Gonka Inference for Builders: Lifecycle, Fees, and Explorer Signals
How AI inference works on Gonka from a builder's view — escrowed fees, epoch phases, validation, and which GNKScan pages and APIs to watch while integrating.
Gonka sells one product: AI inference settled on a chain secured by the same GPUs that run the models. If you are integrating as a client, a participant operator, or a data consumer, the explorer is your ground truth for "is the network in Inference phase, who is active, and is the indexer caught up?" This guide ties the product lifecycle to GNKScan surfaces.
Lifecycle in four beats
- Escrow — Before work, the max possible cost (from max completion tokens × model units × unit price) is locked. Users are not surprised by unbounded bills; leftovers refund.
- Execute — An active participant's API/ML path runs the model. Ordinary chain txs remain gas-free; the fee is the inference market, not blockspace.
- Validate — Outputs are checked at rates that scale with volume. Bad work risks collateral and reputation in the epoch process.
- Settle — At epoch boundaries, work coins (user fees) and reward coins (minted share of work) move according to protocol rules.
Deep dives: tokenomics, proof of compute, epochs.
What phase are we in?
Open GNKScan home or Inference. Phase comes from the gateway's /v1/epochs/latest (for example Inference, or PoC-related phases near the boundary). Progress percent is height-based between PoC start and next PoC start.
Builders should not assume 24/7 identical capacity: near PoC windows, GPUs pivot to proving; live GPU counts fluctuate by design.
Who can serve you?
The Inference page lists active participants with weights, models, and node counts. That is the operational set for the epoch — not the full historical registration list. Model catalog and units-of-compute parameters live on Models.
Explorer signals while debugging
| Symptom | Check |
|---|---|
| Client timeouts / empty model list | Epoch phase; participant inference URLs; status page |
| "Network looks dead" | Blocks still advancing? Indexer lag chip? |
| Fee surprises | Model units-of-compute and epoch unit price governance — not gas |
| Wrong fleet size in dashboards | Use GNKScan live GPUs (active-scoped), not unfiltered hardware dumps |
| Inference count tile empty | Indexer may still be catching inference-labeled txs; on-chain request counters have been stuck at zero historically |
Home Inferences (est.) counts inference-labeled transactions in the current epoch height range when the scanner exposes that stats API — an index estimate, not the dead number_of_requests field.
APIs worth bookmarking
Public, cache-friendly JSON (see response Cache-Control):
/api/overview— aggregated home console payload/api/inference/stats— epoch phase, participants, live GPUs/api/models— registered models/api/market/gnk— price quote/api/blocks/latestand/api/transactions/latest— feeds
For agent/tooling context dumps, prefer /llms.txt.
Security and honesty notes
- GNKScan is an explorer, not a wallet or inference router. It does not custody keys or proxy paid inference as a product surface (paid RPC remains deferred).
- Missing fields render as "—" rather than guessed defaults.
- When integrating alerts, key off indexer lag and phase, not a single GPU number.