 million in Bittensor tokens for treasury, The Block, July 2025. theblock.co/...on-in-bittensor-tokens-for-treasury
  23. ^DCG subsidiary Yuma highlights subnet development in 'State of Bittensor' report, The Block, March 5, 2026. theblock.co/...dcg-yuma-bittensor-report
  24. ^Subnets, Taostats (Bittensor network explorer), accessed August 7, 2026. taostats.io/subnets
  25. ^Subnet 1 Apex, Macrocosmos Developer Guide. docs.macrocosmos.ai/...subnet-1-apex
  26. ^Subnet 13 Data Universe, Macrocosmos Developer Guide. docs.macrocosmos.ai/...subnet-13-data-universe
  27. ^Targon: Confidential Compute Cloud, Manifold Labs. targon.com
  28. ^Chutes: Serverless AI Compute, Rayon Labs. chutes.ai
  29. ^BitMind Forensics ranks among top deepfake detection systems with decentralized AI approach, Crypto Briefing, 2026. cryptobriefing.com/...nsics-top-deepfake-detection
  30. ^OpenRoboto launch announcement (Bittensor Subnet 80), OpenRoboto, X, August 6, 2026. x.com/...2085369343253332175
  31. ^'It is decentralization theatre': Covenant AI exits Bittensor, TAO drops 15%, The Block, April 2026. theblock.co/...covenant-ai-exits-bittensor-tao
  32. ^Bittensor founder targets full decentralization within 18 months, Crypto Briefing, June 22, 2026. cryptobriefing.com/...ralization-roadmap-18-months
  33. ^dTAO mechanism out of control? Bittensor is deviating from the AI track, ChainCatcher, 2025. chaincatcher.com/...2173294
  34. ^The Bittensor Paradigm, Const (core developer), Bittensor. bittensor.com/about

Improve this article

Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.

1 revision · v2 · 2,916 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Reviewer note: Tokenomics, history, incidents, and subnet listings verified against whitepapers, taostats, Wayback captures, and registries on August 7-8, 2026; a subnet-growth framing error was corrected.

Cite this page: AI Wiki. "Bittensor." aiwiki.ai, updated 7 Aug 2026, fact-checked 7 Aug 2026. CC BY 4.0. https://aiwiki.ai/wiki/bittensor

Suggest edit

What links here