Decentralized AI: what Bittensor (TAO) is and how it works

The idea: a market for machine intelligence

Bittensor's whitepaper, published under the pseudonym Yuma Rao, proposes "a market where intelligence is priced by other intelligence systems peer-to-peer across the internet". Instead of one company deciding which model is best, participants score each other, and the scores decide who gets paid.

Bitcoin makes a useful comparison. It pays miners for one narrow, easily checked job: securing its ledger with computing power. Bittensor wants to pay for jobs that are far harder to define, such as answering questions well, and lets each market define its own job.

The blockchain itself, called Subtensor, runs no AI. Built with Substrate, the toolkit behind Polkadot, it produces a block every 12 seconds, records who takes part, collects the scores and issues TAO. The work happens off-chain, on participants' own machines.

Subnets, miners and validators

A subnet is an independent market with its own number (its "netuid") and its own rules. More than 120 were running in mid-2026, each built around a different kind of digital work.

Every subnet has four parts:

• An owner, who registered the subnet and wrote its incentive mechanism: what counts as good work and how it is tested.
• Miners, who do the work. A subnet has 256 slots by default, and a newcomer can push out the lowest-ranked miner once that miner's protection period is over.
• Validators, who test the miners and record scores on-chain, called "weights". By default only the 128 participants with the most stake behind them hold a validator permit.
• A pool pairing TAO with the subnet's own token, called alpha.

Anyone holding TAO can stake it with a validator and share that validator's rewards, minus a commission. A special subnet, number 0 or "root", has no miners at all: it is where TAO is staked directly.

How rewards are decided: Yuma Consensus

Every tempo, 360 blocks or about 72 minutes by default, the chain gathers the validators' weights and runs Yuma Consensus. The whitepaper describes the goal as resisting collusion by up to half of the network's weight.

For each miner, the chain finds the highest score that validators holding at least half of the stake support: a stake-weighted median. Any score above that level is cut down to it, so a validator that wildly over-rates a friend's miner simply has the excess ignored. Each miner's share of the miners' rewards is the stake-weighted sum of these trimmed scores.

Validators are paid through "bonds", which grow when they back miners the consensus rates highly, so those who spot good miners early earn the most. To stop lazy validators from copying others' weights, subnets can switch on "commit-reveal", in its current form since December 2024: weights stay encrypted for several tempos, so a copier only sees stale data.

TAO's supply: 21 million and halvings

TAO borrows Bitcoin's monetary design. There will never be more than 21 million TAO, and issuance halves each time half of the remaining supply has been issued: at 10.5 million TAO, then 15.75 million, then 18.375 million. Until the first halving the chain minted 1 TAO per block, about 7,200 a day.

The trigger is the amount issued, not a block number. Registration and transaction fees are "recycled": subtracted from the issued total so they can be minted again. That pushes each halving back and makes its exact date impossible to predict.

The first halving came in December 2025, when 10.5 million TAO had been issued. Since then the chain has minted 0.5 TAO per block, roughly 3,600 a day. At that pace the next halving, at 15.75 million, comes about four years after the first, and recycling will push it later. Each subnet's alpha token follows its own version of the curve, with its own 21 million cap.

Dynamic TAO: letting the market value subnets

Until February 2025, each subnet's share of new TAO was set by validators' votes, weighted by the TAO staked with them. The Dynamic TAO (dTAO) whitepaper argued that validators could not seriously evaluate a growing number of subnets, and called the old arrangement an "apathetic oligarchic voting system".

dTAO, which went live that February, gave every subnet an alpha token and a pool holding TAO and alpha. Staking TAO into a subnet now means swapping it for that subnet's alpha at the pool's price, and unstaking swaps it back. One subnet's alpha is not interchangeable with another's.

Each block's new TAO is divided among subnets according to demand for their tokens, but the measure of demand keeps changing. From November 2025, shares followed staking flows: TAO moving into each subnet minus TAO moving out. In June 2026 the network returned to prices, using a moving average of each alpha token's price. Since an upgrade in late July 2026, an "emission gate" also sharply cuts what subnets below a demand bar receive; the bar currently sits around the 32nd-ranked subnet.

The TAO a subnet earns goes into its pool, while newly minted alpha is paid to its owner, miners, validators and stakers in the 18/41/41 split.

Criticisms and risks

Concentration

Influence in Bittensor follows stake: validator permits go to the biggest stakeholders, and their scores count in proportion to stake. Rewards are concentrated too. By the developers' own model, the July 2026 gate raised the top eight subnets' share of emissions from about a third to more than half.

Control has been concentrated in emergencies as well. In July 2024, an attacker used a malicious version of the bittensor software package to steal about 32,000 TAO from users. The Opentensor Foundation, which led development, stopped the theft by putting validators behind a firewall in "safe mode", halting transactions. That protected users, but showed that a small group could pause the chain.

Measuring quality

Each subnet's owner decides how work is tested, and miners optimize for whatever the test rewards, which is not always the same as being useful. Weight copying shows that validators, too, can be tempted to follow the crowd instead of testing.

Market-based rewards have their own blind spot: token demand reflects what holders expect as much as what a subnet delivers. Explaining the July 2026 gate, the developers wrote that under the previous rule "a subnet nobody uses still collects a slice of every block".

Volatility and change

TAO and alpha tokens can swing sharply, and a small subnet's pool can be shallow, so large stakes or withdrawals move its price. The rules move too: the way subnets are paid changed three times between November 2025 and July 2026, and each change shifted rewards between subnets.

Where Bittensor sits in decentralized AI

"Decentralized AI" covers several different bets. Some projects rent out computing power, others coordinate the training of a single model across many machines, and Bittensor pays for results.

Bittensor's distinctive choice is to price outputs rather than hardware, and to run many such markets under one token. That makes it flexible, since a subnet can reward almost any digital work, but it puts all the weight on how well each subnet measures what it pays for.