Neuronets

Neuronets are the foundational building blocks of the AETRON network. A Neuronet is a self-contained AI service: a marketplace for one or more AI tasks where miners produce results, the network verifies them cryptographically, and emission flows to whoever produced proven, useful work.

What a Neuronet Is

A Neuronet brings three kinds of participants together around a shared set of tasks:

• An Owner creates it, defines its tasks and parameters, and runs the service.
• Miners provide compute, serve requests, and re-check each other's work.
• Stakers delegate AET to it, which raises its capacity and earns them a share of its emission.

Verification is not a separate role. Miners check each other, and the protocol assigns which work gets re-executed. Results that pass verification earn emission; results that fail earn nothing and cost the offender reputation. See Proof of Intelligence for how this works.

No Slots, No Fixed Seats

Unlike networks that cap each network at a fixed number of miner and validator seats, AETRON has no slot scarcity:

• The number of Neuronets is unlimited.
• The number of miners in a Neuronet is unlimited. Any miner that meets the requirements can join.
• A new miner is not forced to evict an existing one. It joins if it meets the requirements.

Spam is held back by economics rather than slot caps: a registration fee that rises with bursts of activity, a minimum stake to activate a Neuronet, and the rule that only verified work earns emission.

Tasks

A single Neuronet can host several tasks at once, each with its own model and verification settings. A Neuronet might, for example, run a language model, an image model, and a speech model side by side. The runtime allows up to 64 tasks in one Neuronet. Typical task areas include:

• Text and language: generation, translation, summarization, question answering
• Images and video: generation, classification, detection
• Audio and speech: recognition, synthesis, processing
• Inference and training: serving models on demand, training new ones
• AI agents: multi-step tool use and workflow automation

Task parameters are covered in Task Configuration.

Lifecycle

Creation

The Owner submits a creation transaction with the first task's configuration and, optionally, an initial stake. The runtime validates the task config, charges the registration fee, assigns the Neuronet an id, and records the Owner. If an initial stake was supplied, it is delegated to the new Neuronet in the same transaction, so an Owner can bootstrap capacity without a second step.

The fee is not a deposit and is never refunded. It is withdrawn from the Owner's balance and split three ways: 90 percent into the emission recirculation pool, 5 percent to the treasury, and 5 percent to the validator emission pool. The AET is recycled back into future emission rather than destroyed permanently. See Recycling.

Recruiting Miners

A Neuronet exists on chain before any miner joins it. Its tasks describe the model, the hardware class, and the verification settings, and miners read that registry to decide where to work. In a standard public Neuronet miners register themselves once they meet the task requirements, with a small burn at registration. There is no application queue and no per-miner approval step. In a Confidential (TEE) Neuronet the Owner controls access directly and approves, invites, or removes miners.

A Neuronet also needs stake before it is paid. The activation minimum in the runtime is 100 AET of total stake across all its stakers. Below that the Neuronet still exists and still accrues: emission keeps accumulating as a pending balance, and the settlement boundary simply skips the payout and emits a below-threshold event instead. Cross 100 AET later and the accrued balance can still be paid out. See Staking Rewards for what happens to it in the meantime.

Operation

Once active, miners serve requests, build validation artifacts, and re-check each other's sampled work. The Owner can add or remove tasks, change task weights, lower the creator fee, and publish routing details for the gateway and P2P mesh. Every one of these is an on-chain call that takes effect from the next block.

Settlement is per Neuronet rather than on one global clock. Each Neuronet has its own tempo, defaulting to roughly one day of blocks, and Owners can set it within protocol bounds. New Neuronets are scattered across the settlement queue so that they do not all settle in the same block.

The runtime also tracks activity. A Neuronet with no activity for about 7 days is marked as warned, and one with no activity for about 30 days is frozen. A frozen Neuronet stops accruing and drops out of the settlement queue until activity resumes.

Dissolution

The Owner can dissolve a Neuronet at any time. Dissolution unwinds it in a defined order: staked AET is queued back to stakers through the normal unstake cooldown, any emission the Owner accrued but had not yet been paid is settled and paid out, the Neuronet's contribution is removed from the network-wide useful-work denominator, and its tasks, identity, and miner invitations are cleared from storage. The identity deposit comes back. The registration fee does not.

What Creation Costs

Registration is priced on a moving curve rather than a flat rate. The runtime sets a base price for each kind of Neuronet:

Neuronet type	Base registration cost
Standard (public)	1000 AET
Confidential (private, TEE)	5000 AET

Adding a task costs 10 AET on top of that.

The current price starts at the base and doubles after every successful registration of that kind. Between registrations it decays linearly back toward the base over roughly one day of blocks. Two effects follow. A quiet network sits at or near the base price, so an ordinary Owner pays close to 1000 AET. A burst of registrations, which is what a spam attack looks like, drives the price up geometrically within minutes, and the attacker pays for every one.

The doubling is bounded at the base price multiplied by 2 to the power of 20. That ceiling exists to stop an arithmetic overflow, not as a realistic price target: for a public Neuronet it works out above a billion AET, far beyond the 33M AET supply cap. In practice the decay pulls the price back down long before it approaches that bound.

The private base is five times the public one because a Confidential Neuronet consumes scarcer resources and carries stronger access guarantees, so it should not be as cheap to spin up speculatively.

The Owner Role and the Fee

The Owner is the operator of the Neuronet, not a passive rentier. Owners define the tasks, choose the models and verification tiers, keep discovery and gateway routing current, and decide when to add capacity or wind the service down. Ownership is a single field on chain and can be transferred to another coldkey, which is how a Neuronet moves out of a custodial arrangement into the client's own keys.

For that work the Owner takes a percentage cut of the Neuronet's emission. The protocol caps that cut at 10 percent, with a floor of 0. The cap is checked when the task config is validated, so a Neuronet cannot be created with a higher fee in the first place.

The fee moves in one direction only. The runtime exposes decrease_owner_fee (call index 6) and no matching increase. The call rejects any value that is not strictly below the current fee, and rejects anything under the network minimum. An Owner who lowers the fee to attract miners and stakers cannot raise it back once the participants have committed. This is the main protection stakers and miners have against an Owner changing the economics after the fact.

Capacity Comes From Stake

A Neuronet's capacity, the amount of work it can be paid for in an epoch, is set by how much AET is staked on it. The runtime converts stake into capacity at a fixed rate of 100 slots per AET, so a Neuronet holding 1000 AET of stake carries 100,000 slots of paid capacity per epoch. Work is charged against that budget as it is verified, and requests beyond the budget are still served but earn nothing.

This ties three things together. Stake sets capacity, capacity sets how much verified work can be rewarded, and the reward is what draws miners to serve the Neuronet in the first place. Stake alone does not pay: capacity without real demand sits unused, and demand beyond capacity does not earn. A Neuronet is well configured when its stake tracks its actual traffic.

Stake is locked rather than burned and returns after a cooldown, which the runtime sets at roughly 7 days for a standard unstake. See Staking for the staker side.

How a Neuronet Earns

Each block, 90 percent of emission goes to Neuronets and is divided by how much verified useful work each one produced. Within a Neuronet, that share is split again:

• The Owner takes the creator fee, capped at 10 percent.
• Stakers take a fixed 10 percent share, in proportion to the amount each staked.
• Miners take everything left over, weighted by their verified work and reputation.

Because the Owner fee and the staker share are both capped, miners receive at least 80 percent of a Neuronet's emission, and more when the Owner has lowered the fee. A Neuronet only earns when its miners do real, verifiable work, and that income is shared with the people who fund its capacity.

Choosing a Neuronet

For Miners

• Task fit. Does the Neuronet's model and hardware requirement match your setup?
• Capacity headroom. A well-staked Neuronet with real demand has room for your work to earn.
• Reward economics. Check the creator fee and how emission is split, then weigh it against the registration cost.
• Requirements. Confirm you can meet the proof-of-compute threshold and the canonical execution settings.

For Stakers

• Real demand. Capacity only pays when there is genuine usage, so look for Neuronets people actually use.
• Miner health. A Neuronet with active, reliable miners turns capacity into verified work.
• Growth. Rising usage means a larger economy to share in.

What the Owner Does Not Control

Some parameters are network-level settings governed by the protocol rather than knobs an individual Owner turns:

• The registration fee schedule, its base prices, and the doubling and decay behaviour
• The minimum activation stake and the capacity rate per AET
• The Owner fee ceiling and the staker share
• The unstake cooldown
• Verification thresholds, quorum size, and sampling rate

Neuronet Configuration covers the settings an Owner does control, and Neuronet State describes what is stored on chain for each one.