Table of Contents
Key takeaways
- Creating a subnet means registering a new AI market on Bittensor, then running it as the owner who sets the rules.
- Registration requires locking TAO, and that lock cost rises and falls with demand for a limited number of slots.
- The network caps active subnets at 128, with an expansion to 256 planned, so a new subnet can displace the weakest existing one.
- New subnets get a multi-month immunity period that shields them from deregistration while they establish themselves.
- Under Dynamic TAO, registering a subnet also spins up its own alpha token and a TAO/alpha liquidity pool.
Subnets are where the network's AI actually gets produced, and the supply of them is not infinite, so creating one is part engineering project and part economic commitment. This guide explains what really happens when a fresh subnet appears on-chain, and it walks through both the mechanics and the incentives that shape them.
What It Means To Create A Subnet
A subnet is a self-contained market for one kind of AI work, whether that is inference, compute, prediction, training, or data. Creating one is close to launching a startup inside the network, because the person who registers a subnet becomes its owner and inherits real responsibilities. The owner defines the task, writes the incentive mechanism that decides how miners are scored, maintains the code, and works to grow demand around the market they opened.
That ownership carries an economic reward, since the standard emission split sends 18% of a subnet's rewards to the owner, with 41% going to validators and 41% to miners. The reward arrives in the subnet's own token, so an owner who builds something the market ignores ends up holding a share of very little.
The Lock Cost And Why It Moves
Registering a subnet on mainnet requires locking a quantity of TAO, and that amount is not fixed. The protocol runs the price like a Dutch auction, so the lock cost doubles each time a new subnet registers and then falls slowly, block by block, until the next registration resets it higher. When demand for slots runs hot the cost can climb into the millions of dollars, and when interest cools it drifts back down for whoever is patient enough to wait.
An important detail often gets lost in the headlines about how expensive registration can be. The registration cost is a sunk cost since a portion seeds the new subnet's TAO/alpha liquidity pool and the rest is recycled back into the network. It is not returned to the original registrant if the subnet is later deregistered, so a builder should treat the money as spent the moment they commit it.
Winning A Slot In A Capped Network
Bittensor does not let subnets multiply without limit, and the cap is the reason creation is competitive. The network currently holds 128 slots, with an expansion to 256 planned, so a builder who registers when every slot is full is not just adding a subnet, they are triggering the removal of another one.
The protocol decides which subnet leaves by looking at the lowest EMA price, a smoothed average of each subnet's alpha token value over time, and the weakest one outside its protection window gets deregistered when the newcomer arrives. That protection matters for anyone launching, because new subnets receive an immunity period of several months during which they cannot be removed no matter how low their token trades, which buys a team time to ship, attract miners, and build the community support that keeps their price alive.
What Goes Live When A Subnet Registers
Since the Dynamic TAO upgrade went live in February 2025, creating a subnet does also brings a small economy into existence. Each new subnet gets its own alpha token and an on-chain TAO/alpha liquidity pool, and every alpha token carries a 21 million supply cap that mirrors TAO itself.
From that moment the subnet is plugged into the network's market for emissions, where TAO holders stake into the pool, receive the alpha token, and effectively vote with their capital on whether the subnet deserves a larger share of newly minted TAO. A subnet owner therefore does not just launch software, they launch a token market that will judge their work continuously, and that judgment feeds directly back into how much the subnet earns.
The Work That Surrounds Registration
Before paying the lock cost, a serious team designs the task and the scoring logic so that miners cannot easily game it, then tests the whole mechanism on Bittensor's test network where registration is cheap and mistakes are harmless. Getting the incentive design right is the hard problem, because a good scoring system for price prediction looks nothing like a good one for model serving, and a weak mechanism ends up paying the wrong participants.
After registration, the owner still has to attract the miners who produce the work and the validators who score it, and both groups pay their own dynamic burn to claim a slot inside the subnet. The owner also tunes the subnet's parameters, communicates with stakers, and hunts for demand outside the network, since a subnet that depends only on emissions tends to struggle once the initial curiosity fades.
Why The Barrier Is Economic, Not Just Technical
Almost anyone with the engineering skill can write a subnet, but due to the capped allowance for subnets, keeping one alive is a different test entirely. Creation forces a team to put capital at risk through the lock cost, to survive a competitive slot market, and to earn continuous confidence from stakers who can withdraw their support at any time. This is a deliberate structure because it filters out subnets that cannot demonstrate real value and pushes builders toward products that people outside the staking loop actually want to pay for. The teams most likely to hold their slot are the ones that treat registration as the start of the work.
The subnet model is key for understanding how Bittensor operates as a network and is vital for the construction of top-tier solutions that decentralized AI can offer to the world.
