If you are here because somebody mentioned TAO and you nodded along without really knowing what they meant, this is the page for you. Nothing below assumes you have owned a cryptocurrency, run a computer program, or read a technical paper. It does assume you are willing to sit with an idea for a while, because Bittensor is genuinely unusual, and every short explanation of it fails in the same way. They tell you it is decentralised AI, and you come away no wiser than you started.
So we will build the picture slowly, one piece at a time, each with a story attached so it sticks. By the end you should understand this machine well enough to notice when somebody is talking nonsense about it, including when that somebody is enthusiastic and owns a great deal of the token.
One promise first. Nothing here is advice about money. Where something is risky, this page says so plainly, and there is a long section near the end on how people actually lose money here, because a guide that skips that part is a sales pitch.
The problem, before any of the jargon
Think about how artificial intelligence reaches you at the moment. You open an app, you type a question, an answer comes back. Behind that app sits an enormous model that cost hundreds of millions to build, running on specialised computers in a warehouse, owned and operated by one of a very small number of companies. Perhaps five or six of them matter globally. That is the whole of it.
There is a version of this that sounds perfectly fine. Building these things is hard and expensive, so of course only large organisations can do it. But hold the picture a moment and notice what it means in practice.
It means a handful of companies decide what these systems will and will not say. Not through malice, usually, but because somebody has to write the rules and it turns out to be them. It means they decide who is allowed to build on top, and at what price, and they can change that price or withdraw access on a Tuesday afternoon if it suits them. It means the enormous value these systems create flows to whoever owns the servers, and not to the people whose work trained them or the people who improve them.
Imagine if electricity worked this way. Five companies generate all of it, and they also decide which of your appliances are allowed to run. Most days you would not notice. On the day you did notice, there would be nothing at all you could do.
That is the problem Bittensor is aimed at. The idea is to build an open market for machine intelligence, where anybody with a computer can offer useful work, anybody can pay for that work, and nobody at all owns the marketplace. Not a company with a friendly mission statement about openness, but a system with no owner, running on rules anyone can read.
Whether such a system can actually compete with a well funded company is the open question of the entire project, and it would be dishonest to pretend that is settled. But you cannot judge the answer until you understand what is being attempted.
The hardest problem in the whole design
Suppose you wanted to build that open market yourself. You would very quickly hit one obstacle, and everything strange about Bittensor exists because of it.
You want to pay people for producing useful intelligence. Fine. Who decides what counts as useful?
In a normal marketplace this is easy, because a customer decides. You order a meal, you eat it, you either come back or you do not. But an open network cannot rely on a customer showing up for every piece of work, and it cannot rely on a manager either, because the whole point is that there is no manager. It has to hand out money automatically, thousands of times a day, to strangers whose work it cannot inspect.
So the network needs judges. And the moment you introduce judges you have created something far more valuable than the work itself, because a judge who decides who gets paid can simply decide to pay themselves. If I am a judge and I also quietly run three of the competitors, I score my own entries top marks and go home rich.
Every design choice in Bittensor, including the ones that seem baroque and pointless, is an answer to one question: how do you judge the judges? Keep that in your head as you read. When you meet a mechanism later that seems needlessly complicated, it is almost always there to stop a judge cheating, or a competitor gaming a judge, or the two of them agreeing to split the proceeds.
What TAO actually is
TAO is the money the network pays out in. That is the plainest description available and it is the correct one.
Every twelve seconds or so, the network creates a small amount of brand new TAO out of nothing and gives it to the people who did useful work in that window. Nobody has to fund this. There is no company writing cheques. The rules of the software simply say that new TAO comes into existence on a schedule and goes to particular participants according to particular measurements, and every computer running the network agrees on who got what.
There will only ever be twenty one million TAO. That number is fixed in the software, and the rate of creation halves periodically, so the flow slows over time and eventually stops. If that sounds familiar, it is deliberately borrowed from Bitcoin. But the resemblance hides the more interesting difference.
What Bitcoin participants do to earn their coins is guess enormous numbers, very fast, over and over. The guessing has no value outside the system. It is a deliberately wasteful ritual whose purpose is to make attacking Bitcoin expensive. It works, it is elegant, and the work itself is useless by design.
Bittensor tries to pay for work that is useful outside the system. Somebody rents out a graphics card that runs a real customer's job. Somebody hosts a model that answers real questions. Somebody turns live footage of a football match into structured data a broadcaster can use. The network measures the quality of that work and pays accordingly. Whether it succeeds at measuring quality honestly is the whole argument, but the ambition is different in kind.
Now the comparison people get wrong most often. TAO is not a share in a company. When you buy a share you get a legal claim: a slice of the assets, a vote at the annual meeting, a right to a dividend if one is declared, and directors who owe you certain duties. If they misbehave you can sue somebody, because there is a somebody.
TAO gives you none of that. There is no company, no board, and nobody owes you anything. Holding TAO does not entitle you to a share of any revenue anyone earns. A foundation helps develop the software, and some groups hold enough to influence upgrades, but none of them has a duty to you.
So what is it? The most useful picture is a currency rather than a share. It is the only money accepted inside this particular economy. If you want to run a subnet you pay in TAO. If you want to compete as a miner you pay in TAO. If you want a say in how rewards are steered you must hold TAO and lock it up. Its value comes from people wanting to do those things, and from people betting that more people will want to later. Not from a contract, not from a claim, not from a promise. That is a genuinely weaker foundation than a share certificate, and it is worth being clear eyed about rather than glossing over.
A subnet is a competition
Here is the single most useful picture in this whole guide.
Imagine a county fair with dozens of tents. Each tent runs its own competition. One is a baking contest. Another is a sheepdog trial. Another is a competition for the fastest hand at shearing. Each has its own rules, its own judges and its own prize table, and whoever wins the baking contest is entirely irrelevant to the sheepdog trial.
A subnet is one of those tents. It is a competition with a specific task and a specific way of scoring how well that task was done.
That is why there are many rather than one. There is no single way to score intelligence. Judging whether a translation is good is nothing like judging whether a weather forecast was accurate, which is nothing like judging whether a rented graphics card actually stayed online all night. So instead of trying to write one impossible universal scoring rule, Bittensor lets anybody set up a new tent with their own rules, pay the registration fee, and see whether people show up to compete.
The real ones are more concrete than any abstract description. Subnet 51, called Lium, runs a marketplace for renting graphics cards, the specialised chips that AI runs on. Competitors offer up their machines, the network checks the machines are real and working, and customers rent them by the hour without ever meeting the owner. Subnet 64, Chutes, lets people run AI applications on demand without owning a server, the way you might take a taxi rather than buy a car. Subnet 53, engy, serves large open AI models and attaches a cryptographic fingerprint to every answer, so you can check the answer really came from the model you paid for rather than from something cheaper the operator swapped in quietly. Subnet 1, Apex, takes a hard question, puts it to a global crowd of competitors, and routes the best answer back. Subnet 4, Targon, runs AI inside sealed hardware so that even the operator of the machine cannot read what passed through it. Subnet 44, Score, turns live sports footage into structured data in real time.
Six tents, six completely different competitions, one fairground. There are around a hundred and thirty of them at the time of writing, and the quality varies enormously, which is a theme we will come back to.
Anybody can start a subnet. You pay a registration cost in TAO, which is burned rather than paid to anyone, you write the rules of your competition in software, and you open the tent. The market then decides, quite brutally, whether your competition was worth running.
Miners are the competitors
The word miner is an unfortunate inheritance. It makes people picture the Bitcoin thing, warehouses of machines guessing numbers, and that is not what is happening here at all.
A miner on Bittensor is a competitor in one of those tents. A person, or a small team, who has put their name down for a particular competition and is now trying to win it.
What do they actually do all day? They run a computer, or a rack of them, connected to the internet and to the subnet they registered with. That computer sits waiting. Every so often it receives a task from a judge: here is a question, answer it. Here is an image, describe it. Prove that the graphics card you claim to have is real and currently idle. Here is a clip of a match, tell me where the ball is. The computer produces an answer, sends it back, and gets scored on it.
The human behind that computer does what any competitor does. They tune their setup, try a better model, buy faster hardware, watch a dashboard showing their score against everyone else, and grumble when they slip down the table. Some are hobbyists with a gaming machine in a spare room. Some are serious operations with racks in a data centre and an electricity bill that would frighten you.
The crucial difference from Bitcoin is that this is not a lottery. Two miners with identical hardware can earn wildly different amounts because one is simply better at the task. That is the design intent: money should flow towards whoever is genuinely best at the job.
Getting in costs money. To take one of the limited slots in a subnet you burn TAO, meaning you pay it and it vanishes rather than going to anybody. Slots are finite, so if the tent is full and you want in, the worst performing competitor is pushed out to make room. New arrivals usually get a short grace period, an immunity window, so they are not thrown out before they have had a fair chance.
That combination, pay to enter and get pushed out if you are last, is what keeps a subnet sharp. It is also how people lose money without ever buying a token: they pay to enter a competition they are not good enough to win, earn very little, and get pushed out anyway.
Validators are the judges
If miners are the competitors, validators are the judges, and very nearly everything hangs on them.
A validator runs a computer too, but its job is the opposite one. It invents tasks, sends them to the miners, collects the answers and scores them. Then, at regular intervals, it publishes a scorecard: a list saying this miner deserves this share, that one deserves that share. In the jargon this is called setting weights, and it is the only thing the network itself ever sees. The network does not read the answers. It has no opinion about whether a translation was good. It only reads the scorecards.
Which brings us back to the hardest problem. If the scorecards are the only thing that matters, a dishonest judge is worth more than any amount of good work. So how do you stop a judge scoring their own friends?
The answer has two halves, and both are worth understanding.
The first is that judges must have money at stake. You cannot simply announce that you are a judge. Your scorecard counts in proportion to how much TAO is locked behind you. A judge backed by a large amount has a loud voice, a judge backed by very little is barely audible. Most of that backing does not come from the judge's own pocket. Ordinary holders delegate their TAO to a validator they trust, which means lending weight to that judge's voice without handing over the coins, and in return they receive a share of what that judge earns. The judge keeps a cut. So a validator is a little like a fund manager who has persuaded a lot of people to stand behind them, and who now has a great deal to lose by being caught cheating.
The second half is more interesting. The network compares the judges to each other. Rather than simply adding up all the scorecards, it does something closer to a trimmed average, the way Olympic diving throws away the highest and lowest marks. A judge who broadly agrees with the other judges is treated as reliable and earns well. A judge whose scorecard is wildly out of line has the odd parts of it discounted, and earns less.
So the cheating judge who gives their own miner top marks finds those marks trimmed away by everyone else's disagreement, and finds they personally earn less for having tried. Cheating stops paying. That mechanism, and the machinery around it deciding how fast a judge's backing of a particular miner is allowed to build up, is the beating heart of the system.
Be honest about the cost, though, because there is one. A system that rewards agreement punishes the judge who is right before anybody else. If one validator spots that a miner is quietly producing far better work than the crowd has noticed, they are penalised for saying so until the others catch up. Bittensor is therefore, by construction, somewhat conservative. It converges on the consensus view of quality rather than the correct one, and those are not always the same thing. Recent protocol work added tuning knobs for exactly how fast a judge's confidence in a miner may build, which is a quiet admission that this balance is an ongoing problem rather than a solved one.
Where new TAO comes from, and who catches it
Now the money. This part is simpler than people expect.
Roughly every twelve seconds the network creates new TAO. Picture a river with a fixed flow. The network decides how much runs down each irrigation channel, one channel per subnet, and then how the water in each channel is shared among the people standing in it.
Within one subnet the split is fixed by the rules. Something close to forty one percent goes to the miners, divided according to those scorecards. Another forty one percent or so goes to the validators and, importantly, to the ordinary holders who delegated TAO to them, which is how a person with no computer and no technical skill ends up receiving a share. The remaining eighteen percent goes to whoever created and runs the subnet, which is the reward for building a competition worth entering and the reason anybody bothers.
The total flow shrinks over time. Like Bitcoin, the rate halves periodically. Unlike Bitcoin, the halvings are tied not to dates but to how much TAO has actually been issued, and here is a subtlety that catches people out. The TAO burned by miners registering and by people launching subnets is not destroyed forever, it returns to the pot of unissued supply. More burning means longer until the next halving. So the schedule breathes with how busy the network is, which makes confident predictions about supply less reliable than the twenty one million headline suggests.
The thing to hold on to is that this new TAO is not payment from customers. It is newly created money, distributed on a schedule, whether or not a single outside customer paid for anything that day. Some subnets do earn real revenue from real users. Many do not. That gap, between money created by the protocol and money earned from the world, is the most important thing to watch in this entire ecosystem, and we will come back to it.
Alpha, or why every tent prints its own tickets
Here is where most explanations lose people. It is not hard, it just needs the right picture.
Every subnet has its own token, called its alpha token. Subnet 64 has one, subnet 51 has one, and they are separate currencies that are not interchangeable.
Back to the fairground. Each tent sells its own tickets. You cannot use a baking contest ticket to get into the sheepdog trial. If you want into a particular tent you go to that tent's booth and swap ordinary money for that tent's tickets, and the booth sets the price automatically based on how many tickets it has left and how much money has come in.
That booth is real, and it drives everything else. Each subnet has a pool holding two things: some TAO and some of its own alpha. The price of that alpha is simply the ratio between the two. When you stake TAO into a subnet you hand TAO to that pool and receive alpha back, which pushes the price up, because there is now more TAO and less alpha in the pool. When you leave, you hand back alpha and receive TAO, which pushes the price down. There is no order book, no buyer to find, no counterparty. The pool always trades with you, at a price its own contents determine.
Why build it this way rather than paying everyone in TAO? Because it lets the market price each subnet separately, continuously, without anybody voting. If you think subnet 53 is doing extraordinary work you can back it directly, and the act of backing it visibly moves that subnet's price. If you think it is going nowhere you leave, and that moves the price the other way. Every subnet carries a live, public number saying what people currently think it is worth, updated every few seconds.
The practical consequence is that when you back a subnet you are no longer holding TAO. You hold that subnet's alpha, and the emissions you earn arrive as more alpha, not as TAO. To turn any of it back into TAO you go through the pool, at whatever price the pool offers at that moment.
This matters enormously for risk, so it deserves saying flatly. Many of these pools are small. A modest sale can move a small subnet's alpha price a long way, and a large one can move it dramatically, so the amount you can actually get back is not simply your holding multiplied by the quoted price. Everyone can be told their alpha is worth a certain amount and yet not everyone can leave at that amount, because the money to pay them is not sitting in the pool. This is not a scandal somebody is hiding. It is arithmetic, and it is how these pools work everywhere they exist. But it is the most common way people here meet a number smaller than the one they expected.
The moving bar
So we know how money is split inside one subnet. What decides how much a subnet gets in the first place?
Its price, relative to every other subnet's price. That is essentially the whole rule. Add up what the market says all the subnets are worth, work out what fraction of that total is yours, and that is roughly your share of the river. A smoothed average is used rather than the instant price, so a brief spike does not immediately become a payday, but the principle holds.
Back to the fair one last time. At the end of the day there is a prize fund, and it goes to each tent in proportion to how much people paid for that tent's tickets. The tent everybody wanted into gets a big share. The tent nobody visited gets almost nothing.
Now the part worth dwelling on, because it explains a great deal of the behaviour you will see. There is a bar you must clear to be worth running, and that bar moves without you doing anything at all.
Running a subnet costs real money. Servers, engineers, bandwidth, time. Whether your share of the river covers those costs depends not on your own price but on your price compared to everyone else's. If ten new subnets launch tomorrow and each attracts genuine backing, your slice of the same river gets thinner even though you did nothing wrong and your work did not get worse. If a rival's price doubles, yours effectively falls. You can be improving steadily and still sliding down the table, because the bar rose underneath you.
There are mechanisms sitting on top of that. Recent protocol work restored a rule where a subnet withholding part of what it is meant to pay its miners has its network wide share scaled down to match, closing a loophole where an owner could quietly keep more without it costing them anything. And there is a hard floor: subnet slots are limited, so when the fairground is full and somebody wants to pitch a new tent, the weakest existing tent is taken down.
This design has an obvious and much discussed side effect. If your share of the money depends on your token's price, then buying your own token raises your income. Owners can and do support the price of their own alpha, and it is not always possible to tell from outside whether a rising price reflects genuine outside interest or the owner holding it up. The system is reflexive: price drives emissions, and emissions can be used to buy price. Anyone claiming they can reliably tell real demand from manufactured demand in this market is overstating their case.
dTAO, and the day the committee was abolished
It did not always work this way, and the change was the most consequential thing that has happened to Bittensor.
Before February 2025, the split of the river between subnets was decided by a vote. A group of large validators on what was called the root network published their opinions on which subnets deserved what, and those opinions were combined into the allocation. Human judgment, sitting at the very top of the system.
You can guess how that went. The people casting those votes were few, well known and heavily lobbied. Good projects with no connections struggled. Well connected projects with mediocre output did fine. It was, in the end, politics, inside a system whose entire pitch was that it did not need any.
dTAO replaced the vote with the market. The alpha tokens arrived, the pools arrived, and allocation became a price rather than an opinion. Instead of persuading a committee, a subnet now has to persuade anybody at all with money to back it.
What improved is not in doubt. Anybody can back a subnet, so the gate is open to everyone rather than to a dozen well connected parties. Feedback is immediate rather than arriving whenever the committee next thinks about you. Capital can move to a promising new subnet within hours of it appearing.
What got worse is argued about, and honest people disagree. The market is not a wise judge of technical quality. It is a crowd, and crowds respond to a good story, a lively community and an active social media presence at least as readily as to working software. There are subnets doing quietly excellent engineering on a modest allocation, and subnets with a thin product and an excellent marketing operation doing rather well. Small subnets can now die very quickly, which is efficient if they deserved to die and wasteful if they simply needed another six months. And an owner with capital can influence their own allocation in a way no amount of lobbying could achieve before.
The fair summary is that dTAO swapped one flawed judge for another. A committee is slow, political and hard to enter, but it can in principle recognise good work that has not yet found an audience. A market is fast, open and impossible to lobby, but it will happily pay more for the better story. Which is the lesser evil is a real argument with real people on both sides, and you should be suspicious of anybody who presents it as obvious.
Who can change the rules
A word about governance, because it is easy to think of this as a machine that runs itself.
The rules are software, the software gets upgraded, and upgrades arrive roughly monthly. They are signed off by a small group holding the keys, with a second group able to object, so it is neither one person's whim nor a public vote. A recent upgrade changed the conditions under which control of a subnet can be taken from its owner. The old rule added up all the locked backing and set the bar at ten percent, which could be met almost by accident when unrelated parties happened to lock funds at once. The new rule requires one single holder behind more than eighteen percent of the alpha that is eligible to count, which is far harder and far more deliberate. Whether that is a sensible tightening or a barrier protecting incumbent owners depends on where you sit, and both readings have been argued in public.
The point for a newcomer is not the detail. It is that the rules of this economy are genuinely changeable, that they change often, and that a small number of people hold the keys. This is more decentralised than a company and considerably less decentralised than the marketing suggests. Both halves of that sentence are true.
What can go wrong
This is the section that matters most, and the one enthusiastic explanations skip.
The first and most ordinary risk is price. TAO moves violently, in both directions, and it has had falls that would empty most people's stomachs. Nothing in the design prevents that and nothing guarantees a recovery.
The second is that backing a subnet is riskier still, because you take two risks stacked on each other. Your alpha can fall against TAO while TAO is falling against ordinary money, and the two multiply rather than average. People are regularly surprised by this. Alpha is the higher risk instrument here by a wide margin, not the safer one, and a subnet that looks cheap can and does go to very nearly nothing.
The third is that the pool you bought through is the pool you must sell through, and it may be far shallower than you assumed. Look at the size of the pool, not only the price, before deciding what a holding is worth.
The fourth is dilution. New TAO and new alpha are created continuously and given to participants, many of whom sell some to cover their electricity bills. That is a permanent stream of supply arriving whether or not buyers turn up. It is the cost of paying for the work, it is not hidden, and it is real.
The fifth applies if you delegate to a validator. You are trusting somebody. They take a cut, which they can change. They can run their machine badly, go offline, or lose interest. Your funds are not usually at risk of being stolen this way, since the design does not hand them over, but your returns depend entirely on somebody else's competence.
The sixth applies if you compete as a miner. You burn TAO to enter and you may earn nothing at all. Plenty of people have paid to join a competition, found the leaders far ahead, earned very little and been pushed out anyway. That is not a malfunction, that is the competition working.
The seventh is that this is young software with bugs, some of which cost people money. In one subnet, miners paid their registration fee and submitted a sealed, time locked entry, the sort that stays encrypted until a set moment so nobody can peek early. When the unsealing step failed, the system threw the entry away rather than keeping it, so a valid submission simply vanished and the validators saw nothing. Miners paid, did everything right, and got nothing. That was recently found and repaired, which tells you both that problems get fixed and that more are waiting to be found.
The eighth is that nothing here is insured, refunded or reversible. There is no support line, no chargeback and no regulator standing behind it. If you send funds to the wrong place, or lose the secret phrase controlling your wallet, they are gone permanently. Not difficult to recover. Gone.
And the ninth, less a risk than an unanswered question, is whether enough of this economy is supported by real outside revenue rather than newly created tokens. Some subnets have genuine paying customers. Many are largely funded by the protocol's own emissions and by people betting those emissions will continue. That is not fraud, it is how a new network gets started, and it is also how a network fails to ever go anywhere. From outside, in the early years, the two look remarkably similar.
What a failing subnet looks like from the outside
Since subnets fail regularly, it is worth knowing the shape of it, because it is rarely announced.
It starts quietly. The code repository stops changing, at first for a week and then for a month. The team's posts become vaguer, more about vision and roadmap and less about anything actually shipped. Questions in the community chat go unanswered, then the chat itself goes quiet, and the loudest remaining voices are people asking about the price.
Underneath, the mechanics grind on. The alpha price drifts down, so the subnet's share of the river shrinks, so miners find their earnings no longer cover their electricity and start switching off. Fewer competitors means worse output, which means less reason for anybody to back it, which lowers the price again. It is a spiral, and once properly underway it is rarely reversed.
Eventually the tent is taken down to make room for a new one, and the alpha becomes near enough worthless. Nobody sends a letter. There is no announcement. It simply stops.
You can watch all of this in public, for free, without owning anything at all. Anybody who tells you it is impossible to tell a healthy subnet from a dying one has not spent an evening reading commit histories and chat logs.
How a person would begin
This is not a recommendation to do any of it. It is a description of the order in which people who end up understanding this well tend to do things, and the honest observation is that the most useful steps involve no money at all.
The first is to read, for longer than feels necessary. Pick two or three subnets doing things you find genuinely interesting and follow them for a month. Read their public code repository even though you do not code, because you can see perfectly well whether anybody is working on it and whether the changes are substantial or cosmetic. Check whether what they promised in March had happened by August. This costs nothing and teaches more than any amount of price watching.
The second is to learn to read the public dashboards, of which there are several free ones. Look at how emissions are split, which subnets are rising and falling, how deep the pools are, until the numbers mean something rather than being decoration.
Only then does anyone sensibly touch money, and the mechanics are worth knowing regardless. You would need a wallet, which is a pair of keys rather than an account: a coldkey which controls your funds and should stay offline and backed up on paper, and a hotkey which does the day to day signing. The recovery phrase for the coldkey is the most important object in the entire exercise. Lose it and nothing can be done, by anybody, ever. Show it to somebody and they own everything you have.
From there, holding TAO and delegating it to a validator is the lower risk path, because you keep TAO and earn a share of what that judge earns. Staking into a specific subnet's alpha is the higher risk path, because you swap into a smaller, thinner, more volatile thing in exchange for closer exposure to one project's fortunes. Running a miner is a different activity again, closer to a small business with hardware costs and competitors than to holding anything.
The usual advice, general rather than specific to you and not financial advice, applies with unusual force here. Nothing you cannot afford to lose completely. Small amounts while learning, because the first mistake should be a cheap one. Understand the thing before you fund it, not after.
And the underrated option is to do none of it and simply watch for a while. The network will still be there. Watching a full cycle of a subnet rising and falling, with no money involved, teaches more than the same period spent anxious about a position.
The question worth carrying with you
Bittensor is a real, working attempt to build a market for machine intelligence that no company owns. Subnets exist, miners compete in them, judges score them, and money moves accordingly, every twelve seconds, without anybody's permission. That deserves respect.
But it pays for that work mostly by creating new money, and the test of whether it has succeeded is whether outsiders eventually pay for what it produces. So every time you read about a subnet, ask one question. Who outside this system is paying for this, and how much? If the answer is clear and growing, something real is being built. If the answer is that the token went up, you have learned something too, and you did not need to own a single token to find it out.