Staking Overview
AET holders can stake their tokens into a Neuronet to support it and earn a share of its rewards. In AETRON, staking is not a vote for a validator. It is delegated to a Neuronet, and it does something concrete: it sets how much work that Neuronet can be paid for.
Stake Is Capacity
This is the core idea of staking in AETRON. The amount of AET staked on a Neuronet determines its capacity, the volume of work it can earn emission for in each epoch.
capacity = (total staked planck / PlanckPerAet) × CapacityPerAet
Both values are network parameters set in the runtime:
| Parameter | Value | Meaning |
|---|---|---|
CapacityPerAet | 100 | slots granted per whole AET, per epoch |
PlanckPerAet | 1,000,000,000 | planck (the chain's smallest unit) in one AET |
One staked AET buys 100 slots of paid work per epoch. The division by PlanckPerAet is integer division, so the conversion rounds down to whole AET: a Neuronet holding 1,500.9 AET is treated as 1,500 AET and gets 150,000 slots. Fractional remainders add nothing until they round up to the next whole token.
Capacity is not a balance that accumulates. It is a per-epoch budget. The slot counter resets when the Neuronet settles its epoch, so unspent capacity does not carry forward and heavy usage in one epoch does not borrow against the next.
What a Slot Is
A slot is the protocol's unit of accounted demand. It is deliberately model-agnostic, so a Neuronet running a small encoder and one running a large language model are measured on the same scale.
Each task inside a Neuronet declares its own slot_cost_per_request, set by the Owner when the task is configured. The runtime rejects a value of zero, because a free request would let a Neuronet serve unlimited paid work on no stake at all. The rough shape of the costs:
- Inference requests cost 1 slot each. This is the reference point the rest of the scale is built around.
- Shadow replay checks are indistinguishable from real traffic, so they arrive as ordinary requests and consume a slot the same way.
- Training steps cost more than one slot, because a Witnessed Checkpoint step is far heavier than a single forward pass. The exact multiplier is a task parameter rather than a protocol constant.
- Pulse benchmark probes and challenge traffic sit off-budget and consume nothing.
Slots are a gate, not a reward weight. How much emission a miner earns from the work inside the budget is measured separately in compute units, using each task's cu_per_request. A slot decides whether a request is eligible to be paid; the compute unit decides how large that payment is relative to other miners. See Emission System for how the shares are worked out.
How Capacity Is Spent
Capacity is consumed when a Neuronet Owner submits the aggregated batch for an epoch. The batch lists, per miner, how many requests that miner served. The runtime then does three things:
- It computes weighted demand by summing
requests × slot_cost_per_requestover every miner that is actually registered to this Neuronet. Entries pointing at a hotkey from another Neuronet are ignored, so a batch cannot inflate demand with outsiders. - It consumes up to the remaining free capacity. If weighted demand exceeds what is left, only the part that fits is credited and the rest is dropped, with a
CapacityExceededevent recording how many slots were refused. - It distributes the credited slots across the miners in the order they appear in the batch, until the budget runs out.
That third step is worth knowing about: overflow does not spread evenly across a Neuronet's miners. It lands on whoever comes last once the budget is exhausted.
A Worked Example
Take a Neuronet with 1,000 AET staked. Its capacity is:
1,000 AET × 100 slots/AET = 100,000 slots per epoch
Suppose it runs two tasks: an inference task at the standard cost of 1 slot per request, and a training task where the Owner set the cost to 10 slots per step.
| Task | Volume | Slot cost | Slots used |
|---|---|---|---|
| Inference | 60,000 requests | 1 | 60,000 |
| Training | 3,000 steps | 10 | 30,000 |
| Total | 90,000 |
Everything fits inside the 100,000 budget, and 10,000 slots go unused. All 63,000 units of work are eligible for emission.
Now demand grows and the next epoch brings 100,000 inference requests plus the same 3,000 training steps, for 130,000 weighted slots. The budget is still 100,000. The first 100,000 slots are credited and the remaining 30,000 are refused. Those requests were served by the miners and answered for the user, but they earn zero emission for anyone: not the miners, not the Owner, not the stakers.
To cover 130,000 slots the Neuronet needs:
130,000 slots / 100 slots per AET = 1,300 AET staked
So the Neuronet has to attract another 300 AET, either from the Owner or from delegating stakers, before that traffic starts paying again.
Why This Matters
Work beyond a Neuronet's capacity is still served, but it earns nothing. This is a hard limit rather than a soft slowdown: slots over the budget pay zero emission. The effect is a real market for stake:
- A Neuronet with rising demand needs more stake to keep earning on all its traffic.
- A Neuronet with stake but no real usage earns almost nothing, because capacity only pays when there is work to fill it.
- Miners gravitate toward well-staked Neuronets where their work has room to earn.
Stake therefore has value only when paired with genuine demand. This is what links the token to real activity rather than speculation. It also means the two failure modes are symmetric: stake without traffic wastes locked tokens, and traffic without stake wastes compute.
Where to Go Next
- Staking Rewards: what stakers earn and how a Neuronet activates.
- How to Stake: the practical steps, cooldowns, and things to keep in mind.
- Neuronets: how capacity fits into the wider Neuronet economy.
- Emission System: how credited work turns into an emission share.