Bittensor (TAO)

Bittensor (TAO) is the native cryptocurrency of Bittensor, an open network that pays participants in tokens for providing machine-learning services such as text generation, data collection, model training and forecasting. TAO is earned by the computers that do this work and by those that judge it, staked to back particular services, and spent to register new ones. Supply is capped at 21 million TAO, a figure deliberately borrowed from Bitcoin (BTC), and new coins enter circulation only as rewards: there was no pre-sale and no allocation to investors.

The network is divided into subnets, each a separate marketplace for one kind of AI task with its own way of measuring quality. Bittensor's own blockchain, called Subtensor, keeps the accounts and divides the rewards. Its distinctive idea is that the network itself sets the price of machine intelligence rather than simply renting out computer power.

Imagine a fairground full of open contests, each in its own tent. In one tent, contestants write answers to questions; in another, they predict tomorrow's weather. Judges examine every entry and score it, and a daily prize pot is shared out according to those scores. Anyone can enter, anyone with enough standing can judge, and the best performers take home the most.

TAO is both the prize money and the entry ticket. Contestants, called miners, earn TAO for good work. Judges, called validators, earn TAO for scoring honestly, and they must put their own TAO at stake, so rigging the scores would cost them. Visitors who think a particular tent is doing useful work can back it with their TAO, which steers more of each day's prize money towards that tent.

Bittensor was founded by Jacob Steeves and Ala Shaabana, and its development is overseen by the Opentensor Foundation. A first version of the network, Kusanagi, went live in January 2021 and was halted in May 2021 to fix problems with its consensus; it was relaunched as Nakamoto in November 2021. The chain was later rebuilt on Substrate, the software framework also used by Polkadot (DOT). In July 2024 a malicious copy of the Bittensor software uploaded to the Python package index stole users' private keys and about 32,000 TAO, then worth roughly $8 million, and developers briefly halted the chain. Dynamic TAO went live in February 2025, the first halving followed in December 2025, and later that month Grayscale filed to turn its Bittensor trust into a US exchange-traded fund.

Each subnet has an owner who defines the task and the scoring method, miners who produce the work, and validators who query the miners, score their output and record the scores on chain. An algorithm called Yuma consensus combines those scores, giving more weight to validators with more stake and reducing the influence of any validator whose scores differ sharply from the rest. The rewards from each block are then split among the subnet owner, the miners, the validators and the holders who back those validators. People who do not run hardware can still take part through delegated staking, assigning TAO to a validator and receiving part of its rewards. By late 2025 more than 100 subnets were active, covering a wide range of AI tasks, and running a miner or validator works much like being a node operator on other networks.

Until February 2025, a small group of large validators decided by vote how the daily issuance was divided between subnets. Dynamic TAO, which went live on 13 February 2025, handed that decision to the market. Each subnet now has its own token, called alpha, paired with TAO in a liquidity pool. Staking TAO into a subnet means swapping it into that subnet's alpha, and subnets whose tokens attract more demand receive a larger share of new issuance. The change also exposed stakers to the price swings of individual subnet tokens.

New TAO is minted with every block, roughly every 12 seconds, and the block reward halves each time total issuance passes a set threshold, the first being 10.5 million TAO, half of the cap. Daily issuance is:

• $E_{\text{day}}$ is the number of new TAO created per day.
• $86{,}400$ is the number of seconds in a day.
• $t_{\text{block}}$ is the block time in seconds, about 12.
• $R$ is the reward per block, in TAO.

Before the halving the reward was 1 TAO per block. With 86,400 ÷ 12 = 7,200 blocks a day, that gave 7,200 × 1 = 7,200 TAO a day. After the first halving on 15 December 2025, $R$ fell to 0.5, so issuance dropped to 7,200 × 0.5 = 3,600 TAO a day. Registration fees paid in TAO are recycled back into the unissued supply, so the timing of later halvings depends on network activity. The schedule mirrors the 21 million bitcoins model, though Bittensor's blocks are far more frequent.

Critics question how much of the AI work produced on Bittensor is used by paying customers rather than generated mainly to earn rewards, and argue that miners have strong incentives to game scoring methods instead of doing useful work. Governance has also been more centralised than the fair-launch story suggests: protocol changes have been driven largely by the Opentensor Foundation, and stake is concentrated among a relatively small number of large validators. The July 2024 theft showed the risks of depending on community software, and many subnet tokens trade in thin, volatile markets. TAO gives no claim on any company's revenue, and its treatment under cryptocurrency regulation remains unsettled in many countries.

• Bittensor
• Bittensor documentation