What is Bittensor (TAO)?

Bittensor is an open network that pays people to produce artificial intelligence rather than consume it. Participants can contribute models, compute capacity or data, with TAO rewards based on how useful other participants judge their output to be.

Its economics borrow from Bitcoin and apply that model to machine learning. TAO has a hard cap of 21 million tokens, with halvings built into the code. The January 2021 launch included no venture allocation, pre-mine or token sale, meaning every TAO in existence was earned through participation.

The work takes place inside subnets. Each is an independent market dedicated to a task such as inference, model training, compute provision or price prediction, and each defines valuable work for itself. There are currently 128 active subnets, with an expansion toward 256 planned.

The network has also moved beyond theoretical usage. During the first quarter, subnets generated roughly $43 million in AI service revenue. Nvidia's chief executive publicly compared Bittensor's distributed training to folding@home. Meanwhile, TAO trades near $190 after falling around two thirds from its early-2024 peak.

How Does Bittensor Work?

Bittensor coordinates four participant groups through a scoring system that converts judged quality into token payments. Work enters the network through these roles, while rewards flow back according to performance.

1. Subnets

A subnet functions as a self-contained market for one digital commodity. Its creator defines the incentive mechanism: what miners need to produce and the rules validators use to grade it. As a result, every subnet becomes a separate experiment in pricing a particular kind of intelligence.

Each subnet supports up to 256 participant slots. Once those slots are full, a new registration displaces the weakest performer. Competition between subnets follows the same principle against the network-wide cap.

2. Miners

Miners operate the models or hardware that generate a subnet's output. On an inference subnet, they respond to validator queries and are scored for quality, latency and robustness. A training subnet may instead reward gradients, checkpoints or verified model improvements.

Uptime alone does not determine rewards; performance rankings do. Mediocre output earns proportionally less, and sustained underperformance can result in deregistration when a new participant takes the slot.

3. Validators

Validators query miners and score responses according to subnet rules before submitting those weights on chain. Those scores determine how miner emissions are divided, turning subjective assessments of quality into objective payments.

Meaningful stake is required to run a validator. Yuma Consensus reconciles disagreements among validators so one dishonest scorer cannot redirect emissions. The scoring process relies on validator judgement rather than cryptographic proof of the computation. Validators retain a configurable take and distribute the remainder to delegators.

4. Stakers and Delegators

People who want exposure without running infrastructure can participate through staking. Under dTAO, staking TAO into a subnet is not simply a deposit. The TAO enters the subnet's liquidity pool and is exchanged for its alpha token.

That structure introduces a distinct risk. Because the staker holds a subnet-specific asset, the alpha token can decline against TAO even while its subnet continues earning emissions. Returns therefore depend on alpha's price as well as the yield.

5. Emissions and Registration

Emissions are distributed in two stages. Every block injects TAO and alpha into subnet pools under a cross-subnet formula. At the end of each tempo, which lasts roughly 360 blocks, the accumulated rewards are distributed within each subnet through Yuma Consensus.

Registration requires TAO, with the fee adjusting dynamically according to demand. New participants receive an immunity period that protects them from immediate pruning. Once that period ends, a new registration replaces the lowest-emission node that is outside immunity.

What is Dynamic TAO (dTAO)?

The dTAO upgrade launched in February 2025 and changed how the network decides where emissions go. Previously, 64 root validators formed a council that voted on which subnets should receive them, concentrating enormous influence within a small group.

With dTAO, each subnet issues an alpha token paired with TAO in an automated market maker. Staking activity continuously sets the alpha price, which in turn determines the subnet's share of daily emissions. Capital allocation therefore shifts from a committee vote to an open market.

Standard subnet rewards allocate 41% to validators and another 41% to miners, leaving 18% for the subnet owner. By March, all alpha tokens together had reached roughly $1.12 billion in market capitalisation, equivalent to about 27% of TAO's own market cap.

Adoption has come with criticism. Rules for subnet economics have changed repeatedly, including net TAO movement, price-plus-burn and an emission gate that directed more rewards toward higher-ranked subnets. Teams with real products have also been deregistered as parameters shifted.

Leading Bittensor Subnets

Subnet valuations show where the market sees real demand. Measured in late March, the leaders by market capitalisation span training, inference and compute:

• τemplar (SN3), ~$134.9M: Completed Covenant-72B in March, a 72-billion-parameter model trained across more than 70 distributed nodes. Its algorithm compressed gradients by 97% without losing accuracy.
• Chutes (SN64), ~$132.9M: A serverless inference platform that has processed 9.1 trillion cumulative tokens, with daily peaks above 50 billion. By using idle decentralised compute, it prices services roughly 85% below AWS.
• Targon (SN4), ~$91.8M: A confidential GPU compute marketplace from Manifold Labs. It raised a $10.5 million Series A and moved Dippy AI, an application with 8.6 million users, onto decentralised infrastructure.
• Affine (SN120), ~$71.8M: A reinforcement learning environment in which only models on the Pareto frontier receive meaningful rewards. It hosts its own models on Chutes instead of duplicating infrastructure.
• Lium (SN51), ~$52.1M: A peer-to-peer GPU marketplace with more than 500 Nvidia H100 units onboarded. Validators programmatically verify hardware specifications rather than relying on provider claims.
• Ridges AI (SN62), ~$50.8M: Builds autonomous agents for end-to-end software engineering tasks and competes directly with centralised coding assistants.

One development team has a dominant position. Rayon Labs operates Chutes as well as Gradients (SN56) and Nineteen (SN19). Together, the three subnets receive roughly 23.7% of all daily TAO emissions.

TAO Tokenomics and the Halving

Few major tokens follow Bitcoin's supply design as closely as TAO. Its issuance mechanics include:

• Hard cap: Total supply is fixed at 21 million TAO, with roughly 11 million circulating. Governance cannot increase that figure.
• Halving: The first halving occurred in December 2025. Daily emissions fell from 7,200 to 3,600 TAO, reducing daily sell pressure by an estimated $500,000.
• Emission split: Under the dTAO model, block rewards allocate 41% to miners, 41% to validators and 18% to subnet owners.
• Delegator share: Validators pass roughly 82% of their emissions to TAO holders who delegate stake to them, minus a configurable take.
• Staked supply: Around 70% of supply is staked. Reported institutional positions lock up an additional portion, reducing the liquid float available for trading.
• Recycling and burns: Registration fees and unspent rewards return to unissued supply. Subnet owners can also permanently burn alpha, with both mechanisms pushing the next halving further out.

Price and Institutional Positioning

TAO trades near $190 with a market capitalisation of around $3 billion, leaving it roughly two thirds below its early-2024 record near $757. The token returned 21.57% in the first quarter, recovering from $230 to close near $251 before moving lower again.

Institutional capital arrived quickly. Reports put Nvidia's position near $420 million, with 77% staked, and Polychain Capital's at roughly $200 million. Neither company has confirmed those figures on the record, however, and no on-chain verification has been published.

Regulated access remains the pending catalyst. Grayscale filed to convert its Bittensor Trust into a spot ETF on NYSE Arca, while Bitwise submitted a competing product. An SEC decision window opened in August. Grayscale also increased TAO's weighting in its AI-focused fund to 43.06%.

Institutional Adoption and the ETF Question

Bittensor has drawn more attention from traditional finance than most decentralised AI projects. Regulated wrappers matter because they determine whether pension and wealth capital can hold the asset at all.

Grayscale's Bittensor Trust trades under the GTAO ticker and has traded at a substantial premium to net asset value, a pattern that preceded Bitcoin's own ETF conversion. At the end of June, its most recent filing reported 46,331 TAO against $9.4 million in net assets.

ETF approval is far from assured. TAO does not have a regulated futures market on a venue such as the CME. Grayscale has also abandoned applications linked to Cardano, Hedera and Polkadot in the past, leaving the AI crypto sector's first spot ETF an open question.

Custody infrastructure has expanded alongside these filings. BitGo partnered with Yuma to provide institutional-grade custody and staking. Kraken completed native dTAO integration as well, with plans to list subnet alpha tokens directly and broaden access beyond the native wallet stack.

How to Use Bittensor

Participation ranges from relatively simple staking to operating an entire subnet, with each role requiring a different level of commitment and technical skill. The process starts by buying TAO on a crypto exchange.

From easiest to hardest, the available paths are:

1. Stake TAO: Delegate to a validator or stake into a specific subnet. Both can earn a share of emissions, but choosing a subnet instead of the root also introduces alpha price risk.
2. Buy alpha tokens: Purchase TAO on an exchange and transfer it to a compatible wallet. From there, swap it for a subnet's alpha token through Taostats or the wallet interface.
3. Run a miner: Register on a subnet and operate the required models or hardware. Competing against other miners for emissions requires GPUs and continuous optimisation.
4. Operate a validator: Acquire enough stake to run validation infrastructure, then score miner output under the subnet's rules while earning a take from delegators.
5. Launch a subnet: Create an incentive mechanism and pay the registration cost, then attract miners and validators while competing with existing subnets for one of the capped slots.

Wallets and Key Management

Bittensor assigns different functions to separate keys. The coldkey protects the TAO balance and remains offline. Active operations such as mining and validating use the hotkey instead, which means compromising a hotkey does not expose the underlying holdings.

Storage, transfers and staking are supported through the official browser extension, while Talisman and SubWallet provide alternatives. Anyone with a meaningful balance should protect coldkey storage with the same care used for any self-custodial wallet.

Developer and Analytics Tools

For developers, the Python SDK is the network's primary building tool. It supports subnet creation, participant registration and scripted interaction. Those who prefer a terminal can use the command line interface for wallet creation, key management and staking.

Taostats has operated as Bittensor's block explorer since 2022, tracking emissions and subnet performance as well as validator rankings and tax reporting. When evaluating a subnet, net staking movement and liquidity pool depth are more important than headline market capitalisation.

Bittensor Founders and Governance

Jacob Steeves and Ala Shaabana co-founded Bittensor and launched mainnet in 2021. After studying mathematics and computer science, Steeves worked as a software engineer at Google. He has worked at the intersection of Bitcoin and machine learning since 2015.

Shaabana holds a doctorate in computer science and was an assistant professor at the University of Toronto before co-founding the network. Both founders continue contributing through the Opentensor Foundation, which stewards protocol development.

In practice, governance has been contentious. The Covenant AI operator exit caused a 38% drawdown in the token. The community later restored affected subnets without central intervention, an outcome supporters framed as evidence of resilience while critics saw concentration risk.

Risks to Understand

Bittensor combines early-stage AI economics with a volatile token, creating exposures that are specific rather than generic. Consider these risks before committing capital:

• Alpha token risk: Staking into a subnet means holding its alpha token. It can decline against TAO even when the subnet earns emissions, so receiving yield does not guarantee a positive return.
• Rule instability: Emission and registration parameters have changed repeatedly. Teams that based hiring and infrastructure decisions on previous rules have subsequently been deregistered.
• Concentration: One development group controls roughly 23.7% of daily emissions through three subnets. Reported institutional positions concentrate an additional portion of supply.
• Deregistration: Underperforming subnets and participants can be removed, meaning builders risk losing a slot regardless of product quality.
• Unverified figures: Widely reported institutional investment figures have neither on-chain confirmation nor company statements behind them. Headline inflows should therefore be treated as reported rather than established.
• Revenue scale: Quarterly AI service revenue of about $43 million is real, but remains small relative to a $3 billion market capitalisation. Valuation therefore depends on future demand.
• Liquidity: Smaller alpha tokens trade in automated market makers whose pools can fall below $1 million. Large orders can move prices sharply, and exit liquidity may disappear during volatile periods.

Final Thoughts

Bittensor has built something genuinely unusual: a market that rewards machine intelligence according to judged quality rather than promises. dTAO extended that model by shifting decisions over which AI tasks receive funding from a validator council to capital.

There is evidence of real usage. A 72-billion-parameter model has been trained across distributed nodes, while trillions of inference tokens have been served and enterprise applications have migrated onto subnets. Those examples point to a network performing work rather than merely issuing tokens.

Durability remains the open question. Rule changes have unsettled builders, a small number of teams capture a concentrated share of emissions, and quarterly revenue is still modest relative to valuation. For anyone following decentralised infrastructure, Bittensor remains the most advanced attempt to price intelligence openly. The ETF decision will shape how much capital can follow that thesis.