Open entry, priced survival

Bittensor is a blockchain that pays people to produce machine intelligence.

The part I want is not the intelligence claim. It is the way entry stays open while survival stays priced. Let's look at how the machine pays, and then take the idea off the blockchain.

I was pulling at a version of this back in 2022, when the network was small enough that it was mostly a thought experiment.

Price-governed permissionlessness

"Price-governed permissionlessness" - anyone may enter, nobody needs to be known, and the crowd's moving price helps decide what keeps getting paid.

The door

Bittensor has grown to more than a hundred separate subnets. Each one pays people to do a narrowly defined job with no manager.

Anyone can start a subnet by burning $TAO to claim a slot and writing two pieces of code: one that defines the desired work, and one that scores whether the work was good.

That is the open part. The door does not check a resume. It does not ask where the work came from. It asks for a slot, a job, and a scoring rule.

Alice and Bob

So here's Alice. Alice wants to run a subnet. She burns $TAO, claims a slot, and writes:

Bob shows up to mine on Alice's subnet. Alice never learns who he is. A miner is anonymous and competes with every other miner on that subnet for the same pool.

Bob does the work and submits it. Validators score miners and publish weights. Then the chain takes the stake-weighted median of the validator scores. If a validator scored Bob above that consensus, the chain clips that validator back down to it.

The median ignores small disagreements and only catches large ones. A validator can be early on Bob, maybe even right, and still get pulled back toward the middle. Validators cannot weight themselves, and only the top validators by stake are permitted to set weights.

When the scoring settles, the emission pool splits automatically. 41% goes to miners like Bob. Another 41% goes to validators and their stakers. The remaining 18% goes to the subnet owner, which here is Alice.

The jobs are concrete too. Chutes runs serverless inference. Templar does distributed pre-training. Targon sells confidential GPU compute. Vanta blends trading signals from competing strategies. Hippius does storage.

The priced slot

A subnet does not keep its slot only because the code still runs. When subnet slots are full, a new registration removes whichever existing subnet has the lowest moving-average token price.

If that lowest price belongs to Alice's subnet, her slot is the one the new registration takes. No manager has to say she failed. The rule compares moving-average prices and removes the lowest one.

New subnets receive a few months of immunity before they can be replaced. The grace period ends. That seems important to me, because a strange job may need time before its work is easy to judge.

In February 2025, dTAO gave every subnet its own token and its own liquidity pool. As of June 2026, a subnet's share of emissions is proportional to its moving-average token price.

Read that slowly. Price is not only a reaction to the subnet after the work is scored. It also sets how much new work the subnet can buy.

The root

The same pressure shows up at the top of the chain. The root upgrade shipped over the summer. Root dividends previously left the system by being claimed and sold for $TAO. Now each validator directs those dividends into a curated basket of subnet alpha that compounds.

Estimates put the reduction in automatic annual sell pressure from those root dividend changes at up to a third. What that's worth, I couldn't tell you. This is not investment advice. At least one validator has publicly argued that the root design carries substantial risks.

For the pattern, the useful point is narrower. Flow that used to leave now gets pointed back toward priced subnets.

The wrapper

Outside the chain, there have been filings to wrap Bittensor exposure in exchange products. As I write this, none of those filed products trade as ETFs.

If a wrapper ever trades, it probably makes exposure easier to hold. It does not make the inside less anonymous, less scored, less clipped, or less priced. The held thing would still be a selection process that pays for intelligence and can remove the subnet priced lowest by the crowd.

Off-chain versions

Alright. Let's back away from the trees. Here's a couple off-chain versions to explain:

A community garden could run this way. Anyone claims a plot by posting a refundable deposit. The shed door says what counts as maintained, who checks, how complaints get averaged, and what happens when space runs out. New gardeners get a protected start. After that, the plot with the weakest upkeep record goes back to the waitlist.

An open mic can carry it too. Signup is open to anyone who can write a name on the list. The score is not one judge's taste. It is a rough middle of laughter, silence, and attention, with the host clipping the regular who calls every set historic.

A shared workshop has the same bones. Members bring projects, tools are scarce, and booking protects new members for a while before it starts favoring benches whose work gets used, borrowed, repaired, and asked about.

The portable rule is: charge something to enter, write the scoring down before the work starts, let many judgments meet in the middle, and make survival follow a public signal instead of a manager's mood.

The open question

I don't know whether moving-average price discovers good work here, or whether it mostly discovers work the current crowd can recognize quickly. Those may look identical during ordinary weeks. They may separate when a subnet is young, strange, and still inside its immune months.

I also don't know how much protection a new idea deserves before the market is allowed to stop paying for it. The mechanism answers with a number. I don't have that number settled.