Bittensor (TAO) Explained: How the AI Marketplace Works and Its Risks

Many investors file TAO alongside dozens of other coins with “AI” in the name. A new Bullrunners video, hosted by George Lee, argues that is a mistake. Bittensor, the network behind TAO, is designed as a marketplace where computers anywhere in the world compete to produce useful AI and get paid in tokens for it.

That design is different from most AI-branded projects, and so are the risks. The video walks through where Bittensor came from, how its subnets and tokens work, and the very public dispute in April 2026 that sent TAO sharply lower within hours. If you hold TAO, or someone has told you to buy it, both sides matter.

What is Bittensor, in plain terms?

Bittensor is an open network that rewards computers for producing AI work, rather than a company building one AI product. Building frontier AI costs billions, and the video notes that only a handful of large companies and governments can afford it. Bittensor’s bet is that an open, competitive network can offer another path.

In the network, two kinds of participants do different jobs:

• Miners do the work, such as answering queries, training models or supplying computing power.
• Validators score that work.

The network then pays out rewards based on those scores.

Who founded Bittensor?

Bittensor was founded by Jacob Steeves, known online as Const. In a talk shown in the video, he says he studied mathematics and computer science, worked for DARPA building neuromorphic chips (computer chips modeled on how the brain works), became interested in Bitcoin in 2016, and later worked at Google Brain, which he describes as the leading AI research lab of its time. He then left to build Bittensor.

That mix of Bitcoin and AI is the core idea. In his words, the question was how to connect “the most powerful computer in the world” to “the most important computational problem” of the century, which he sees as artificial intelligence.

How does Bittensor copy Bitcoin?

Bittensor borrows Bitcoin’s basic model: post a reward and let anyone compete for it. Bitcoin doesn’t hire anyone. Anyone can plug in a machine, compete to solve a puzzle, and the winners are paid in new Bitcoin.

Const’s question was what would happen if that open competition were pointed at AI. Instead of rewarding machines for solving a puzzle whose only job is securing the network, Bittensor rewards them for producing something useful.

The hard part is measurement. Checking a Bitcoin puzzle is simple because the answer is right or wrong. Judging whether one AI answer is better than another is far harder, which is why validators and scoring sit at the center of the design.

Bittensor also copied Bitcoin’s money rules, according to the video:

• Supply cap: there will only ever be 21 million TAO.
• Fair launch: it launched in 2021. Grayscale Research, which runs its own Bittensor fund, says this meant no venture capital pre-sales and no insider allocations. Critics point out that a small group of insiders and early participants mined much of the supply in the early years.
• Halving: in December 2025, TAO had its first halving, which cut new daily issuance in half from about 7,200 TAO a day.

What are Bittensor subnets?

Subnets are where the AI work actually happens. Each subnet is its own competition set up to produce one kind of thing.

Const explains it with a simple example. To join a subnet, you burn a small amount of token, like an entry fee, to show you are willing to play. Your registered computer then does work: for example, running an AI model that answers questions sent by users, such as “the capital of Texas is Austin.”

According to the video, by mid-2026 there were around 128 subnets, the most the network allowed at that point. They cover AI models, computing power, data storage and more. Some outside companies have started using them. In April 2026, Erik Voorhees, the founder of ShapeShift, said that the private AI service Venice had started working with Bittensor subnets and that one of its new models was trained on a subnet.

What changed with Dynamic TAO?

Dynamic TAO gave every subnet its own token and handed the market control over rewards. The upgrade arrived in February 2025. Each subnet now has an alpha token that trades against TAO.

From then on, the market rather than a small group of validators decides how much TAO each subnet receives. Since late 2025, per the video, that has been based on how much TAO flows into each subnet.

Const’s own description of alpha tokens is notably candid. He says a subnet token “only has value if at the end of the day, people are going to want to buy that token to get access to the computers on this network.” The video calls that the right question for anyone holding a subnet token or TAO itself.

This also changed what holding TAO means. Holders can:

1. Stake on the root network and keep their position in TAO, or
2. Stake into a specific subnet, which swaps TAO into that subnet’s alpha token and backs it directly.

According to FalconX, around a fifth of circulating TAO was staked into subnets by May 2026. The second option is a real way to take part, but it ties you to that subnet’s fate. If the team behind it leaves, its token can fall quickly.

What happened with Covenant AI in April 2026?

In April 2026, a team called Covenant AI, which ran three well-known subnets including one called Templar, announced it was leaving Bittensor. Its founder accused Const of having too much control over a network that calls itself decentralized.

Covenant argued that when a single actor can suspend a subnet’s rewards, override its owners and use token sales to pressure teams, that amounts to “centralized control with decentralized branding.” Covenant then sold its subnet tokens. According to FalconX, some of those tokens fell by as much as 40% almost immediately, and TAO fell about a quarter within hours.

Const disputed the claims. According to reports cited in the video, he said he had no power to suspend a subnet’s rewards on his own and had sold less than 1% of the tokens he had put into Covenant’s projects. FalconX also reported that Covenant’s departure message appeared to have been written with AI help, which it said could cast doubt on the motives.

Some context from the video:

• In February 2026, Const stepped down as chief executive of the Opentensor Foundation, and a fellow co-founder left an executive role, a move billed as a step toward decentralization. Const said he would still be on the same calls and still writing code.
• In May 2026, Bittensor sped up a governance change it calls “conviction.” It locks up subnet owners’ rewards and makes any exit visible on-chain well before most of it can be sold.
• In July 2024, a malicious version of Bittensor’s official software was uploaded to the main public store for Python code, and attackers used it to steal around $8 million worth of TAO from users. The Opentensor Foundation put the network into safe mode. It said the blockchain itself was not compromised, but critics noted that a single foundation could pause the network.

The bull case and the bear case

The video presents both sides, and the core question remains open.

	Bull case	Bear case
Access	Anyone with skill and hardware can compete from anywhere, without permission	Bad actors can join too, so scoring must work very hard
Funding	Grayscale Research likens it to a Y Combinator for decentralized AI, with TAO funding subnets	Paying people in new tokens is not the same as paying customers
Demand	Outside users such as Venice are using subnets	Some subnets may chase rewards more than real users
Control	Leadership changes and conviction aim to spread power	The Covenant exit and 2024 pause raise questions about who really holds power

Const himself has acknowledged the scale of the challenge. In April 2026 he wrote that “we can’t beat Anthropic at their own game” and that new systems for training models are needed. He is building a subnet of his own aimed at measuring and rewarding intelligence itself, which he admits is much harder to measure than compute or storage.

Three lessons for any coin you hold

The video closes with a framework that applies well beyond TAO.

1. The name isn’t the product. Many coins put AI in their pitch. Ask what the network actually produces and who uses it. With Bittensor, you can point to subnets and outside users like Venice.
2. Follow the demand, not the rewards. As Const said, a subnet token only has value if people want what the network produces. Rewards paid in newly created tokens are not the same as customers paying for a service.
3. Decentralization is a claim, not a fact. The Covenant exit and the 2024 pause both raise the same question: who can step in, and how much power does one person or foundation really have?

The video also suggests a one-sentence test. For any coin, try to explain who pays for what it produces. If the only answer is “people buying the coin itself,” that is worth knowing before you buy. For Bittensor, the video’s honest answer is a mix: some real outside users and a lot of reward funded by new TAO.

What to watch next

The key signal, according to the video, is how that mix shifts over time. Watch whether more outside companies like Venice pay to use subnets, whether the conviction rules reduce sudden team exits, and how much influence Const and the Opentensor Foundation keep after the leadership changes. If you stake into a subnet, check who runs it and what it actually produces, since its alpha token carries that team’s risk, not just TAO’s.