Bittensor is the best-known attempt to build an open market for AI on a blockchain. Instead of one company training and serving models, Bittensor pays independent “miners” in its token, TAO, to do AI work such as running inference, training models or scraping data, and pays “validators” to score that work. In 2026 it has been both the AI-crypto project with the most real usage and one of the most controversial, thanks to a public governance fight in April. This guide explains how Bittensor actually works, what changed in 2025 and 2026, and the risks you should understand. If you want a broader overview of AI tokens first, read my AI crypto tokens explained guide.
- What it is: a blockchain (Subtensor, built on Substrate) that rewards AI work done off-chain in competing “subnets”.
- Subnets: 128 active subnets as of October 2026 (capacity raised to 256 in May 2026). Each is a mini-market for one AI task.
- Dynamic TAO (since Feb 2025): each subnet has its own alpha token; staking TAO into a subnet steers emissions to it.
- Supply: 21 million TAO max, about 11.68 million circulating. First halving in December 2025 cut emissions from about 7,200 to 3,600 TAO a day.
- Market (early November 2026): roughly $230 to $260 per TAO, about 69% below its $757.60 all-time high, market cap near $2.7 billion.
- Main risks: concentrated governance, emissions tied to market flows, past security incidents and high volatility.
How Bittensor Works
Bittensor has two layers:
- The chain (Subtensor). A Layer 1 blockchain built on Substrate, the framework behind Polkadot, producing a block about every 12 seconds. It records stakes, registrations, validator scores and token emissions. It does not run AI itself.
- The work (subnets). AI tasks happen off-chain on miners’ own hardware. The chain only decides who gets paid, based on validators’ scores.
Think of it as a set of permanent competitions. Each competition (subnet) defines a task and a way to grade it. Miners compete to do the task best, validators grade them, and the network pays out newly issued TAO according to those grades.
Subnets, Miners and Validators
| Role | What they do | How they earn |
|---|---|---|
| Subnet owner | Designs the task and scoring rules, runs the subnet | A share of the subnet’s emissions (0 to 33% since the October 2026 eTAO upgrade) |
| Miner | Does the AI work: inference, training, data, predictions | At least 33% of emissions, split by validator scores |
| Validator | Tests and scores miners, sets weights on-chain | Up to 50% of emissions, weighted by stake and agreement |
| Staker | Delegates TAO to a validator or subnet | A share of the validator’s rewards |
Scoring uses Yuma Consensus, which rewards validators whose scores agree with other validators, weighted by how much stake they hold. In October 2026 runtime 475 added an opt-in “Null Consensus” mode in which the highest-stake validator sets the weights, which critics say raises centralization concerns. Mining requires real hardware, typically GPUs; if you are curious about the hardware side, see my best GPU for local LLMs guide.
Dynamic TAO and Alpha Tokens
The biggest change in Bittensor’s history went live in February 2025: Dynamic TAO (dTAO).
- Every subnet got its own alpha token, paired with TAO in an automated market maker pool.
- Staking TAO into a subnet is effectively a swap into that subnet’s alpha.
- More TAO flowing into a subnet increases its share of new emissions; withdrawals reduce it.
- Since November 2025 (“Taoflow”), emissions follow net TAO inflows rather than token prices.
- The May 13, 2026 emissions refactor concentrated rewards among roughly the top 30 subnets.
The idea is to let the market decide which subnets deserve rewards, replacing a small group of root validators who used to vote. The criticism is that market demand for an alpha token does not always reflect real customer usage of the subnet.
TAO Supply, Emissions and Halving
| Item | Value |
|---|---|
| Maximum supply | 21,000,000 TAO |
| Circulating supply | About 11.68 million TAO (November 2026) |
| Emission before first halving | 1 TAO per block, about 7,200 per day |
| First halving | December 2025, at 10.5 million issued |
| Emission now | 0.5 TAO per block, about 3,600 per day |
| Halving rule | Triggered by issuance thresholds, each time issuance crosses the midpoint of the remaining supply |
The 21 million cap and halvings are deliberately modelled on Bitcoin. Price snapshots from CoinStats on November 1, 2026 put TAO at about $233 to $236 with a market cap near $2.7 billion (rank around #47 to #48). TAO News reported a price of $259 on October 8 after a 14% weekly fall. Prices move fast; check a live source.
Notable Subnets and Real Revenue
| Subnet | What it does |
|---|---|
| Chutes (SN64) | Serverless AI inference marketplace; the top revenue-generating subnet in Q1 2026 |
| Targon (SN4) | Confidential enterprise compute, built by Manifold Labs |
| Apex (SN1) | LLM inference and agent workflows, one of the original subnets |
| Templar (SN3) | Decentralized pre-training; its team, Covenant AI, left in April 2026 after training Covenant-72B |
| Others listed on Kraken | Hippius, Lium, Score, Ridges AI, Vanta |
Subnets that sell real products generated an estimated $43 million in AI service revenue in Q1 2026, led by Chutes and Targon. That is meaningful for a crypto network, but small next to centralized inference companies: fal.ai alone reports more than $800 million in annualized revenue (see my fal.ai guide).
2026 Timeline: Governance Crisis and Upgrades
- February 13, 2026: co-founder Jacob Steeves (“Const”) said he stepped down as CEO of the Opentensor Foundation; COO Ala Shaabana also stepped aside.
- April 9, 2026: Covenant AI founder Sam Dare announced his team was leaving and alleged that Steeves effectively controlled the three-signer “triumvirate” that approves protocol upgrades. Steeves denied being able to suspend emissions. TAO fell about 15% in roughly two hours, from around $338 to about $285. These allegations come from one party and are disputed.
- May 2026: “Robin” upgrade doubled subnet slots to 256; emissions refactor on May 13.
- June 22, 2026: Opentensor published an 18-month roadmap targeting full decentralization of governance by around December 2027.
- August 4, 2026: “Root Reborn” changed how root stakers earn yield.
- October 2026: runtime 475 (Null Consensus option), owner control of registration difficulty, and the eTAO upgrade letting subnet owners set emission splits within limits.
As of October 2026, the three-person signer group still controls protocol upgrades, so the decentralization promise is a roadmap, not a done deal.
Grayscale ETF Filing and Exchange Listings
- Grayscale formed a Bittensor Trust in April 2024 and filed an S-1 on December 30, 2025 to convert it into a spot ETF under the ticker GTAO on NYSE Arca. An October 23, 2026 amendment names BitGo Bank & Trust as main custodian and Coinbase Custody as second custodian. A filing is not an approval.
- Kraken completed a native dTAO integration and announced listings of subnet alpha tokens.
Risks You Should Understand
- Governance concentration. Upgrades still depend on a small signer group, and a Grayscale filing noted that most Subtensor nodes were run by the Opentensor Foundation as of December 2024.
- Market-driven rewards. Emissions follow money flows into alpha pools, which can reward hype over real usage.
- Security. In July 2024 a malicious release of the
bittensorPython package (version 6.12.2) led to wallet drains of about $8 million. Only install software from official sources and never paste seed phrases into tools. - Alpha token risk. Subnet tokens are thinner markets than TAO and can move far more sharply.
- Volatility. TAO trades about 69% below its peak; double-digit weekly swings are common.
- Regulation. Crypto rules differ by country and can change quickly.
How to Evaluate a Subnet
- Does it sell something real? Look for paying customers or a public API, not just emissions.
- Is the scoring hard to game? Read how validators test miners.
- Who runs it? Check the owner’s track record and how much of the emissions they take.
- How liquid is the alpha pool? Thin pools mean large price impact.
- Compare with centralized options. If an API like OpenRouter or fal does the same job cheaper and more reliably, the subnet needs another edge. My OpenRouter guide is a useful benchmark.
How to Check a Subnet’s Stats on Taostats
You do not need to install anything to see what a subnet is doing. Block explorers such as Taostats read the chain and show each subnet’s numbers in a normal browser. This is the routine I would follow before forming any opinion about a subnet. It is research, not a buying guide.
- Open the subnet list. Go to taostats.io and open the subnets section. Each subnet is listed by its number (netuid) and name, for example SN64 for Chutes.
- Read the summary at the top of the subnet page. Taostats shows the GitHub link for the miner and validator code, when and by which hotkey the subnet was registered, its share of total emissions, the amount of TAO recycled (historical and last 24 hours), the current registration cost for a neuron, and how many miners and validators are registered.
- Look at the emissions share over weeks, not just today. Since Taoflow, emissions follow net TAO flows into the subnet, so a sudden jump can reflect money moving in rather than more customers.
- Open the GitHub repository. Look for recent commits, a readable README and a clear explanation of how validators score miners. An abandoned repository is a warning sign.
- Check the miner and validator breakdown. If a handful of hotkeys collect most of the rewards, or very few validators are active, the scoring may be easy to game or poorly policed.
- Note the registration cost. This is what a new miner pays to join. It floats with demand, so a very high cost tells you competition for slots is intense.
- Cross-check outside the explorer. Search for the subnet’s product, public API or paying customers. On-chain numbers show incentives; only off-chain evidence shows real demand.
Explorer layouts change often, so treat these labels as a guide rather than exact button names.
Troubleshooting: Common Problems and Fixes
If you go beyond reading the explorer and try running a miner or validator, these are the problems people hit most often. The error names and defaults below come from the official Bittensor documentation; commands change between releases, so always check the current docs at bittensor.com before running anything.
| Problem | Likely cause | Fix |
|---|---|---|
| Registration cost much higher than expected | Each subnet has a floating registration cost that rises with demand, and the exact charge is only known when the transaction runs | Preview the fee first with a dry run (the current docs show btcli tx burned-register --netuid <id> --dry-run). The cost is burned or recycled, so assume you will not get it back; on some subnets part of it is locked as miner collateral that is only released through earned emissions. |
Registration fails with SubNetRegistrationDisabled |
The subnet owner has switched registration off | Nothing on your side will fix it. Wait for it to reopen or choose another subnet. |
| New miner deregistered after a while | On a full subnet, each new registration evicts the lowest-emission neuron that is outside its immunity period | Make sure your miner scores well before immunity ends, and watch your UID and emissions on the explorer. |
Weights rejected with IncorrectWeightVersionKey |
Your validator code is older than the version the subnet requires | Update to the subnet’s latest release from its official repository. |
NotEnoughStakeToSetWeights |
Stake on the hotkey is below the chain’s stake threshold (roughly 1,000 TAO-equivalent on mainnet, per the docs) | Validating needs substantial stake. Most newcomers are better off running a miner on a test network or just observing. |
| No validator permit | Permits go only to the top K neurons by stake weight (K defaults to 128) | Check permit status in the metagraph. Losing a permit stops dividends and deletes bonds, though it does not deregister you. |
| Weight submissions rate limited | Each UID must wait a set number of blocks between submissions (100 blocks by default) | Wait the number of blocks the preflight reports. Keep the validator running, because one that stops setting weights is treated as inactive. |
| Weights on the explorer look out of date | The subnet uses commit-reveal, so weights are timelocked and revealed later | This is expected. Changes appear after the reveal period. |
| Wallet commands fail on Windows | Validating is not supported on Windows | Use WSL 2 for wallet operations, as the docs recommend. |
| Funds missing from a wallet | Unofficial or tampered software, or a leaked seed phrase | Install only from official sources, keep the coldkey offline, and never paste a seed phrase into a website, bot or script. |
A practical tip: try new setups on a local or test network first, so a mistake costs time rather than real TAO.
Frequently Asked Questions
What is Bittensor in simple terms?
A blockchain that pays people in TAO for doing useful AI work, organized into competing subnets where validators grade the work.
What is the max supply of TAO?
21 million TAO. About 11.68 million were in circulation in November 2026.
When was the Bittensor halving?
The first halving happened in December 2025, cutting daily emissions from about 7,200 to 3,600 TAO.
What is dTAO?
Dynamic TAO, live since February 2025. Each subnet has its own alpha token, and staking TAO into a subnet directs emissions to it.
Is Bittensor a good investment?
I cannot tell you that, and this article is not investment advice. It is a high-risk, volatile asset with real usage but unresolved governance questions.
Can I use Taostats without a wallet?
Yes. Taostats is a block explorer, so you can browse subnets, emissions, registration costs and validator data in a browser without connecting a wallet or holding any TAO.
How much does it cost to register a miner on a subnet?
There is no fixed price. Each subnet has a floating registration cost that changes with demand, shown on its explorer page. The TAO paid is burned or recycled, so treat it as a sunk cost and preview it with a dry run first.
