According to the whitepaper (Rao 2022), Bittensor is “a market where intelligence is priced by other intelligence systems peer-to-peer across the internet”.

Statistics and the current state:

• TAO explorer: price, market cap (1b at 2026-01-10), subnets (around 200) and more, very discoverable dashboard, even has its chatbot.

Further reading:

• Documentation (and glossary).
• LearnBitTensor: among other stuff, contains detailed explanations how some large subnets work (e.g. see Templar).

Applications

The DAO consists of numerous subnets each being a separate startup generating value for its clients. There are about 200 of them now (2026-01-10), grouped into several dozen types like Agents, 3D, Programming etc. Subnets vary in their quality and maturity. Few examples of top subnets:

• 404, a company providing quick 3d assets AI-assisted generation (also see its CEO interviewed). There is a ton of info: twitter, discord, blender and unity plugins, demos, even a game made in a month as a demo project. Bittensor-related code is kept in a github repo.
• Chutes, an AI models infrastructure which incentivises its "miners" to keep models "hot" (loaded into memory) and thus provide low latencies.

And the "bottom" ones:

• Talisman, supposedly some sentiment analysis network, but appears to be a part of some indistinguishable crypto wallet startup, one of thousands alike. No info about the subnet and its features is given.

Bittensor acts as a marketplace for their subnets: attracts miners and validators, provides rewards. Consider this quote from 404 CEO interview:

By operating within Bittensor’s decentralized AI network, we ensure that our platform scales organically, improving as more contributors join. This system incentivizes innovation—miners are rewarded for improving the models, meaning that our AI is constantly evolving based on real-world demand. It also removes reliance on a single entity, ensuring that creators have open, unrestricted access to cutting-edge 3D tools, rather than being locked into proprietary ecosystems.

Architecture

The whitepaper (Rao 2022) builds a formal model and proves its properties. The network consists of peers, each providing its own [ML] model trained on a dataset with a loss function. Peers evaluate models of each other, assigning weights to them. Additionally, peers have stakes. When a model gets a stake-weighted review from 50% of a network, it is said to get consensus. There are incentives to avoid collusions and spur consensus level growth. All this mechanics supposedly maximizes the amount of intelligence provided.

Unfortunately the paper is very terse telling how things would work together in a real world: what problems would peers solve, who will use their work, how peers would evaluate others models quality.

Possibly related field: mixture-of-experts models.

The network has two types of nodes (neurons): miners (servers, axons) and validators (clients, dendrites). Servers do the computations and validators validate their work, assigning weights to them according their performance. Validators have stakes at play but servers don't.

Consensus

The consensus mechanism (so-called Yuma consensus, by the name of an original whitepaper author) is described in details, with a simulation.

The system is initiated with a set of hyperparameters:

• $\kappa$, the ratio of stake that decides consensus. Its typical value is 0.5. The lower it is, the lower consensus is required;
• $\beta$, the bonds penalty;
• $\xi$, the emission ratio.

Glossary:

• Validator stake $S_i$: amount of network tokens validator $i$ has.
• Server consensus $W_j$: a weight value that maximizes the total stake-adjusted weight which surpasses $\kappa$.
• Consensus score $C_j$: ??
• Consensus-clipped weight $\bar W_{ij}$: a maximum of a weight given by validator $i$ to server $j$, and a consensus for this server.
• Server rank $R_j$: stake-adjusted sum of consensus-clipped weights. (Not actually a statistical rank but just another weight.)
• Server incentive $I_j$: percent of the rank of a server to the total ranks given in the turn.
• Validator bond $B_i$: ??
• Validator reward $R_i$: equals incentive-weighted bonds given to servers by a validator.
• Emission $E_i$: new stake issued after each turn and given to validator $i$.

Each turn validators assign weights to servers, adjusted weights are calculated, and incentives and bonds are emitted (new tokens are minted) and doled out to servers and validators. The reward is distributed in the following proportions: 18% to subnet owners, 41% to the miners, and 41% to the validators.

The adjustment algorithm keeps the stake of honest majority of the net to this majority subject to wide ranges of malicious behaviour of network minority, i.e. emissions to the honest part keeps being greater than the stakes of this majority: $E_H \ge S_H$ . For instance, simulations show that a honest network with 60% stake can retain its stake if if generates 70% or more utility.

The reason these guarantees are weaker than e.g. Bitcoin 51% limit is that the network doesn't have a provable objective utility, like Bitcoin's PoW, and thus it must rely on bids of the peers which are at least partly subjective. (There is a pack of plots showing the simulation results under various levels of subjectivity.)

Governance

A section in the docs describes the governance process. The control over the network was initially held by one person. Then it transitioned to a system of a Triumvirate (three members able to create proposals for changes) and a Senate (up to 12 members who vote to approve or discard a proposal). The next stage is promised to be more decentralized/democratic.

Each member of the Senate has a stake of network tokens in game. Additionally, other members may delegate their stakes to a senator of choice. To be elected, one must aggregate at least 2% of the total network value in own funds and delegated stakes.

Dynamic TAO

The proposed next stage of the governance is Dynamic TAO (Rao 2025). It changes incentives allocation across subnets to make weights distribution among subnets more optimal. The idea is to introduce subnet tokens and make them to be openly traded in a constant-product automatic market makers (cpAMMs), thus eliciting subnets value via the market.

AAMs usually need liquidity providers (LPs) who stake their TAO tokens into AMMs, and also attracts speculators arbitraging the prices among AMMs. However, Bittensor AMMs will gain liquidity only through emission, so no LPs and no swap fees present here.

Glossary:

• APY: annual percentage yield
• Root validators / root network: ??

Development

There is an org on Github: opentensor. The blockchain layer, subtensor, is written in Rust.

There is an SDK for users.

Conclusions

There are the following groups of stakeholders in the DAO:

• servers (miners) who produce the common good (the computations) and gain 41% from the network emission, and also benefit (or suffer) from the network token price movement;
• validators who make the network operate in zero-trust environment also get 41% of the reward and also have interest in maintaining DAO tokens prices high;
• subnet owners, of which I don't know much yet;
• users who order computations from the servers and pay for them with the network tokens, thus keeping the tokens price high.

Their incentives are clear.

Then, there are governing structures: the Triumvirate and the Senate, which accumulate network token stakes. Supposedly, these people are from the companies/organizations running servers and/or validators and thus having stakes.

Also, there are developers, analysts and others who improve the network. They possibly also have stakes, run servers, own stakeholding entities or work for such.

References