Gonka is a blockchain where the mining is AI work. Instead of burning electricity on hash puzzles or locking up capital in staked tokens, the hardware providers who secure the network earn their consensus power by running large language models — generating proofs of compute between epochs and serving real inference requests the rest of the time.

That single design decision drives almost everything else about the network: how validators are chosen, how fees are priced, how rewards are minted, and what you see when you open a block in the GNKScan explorer. This guide walks through the whole system, from the problem Gonka sets out to solve down to the individual transaction types you can watch land on chain.

The problem: idle compute and expensive trust

The whitepaper behind Gonka ("Decentralized AI: Meaningful utilization of computational power for real-world application") starts from a simple observation: demand for AI computation keeps climbing, while enormous amounts of GPU capacity sit idle or underused outside the big clouds. Classic blockchains do not help here. Proof-of-work networks consume compute on deliberately useless puzzles, and proof-of-stake networks tie security to locked capital rather than to anything productive.

Gonka's answer is to make useful AI computation the scarce resource that secures the chain. Hardware providers — the whitepaper calls them Hosts, the chain calls them participants — prove how much ML compute they control, receive consensus voting power in proportion to it, and then spend the bulk of every epoch doing paid inference work for users.

Three kinds of nodes, one network

Under the hood, a Gonka participant runs a small stack of cooperating services:

• ML nodes do the actual GPU work. They generate proof-of-compute batches during the competition phase and execute inference (and validation of other participants' inference) the rest of the time.
• Decentralized API nodes orchestrate. They manage a participant's fleet of ML nodes, track the chain's epoch phases, relay proof batches to the blockchain, and accept user inference requests.
• Chain nodes run the blockchain itself — a modified Cosmos SDK application whose staking module has been rewired so that validator power comes from computational results instead of bonded tokens.

The modification to the staking layer is the heart of the design. A function called SetComputeValidators replaces token-weighted voting power with compute-weighted power at the end of every epoch. No tokens are bonded for consensus at all; slashing for validator misbehavior reduces a participant's compute-derived score and hits the separate collateral system instead of burning a stake.

Epochs: the network's heartbeat

Gonka time is divided into epochs, and each epoch cycles through the same stages:

1. Proof-of-compute generation — every participant's ML nodes race to generate proof batches seeded by a recent block, a structured competition the whitepaper calls a Sprint.
2. Validation — participants cross-check each other's submitted batches by re-generating a sample of the claimed proofs.
3. Results computation — the chain tallies validations. A participant's work is accepted when more than half of the network, weighted by compute, vouches for it.
4. Validator set update — accepted participants get consensus voting power proportional to their proven compute, and the next epoch begins.

Between those bookends, the network does its real job: serving inference. Requests are routed to participants in proportion to their proven compute, so the same measurement that secures the chain also decides who gets paid work. If you want the stage-by-stage detail, we cover it in How Gonka's Proof of Compute Works and Gonka Epochs Explained.

How the money works

Gonka's fee model has an unusual property: ordinary blockchain transactions are free. The chain charges no gas. All economic weight rests on inference fees, which are priced in an abstract unit:

Final fee = (prompt tokens + actual completion tokens) × units-of-compute-per-token × unit-of-compute price

Each model registered on the chain carries its own units-of-compute-per-token figure, and the price of a unit of compute is set every epoch as a weighted median of participant proposals. When a user submits a request, the maximum possible cost goes into escrow; after the response is generated, the participant is paid for the tokens actually processed and the remainder is refunded automatically.

Participants earn two streams at the end of each epoch: work coins (the escrowed fees for inference they actually served) and reward coins (newly minted GNK, distributed in proportion to each participant's share of the epoch's total work). A subsidy schedule tapers the minted rewards over time, and in production configurations rewards vest gradually — on the order of 180 epochs — rather than unlocking at once. The full picture, including the collateral and slashing system, is in Gonka Tokenomics: How GNK Fees and Rewards Work.

The network today

As of early July 2026 (around epoch 318), the network's current epoch has roughly 35 active participants operating on the order of 196 ML nodes, backed by roughly a thousand live GPUs — a fleet dominated by NVIDIA B200, H100, H200, B300 and A100 hardware. Participants serve large open-weight models registered through on-chain governance — the current registry lists MiniMax M2.7, Kimi K2.6 and GLM 5.2.

Two numbers are easy to confuse when you first look at Gonka dashboards: the count of active participants in the current epoch (a few dozen entities actually doing work) and the cumulative count of registered participant addresses since genesis (thousands). GNKScan's participants page and network analytics report the active set, because that is what secures the chain right now.

Reading the chain

Because inference is the network's core activity, Gonka blocks look different from a typical Cosmos chain. Scroll the live transaction feed for a few minutes and you will mostly see the inference lifecycle — start-inference and finish-inference messages — punctuated by proof-of-compute batch submissions near epoch boundaries, validation messages, reward claims, and ordinary transfers.

A few pointers for exploring:

• Blocks shows the chain tip with per-block transaction counts and gas totals. Empty blocks are common and normal — block production never stops, inference traffic ebbs and flows.
• Transactions is the firehose, with each message type labeled.
• Participants lists who is doing the work and with how much weight.
• Inference and Models show the AI side: which models are live and how much inference the network is processing.
• Analytics tracks longer-run trends.

If a field on a block page shows "—", that is deliberate honesty rather than missing polish: some upstream data sources expose only partial block data, and GNKScan never fabricates values it cannot verify. We wrote about what every field means — and where each one comes from — in How to Read a Gonka Block.

Why it matters

Plenty of projects have promised decentralized AI compute. Gonka's contribution is architectural: it does not bolt an AI marketplace onto an existing chain, it makes verified AI compute the consensus mechanism itself. The same GPUs that vote on blocks serve the models; the same measurement that allocates voting power allocates paid work; and the fee system prices exactly one thing — tokens processed by a model — rather than layering gas markets on top.

Whether that design wins in the long run is an open question, and one you can watch play out in real time. The chain is public, the workload is visible, and every epoch's competition, validation and settlement lands in blocks you can inspect.

Keep reading

• How Gonka's Proof of Compute Works — the Sprint, batch validation, and the dual power system.
• Gonka Tokenomics: How GNK Fees and Rewards Work — units of compute, escrow, minting, vesting, collateral.
• Gonka Epochs Explained — the full epoch lifecycle.
• How to Read a Gonka Block — a field guide to the explorer.
• Gonka FAQ — quick answers to the questions we get most.