inferva

Getting started

How the flywheel works

What happens to a single request, and where its money goes.

Every request to the endpoint goes through the same five steps. Nothing is paid out that did not come from an agent's payment.

  1. 1

    Authenticate and rate limit

    The agent key identifies the paying account. The account's stake decides its tier, and the tier decides its rate limit.

  2. 2

    Route

    "model": "auto" goes to the cheapest active provider. A specific model id goes to the provider that serves it. Each model id is served by exactly one active provider.

  3. 3

    Check funds and capacity

    Before generating anything, INFERVA checks the worst-case cost against the payer's USDG balance and makes sure a sponsored capacity position has room. If either check fails the request is rejected and nothing is charged.

  4. 4

    Meter and charge

    Token counts from the reply set the price. The payer is charged, and the capacity position with the most room left is drawn down by the provider cost.

  5. 5

    Split and settle

    The protocol fee is split between stakers, the sponsor and the treasury, and a settlement record is written that hashes the previous record.

The price of a request

provider_cost = tokens × provider price per 1k tokens / 1000
discount      = provider_cost × tier discount
protocol_fee  = provider_cost × fee rate
agent pays    = provider_cost − discount + protocol_fee

The protocol fee is then split:

RecipientShare of the fee (sandbox value)How it is paid
Stakers50%Added to a fee index; each staker claims in proportion to stake
Capacity sponsor40%Credited to the owner of the position that served the request, together with the provider cost
Treasury10% (the remainder)Held by the protocol
The sandbox fee rate is 10% of provider cost. If nobody is staked when a request settles, the stakers' share goes to the treasury instead.

Worked example

An agent at Tier 1 sends a request that uses 1,000 tokens and is routed to Basalt Inference, a seeded sandbox provider priced at 0.0018 USDG per 1k tokens. With sandbox parameters:

LineUSDG
Provider cost0.0018
Tier 1 discount (5%)−0.00009
Protocol fee (10%)+0.00018
Agent pays0.00189
To stakers (50% of fee)0.00009
To sponsor (40% of fee)0.000072
To treasury (rest)0.000018

Amounts are tracked as integers in micro units (1 USDG = 1,000,000), so the provider cost and fee round up and the shares round down.

Why volume matters

Stakers and sponsors are paid per request. More agents sending more requests means more fees to split. There are no emissions behind the rewards: if agents stop paying, rewards stop.

Was this page helpful?